As data centers are rapidly constructed around the world, people are inevitably beginning to question their sustainability and the impact they have on the communities that surround them.

From an enterprise perspective, there is even more to wrestle with. Data centers are long-term investments that need upgrades and maintenance to continue operating. Is there a cheaper and more environmentally friendly way to do this?

Chris Carreiro, the CTO of global IT infrastructure services firm Park Place Technologies, joins Jane and Bobby to discuss the pros and cons of flatpack data centers and how modular computing can help enterprises scale up.

Highlights

"These are now purpose-built for higher workloads – a few years ago we were building data centers for 5 and 10 kW cabinets; every rack was about 10KW – now we're going into the 40-40 kW on a regular build, and an AI facility, AI factory, is demanding 100 kW per rack now."

"A traditional data center, two years of trucks and construction carry us forward into maybe 5 to 7 years down the road without adding additional cooling capacity or additional pieces. So I think there is a little bit of runway for some of those systems. But as far as upgrading, if the design is chasing the 500 kW cabinet that Nvidia is talking about, as long as there is space for those types of cooling elements, you would be able to leverage just upgrading a few components inside of a facility."

Links

• The great data center backlash threatens to disrupt the future
• Data center markets are booming in these EMEA countries
• AI data center growth stymied by power constraints

Enterprises still face huge problems pushing AI projects from pilot to production, according to new research from Dell Technologies, with a confluence of factors hampering success.

Infrastructure deficits, regulatory friction, and fragmented modernization efforts are creating major barriers for enterprises, the study noted. Indeed, nearly two-thirds (64%) of EMEA IT leaders said their infrastructure isn’t ready for production-scale AI and analytics workloads.

More than half (55%), meanwhile, pointed to security and compliance-related issues as a leading factor behind stalled projects.

Notably, UK decision-makers are feeling the strain more than their European peers.

Around 84% said that running AI, workplace, and data center modernization in silos creates severe operational challenges, compared with 69% across EMEA as a whole.

Just over six-in-ten said they lack a clear, unified roadmap across AI, data, and security, compared with an EMEA average of 46%.

UK leaders are also more likely to require partners to demonstrate clear business impact and risk management before progressing new AI projects, at 77% versus 66%.

"Businesses are entering the next phase of AI adoption. The focus is shifting from exploring what's possible to connecting the infrastructure, data, security and workplace technology needed to turn AI into real business value at scale," said Steve Young, SVP and UK & Ireland managing director for Dell Technologies.

"Organizations that build these foundations today will be the ones that lead tomorrow."

Uncertainty is hampering AI adoption

Across Europe as a whole, economic uncertainty continues to influence technology investment decisions, with eight-in-ten leaders admitting they are more selective about investment.

There's a growing trend towards simplification and consolidation, with 81% saying they'd prefer to meet more of their technology needs through a single strategic partner.

IT leaders are also demanding more from their providers, with 77% shifting investment towards partners with greater scale, ecosystem, and supply chain capabilities.

Security, compliance ,and privacy have emerged as critical considerations, with 64% of leaders saying they're concerned about complying with government regulations when using data for AI.

More than half (55%) also report that security and compliance concerns have already slowed or delayed AI adoption initiatives.

Legacy hardware pains

With infrastructure capabilities still lagging, IT leaders are now scrambling to bridge gaps and shortcomings, according to Dell.

Indeed, 86% of EMEA firms said they are making AI PCs central to their future strategy with AI. Overhauling legacy hardware is a significant challenge for many, however, and 79% noted that it has a direct impact on AI adoption.

"EMEA organizations are navigating a complex environment shaped by economic uncertainty, accelerating AI adoption, growing cyber threats, and evolving regulations," said the firm.

"As organizations look to balance innovation with resilience and compliance, they are increasingly focusing on technology investments that deliver measurable business outcomes and support long-term competitiveness.”

FOLLOW US ON SOCIAL MEDIA

A lengthy Google Cloud outage has raised questions over the UK’s reliance on a handful of cloud providers, with organizations across a range of industries having been affected.

The incident began just after 7am UK time on 20 October and affected Google Cloud’s europe-west2-b region, one of three local availability zones in London.

Notably, the outage lasted roughly eight hours, with Google mitigating the issue at 3pm and services coming back online an hour later. By 6pm, the hyperscaler said services were restored for most customers.

Only a “very small number” of those impacted in the outage were still experiencing issues, the company added.

In a statement in the wake of the outage, Google Cloud said it will publish a full analysis of the incident once it has completed an internal investigation, urging any customers still affected to contact its support team.

Who was affected in the Google Cloud outage?

Organizations across an array of industries were impacted by the Google Cloud outage.

One of the worst-affected was online bank Revolut, with customers unable to make card payments, withdraw cash, or make transfers for several hours.

The neobank said customers’ funds and data remained secure throughout the incident.

Customers of Lloyds Bank, Halifax, and Bank of Scotland, all part of the Lloyds Banking Group, also reported issues, as did HSBC, with some customers unable to log in via the app and online.

The outage also hit the video system used by defendants, lawyers, and other parties to attend remote court hearings across England and Wales. DPD and Royal Mail are also believed to have been affected, along with Screwfix and Rightmove.

What was the cause?

As noted, Google Cloud is still to publish a full post mortem of the outage, meaning that at this stage there is very little information on the exact cause of the incident.

What we do know so far is that Google has blamed the incident on a “power event”, which is believed to have taken place at its London cloud facility.

Google is preparing a report on the outage, which should be available in the coming days or weeks. ITPro will update this article accordingly when a full explanation from Google has been published.

What happens next?

The Google Cloud outage has triggered the interest of the Financial Conduct Authority (FCA) and the Bank of England’s Prudential Regulation Authority.

The company is designated under the Critical Third Parties regime as critical to the financial services sector, and as such the two authorities have requested more information.

Muhammad Yahya Patel, vCISO and cybersecurity advisor for EMEA at Huntress, said the incident will naturally prompt serious discussions about its impact and resilience.

“[The] outage will generate a significant number of post-incident reviews and likely some difficult conversations between firms and their regulators about whether their cloud resilience testing adequately reflected a scenario where a major provider goes down in a key region.”

What can we learn from the Google Cloud outage?

As with the AWS outage this time last year, experts have warned the incident highlights how many essential services rely on a handful of infrastructure providers.

Scott Dawson, CEO at DECTA UK, said there will undoubtedly be “uncomfortable conversations in IT departments” as a result of the outage.

“Banks and payment providers spend a fortune on keeping their systems running, but how many know whether their backup supplier is using the same Google infrastructure as their main one? That's something I'd want an answer to," he said.

"We've known for years that cloud services can fail, and there's really no excuse for waiting until customers are locked out of their accounts to find out whether your contingency plans work.”

Other industry experts have reiterated that the Google Cloud outage shows the importance of digital sovereignty and stronger regional infrastructure capabilities.

Mark Boost, CEO of UK-based cloud provider, Civo, said the outage highlights the risk of overreliance on foreign infrastructure providers.

"True operational resilience simply cannot exist when critical workloads are concentrated in a handful of foreign platforms; it requires open, hybrid, and multi-cloud environments,” he said.

“When the failure of a single foreign hyperscaler has a large-scale impact on the UK’s digital economy, then the tech stack is not robust; it is merely a business risk.”

Boost added that the outage should serve as a “clear directive” that greater investment is needed in UK sovereign capabilities.

Jon Lucas, director and co-founder of Hyve Managed Hosting, echoed Boost’s comments regarding overreliance on foreign cloud providers.

“The incident highlights how businesses relying too heavily on a single cloud provider risk being left without a viable alternative when services fail,” he said.

Lucas noted, however, that diversification of services isn’t quite enough.

“Workload portability is essential,” he said. “Businesses must be aware and avoid unnecessary vendor lock-in and ensure they can move critical applications and data between environments without major disruption.”

FOLLOW US ON SOCIAL MEDIA

Microsoft has reportedly pulled out of a deal with Nscale under which it would have been the anchor tenant for a major UK data center, over concerns about the power supply.

The development, in Loughton, Essex, was set to deliver 50 MW of AI capacity, scalable to 90 MW, and to initially house 23,040 NVIDIA GB300 GPUs. It's expected to create 250 permanent jobs.

But now, according to the Daily Telegraph, power supply problems mean that the project will face significant delays. UK Power Networks (UKPN), which owns the electricity grid across London and the south-east of England, is believed to have told Nscale that it won't be able to supply enough power to the site until the early to mid-2030s. The development was originally set to be finished early next year.

The two companies are still working together on projects in Norway, Portugal and the US – although the Norway facility is believed to be behind schedule too.

"We remain committed to our partnership with Nscale, including work to build robust digital infrastructure and support AI growth globally," a Microsoft spokesperson told ITPro. "This decision relates to a single site."

Supplying adequate power to data centers is becoming an increasing problem, with Gartner warning as far back as two years ago that 40% of data centers could face constraints in power availability by 2027.

And late last year, a report from Savills found that the issue is slowing data center construction, with only 850MW of power capacity delivered across EMEA since the start of 2025, down by 11% compared with the same period in the previous year. New transmission, distribution, and generation capacity could take years to come online, said the firm.

UKPN recently signed up as a strategic partner of the Data Centre Alliance (DCA) as part of efforts to strengthen engagement with the data center sector and ease supply problems, publishing an Introduction to Connections guide on the DCA's member portal.

To try and deal with power supply constraints, it said it could offer a range of connections beyond non-curtailable or unconstrained connections, including ramped connections, with flexible capacity that increases over time; flexible connections managed in real time; and profiled connections, in which customizable capacity profiles are created to match operational patterns.

Energy regulator Ofgem, meanwhile, has complained that many applications are speculative in nature, and is proposing that data center projects should have to pay between £237,500 and £712,500 per megawatt when they accept a connection offer.

Microsoft is currently the only customer to have signed up for the Loughton data center, and its decision to pull out could have an effect on Nscale's planned flotation in the US later this year, where it is targeting a valuation of $35 billion.

But the move is not only a problem for Nscale, but a potential embarrassment for the UK government. Last year, as part of its AI Opportunities Action Plan, it trumpeted plans from Nscale to invest $2.5 billion to support the UK's data center infrastructure over three years. These plans included the building of the Essex data center, described as the UK's largest sovereign AI data facility.

In a statement, Nscale told ITPro: "We remain committed to both the Loughton project and the UK and we are exploring alternative power sources for the site, in addition to other expansion sites in the UK. Our relationship with Microsoft remains strong and long-term, and this unfortunate London-wide grid issue has no impact on our other contracts."

FOLLOW US ON SOCIAL MEDIA

Ugreen's NASync desktop appliances are gaining a strong reputation in the NAS space as they score highly for build quality, hardware specifications, and prices. The premium Ultra series steps in at the top of the range and moves the focus up to growing businesses and power users that want a combination of high storage capacities, top performance, and plenty of features.

This is a review of Ugreen's flagship NASync DXP8800 Ultra 8-bay model, which has a modern Intel Core 7 150U mobile CPU in the driving seat. And eight efficient plus two multi-thread performance cores, a maximum 5.4GHz Turbo frequency, and a minuscule 15W base TDP. Along with a faster CPU, the only other notable difference with Ugreen's older DXP8800 Plus is that the maximum DDR5 memory capacity has been increased by 50% to 96GB.

The DXP8800 Ultra employs exactly the same solidly built aluminium chassis as the Plus model, so you get a veritable port-fest. A standout feature is the pair of 40GbE Thunderbolt 4 USB-C ports, but there are also dual 10GbE multi-Gigabit, USB-A 3.2 Gen2, and USB-A 2.0 ports, along with an integral SC4 card reader, while 8K video output from the CPU's Iris Xe Graphics 96EU chip is routed to an HDMI 2.1 port.

Ugreen NASync DXP8800 Ultra review: Hardware features

The main drive bays use tool-free telescopic carriers and you can pick from a wide choice of devices as Ugreen's compatibility list currently shows 57 different models which include plenty of NAS-specific HDDs from Seagate, Western Digital and Toshiba.

You'll also find two M.2 NVMe SSD PCIe 4 slots lurking behind a spring-loaded cover in the base which can be used as high performance storage pools or caches. Alongside them are two SODIMM slots with one occupied by the base 8GB of 5,600MHz DDR5 memory.

The rear magnetic dust cover has two 12cms diameter cooling fans behind it and during testing, we found they were whisper-quiet. There's also a PCIe Gen4 expansion slot alongside but although the user manual provides clear instructions, patience is required as it's very difficult to access.

Ugreen NASync DXP8800 Ultra review: UGOS and apps

For testing, we loaded four 4TB Seagate IronWolf NAS HDDs and used Ugreen's web portal to discover the appliance and install the latest UGOS software. The web console is easy to use and aids storage pool creation with sage advice on the various RAID options based on the drives selected.

We chose the recommended RAID5 volume and were prompted to create personal and shared folders and configure remote access, which creates a unique Ugreenlink ID and secure web link. We opted for a Btrfs volume to enable storage quotas and snapshots and noted that the latter feature still only supports NAS shares and not iSCSI LUNs.

Ugreen lags behind Qnap and Synology for app choices, but it is catching up as 35 are now available for the DXP8800 Ultra. The Model Manager and OpenClaw agent apps are of particular interest as they bring smart AI tools into play.

Model Manager provides a central location for managing AI features, but Ugreen's AI Console, Uliya, and Voice Memos apps are only supported by the iDX models. For the DXP8800 Ultra, the Model Manager integrates with the Photos and Surveillance Center apps and provides facilities for authorising various AI models for them, including people, text, and image recognition for photos, plus facial recognition and general detection for surveillance.

OpenClaw uses the Docker app and the latest UGOS updates its engine to 29.6.2, which addresses a number of security issues. If you're planning on using OpenClaw, you'll find Darien Graham-Smith's excellent in-depth tutorial in Issue 383 of PC Pro.

Other apps of distinction include Sync & Backup, which secures local data to a remote Ugreen NAS and vice versa, and also supports Rsync servers. The Cloud Drives app provides file syncing for ten cloud storage services including Dropbox, OneDrive and Google Drive, while Ugreen's NAS Windows and macOS desktop tools use sync tasks to secure selected data to shared folders and update it in real time.

The appliance's HDMI port comes into play only when you use Ugreen's mobile app. Load the Theatre app, pick a video to stream, select the HDMI Play projection mode, and sit back and enjoy.

Ugreen NASync DXP8800 Ultra review: 10GbE performance

For our speed tests, we hooked the appliance up to the lab's Zyxel XMG1930-30HP multi-Gig switch and mapped a share over 10GbE to a Dell PowerEdge Windows Server 2025 host. Ugreen added UGOS support for 9000-byte Jumbo frames earlier this year, and with these enabled, we saw good Iometer sequential read and write rates of 9.2Gbits/sec and 8.8Gbits/sec and random speeds of 9.2Gbits/sec and 0.7Gbits/sec.

For our real-world tests, copies of a large 25GB test file between the appliance and server delivered sustained read and write rates of 3.8Gbit/sec and 3Gbits/sec. Securing a 22.4GB folder with 10,500 small files to the share averaged a very respectable 2Gbits/sec.

From the SAN Manager app, we created a 500GB iSCSI target and, after mapping it to our host server, watched Iometer report back with very similar sequential and random read and write rates as the NAS share. We noted that although the SAN Manager clustering feature allows multiple clients to connect to the same iSCSI LUN, MPIO (multipath I/O) for redundant, load-balanced links is not currently supported.

For cache testing, we installed two Kingston M.2 NVMe SSDs and assigned them to our data volume as a mirrored read/write cache. This made significant improvements to random write operations with our NAS share and iSCSI LUN, seeing respective speed increases of 245% and 290%.

Ugreen NASync DXP8800 Ultra review: Is it worth it?

For an 8-bay NAS with dual TB4 and 10GbE ports, the NASync DXP8800 Ultra looks like a good value, with it currently listed on Ugreen's US site for $1,799. Competition comes mainly from Qnap's new TX-h966TX 9-bay hybrid NAS, which sports dual TB4 ports, a 6-core Core i3 CPU, a single 10GbE port, and support for up to four U.2 PCIe NVMe SSDs.

Ugreen's UGOS OS is a work in progress, but all the important apps are present, while the OpenClaw agent and Model Manager show its continued focus on developing AI-powered capabilities. At first glance, it does look to be a modest hardware upgrade over the DXP8800 Plus model, but the Core 7 150U CPU is noticeably faster across the board than the older Core i5-1235U, and memory capacity is 50% higher, so it can easily cope with current and future workload demands.

Ugreen NASync DXP8800 Ultra specifications

PC sales have fallen yet again as AI-driven price increases continue to bite, according to Gartner.

The trillion-dollar AI infrastructure rollout has led to shortages of key components, notably memory and storage, with one report suggesting AI will use a fifth of DRAM capacity this year.

Gartner said that global PC shipments were 55.4 million in the third quarter of 2026, down 22.4% from the 71.3 million shipped in the same quarter last year. That's on par with shipments from the first quarter of 2023, the analyst firm noted – and the decline is double the 10% annual reduction in PC shipments predicted by Gartner earlier this year.

That fall in shipments is largely down to rising costs of components driving up overall prices, Gartner said, leading to slower sales.

"The impact of higher memory and storage costs is now fully reflected in PC pricing," said Namrata Banerjee, principal analyst for research at Gartner. "DRAM costs surged five times year-over-year in the second quarter, and PC vendors passed those higher component costs on to customers, driving the steep decline in PC shipments during the third quarter."

Scarcity rather than oversupply

While sales have hit similar levels as they were two and a half years ago, the dips are driven by different causes, Gartner added.

"PC shipments have fallen to levels last seen in early 2023, though the market dynamics were noticeably different," said Banerjee. "Rather than excess inventory, the decline in the third quarter of 2026 was driven by higher prices, vendor-led reprioritization of scarce components and a shift toward higher-priced configurations."

The AI-driven components shortage has led PC makers to focus on higher-end models in order to help maintain margins and absorb some rising costs, with a separate report from Omdia earlier this year predicting reduced availability of low-budget devices alongside a wider rise in prices.

Banerjee suggested the absence of budget models was leading customers to hold off upgrading, while businesses focus on AI-focused investments. "The resulting reduction in the availability of lower-cost devices made replacement purchases less affordable for consumers, while businesses focused spending on targeted investments, such as workstations and AI PCs rather than broad-based PC deployments," Banerjee said.

Sales slide across the board

The sales decline hit all of the top PC makers, Gartner added.

Apple saw the smallest fall in shipments at 7.1% between the third quarter of this year and last year, but its PC rivals posted double-digit declines. Market leader Lenovo fell 22.4% over the same period, while second-ranked HP tumbled by nearly a third at 30.9%, and Dell wasn't far behind at 25.4%. Asus and Acer had slower slides at 13% and 11.2% respectively.

The report from Gartner follows a separate report from Counterpoint research showing global smartphone shipments had fallen 20% on-year to their lowest for the third quarter since 2021 – also driven by higher memory costs and supply shortages of other components. Over the summer, Apple admitted that the AI buildout would lead to more expensive iPhones.

HPE has launched a range of new servers optimized for enterprise inference and agentic AI.

The firm said the HPE ProLiant Gen13 Servers, powered by sixth-generation AMD Epyc Server CPUs, deliver higher core density, memory bandwidth, and I/O.

These servers are primarily targeted at compute-intensive workloads such as agentic AI and inference, analytics, virtualization, simulation, and modeling, the company said in an announcement.

“Organizations need to turn their investments in agentic AI and inference into business value without adding operational complexity or risk,” said Krista Satterthwaite, senior vice president and general manager for compute at HPE.

“For these workloads, HPE ProLiant Gen13 brings together optimized performance, built-in, automatic security, intelligent operations, and proven reliability. It gives customers a trusted foundation to modernize and innovate today and adapt as their requirements evolve.”

Under the hood of HPE's ProLiant Gen13 servers

The HPE ProLiant DL585a Gen13 Server is a high-density platform designed to support up to eight double-wide PCIe GPUs, two sixth-generation AMD Epyc Server processors, and PCIe Gen6 connectivity.

Conversely, the HPE ProLiant DL525 Gen13 Server is a more compact, performance-optimized 1U, single-socket system powered by one sixth-generation AMD EPYC Server processor with up to 256 cores.

HPE said this features high bandwidth 16-channel memory, with support for high-speed and high performance –DDR5 and MRDIMM – and PCIe Gen6 connectivity.

Notably, this is suited more for workloads with extremely high data throughput, such as fraud detection, market-data processing, risk and pricing models, electronic design automation (EDA), simulation, and AI inference.

Meanwhile, the HPE ProLiant XD245 Gen13 and HPE ProLiant XD285 Gen13 are designed to maximize compute performance within a smaller footprint. The two OCP-compliant servers combine high-density compute with efficient shared power and cooling to support data-intensive AI and HPC workloads.

Finally, the HPE ProLiant XD245 Gen13 Server uses a one-Open Rack Unit (1OU), four-node direct liquid-cooled design, and is powered by up to eight sixth-generation AMD EPYC Server processors; while the HPE ProLiant XD285 Gen13 Server uses a two-Open Rack Unit (2OU), two-node air-cooled design.

More on the horizon for HPE

According to HPE, there are more systems planned in the near future. The new servers combine automatic security from HPE Integrated Lights-Out 8 with enhanced HPE Compute Ops Management capabilities to protect data and system integrity.

The company will also introduce new features, including automatic self-encrypting drive (Auto-SED) server capabilities to protect data at rest from first boot without added configuration, external key managers, or manual activation.

Multi-person authorization (MPA) requires two authorized users to approve high-impact actions, including key provisioning and secure erase, while expanded post-quantum cryptography (PQC) readiness now extends beyond firmware to help protect server communication paths from emerging threats - a first in the industry, HPE said.

“AI is moving into the applications and operations that run the enterprise, fundamentally changing what organizations demand from their infrastructure,” said Dan McNamara, senior vice president and general manager, Compute and Enterprise AI, AMD.

“Across generations of AMD Epyc CPUs and HPE ProLiant servers, our teams have worked side by side to turn advances in processor and system design into practical value for customers,” McNamara added.

“With sixth-generation AMD EPYC processors and HPE ProLiant Gen13, we are delivering the performance, efficiency and flexibility enterprises need to scale AI while modernizing the critical workloads their businesses depend on.”

The HPE ProLiant DL525 Gen13 server will be available later in 2026, with the HPE ProLiant DL585a Gen13 server rolling out in early 2027 and the HPE ProLiant XD245 and XD285 Gen13 servers later that year.

FOLLOW US ON SOCIAL MEDIA

Is it time to upgrade the fleet of endpoints across your organization? The answer to that question may well have once come from a simple check of the calendar. In 2026, however, the concept of a three-to-four-year refresh cycle feels like something from a bygone era, like looking back on the charm of an old sitcom.

Life was simpler then; prices were more stable, and the improvements between hardware generations were more predictable. In the age of mobile computing, AI PCs, a components (specifically memory) supply crisis, and Microsoft's Windows OS proclamations, among other factors, the conventional hardware refresh cycle has shifted, experts tell ITPro.

In its place is a complex picture of different and dynamic cycles that vary based on not just the type of device in question, but its role within the business and how friendly any upgrade can be on IT budgets. In 2026 and beyond, organizations will be considering multiple angles – not just the calendar – when opting into an upgrade program.

Which factors are affecting the refresh cycle?

The cycle is dead, or maybe it isn't – perhaps it's somewhere in between. Nobody can agree on a consensus, least of all the experts that ITPro consulted with.

Michael Rosholt, SVP at M&A specialist and IT industry investment bank martinwolf, falls into the former camp but adds certain caveats: "It depends on the type of technology," he says, adding: "Rather than one universal refresh cycle, we're seeing the market increasingly segmented by technology and use case."

It might not even be describable as a cycle, says Adam Bowles, who runs the managed IT services company Act360 in Canada, telling ITPro that it's split by role and device type – where employees with less intensive performance requirements are hanging onto hardware for longer to save on budget – while 'power users' may see a new device every couple of years due to the recent acceleration in functionality.

So how did we get here? The earliest factors that drove some of the shifts we are seeing today first emerged with Microsoft and its Windows 11 hardware requirements, "forcing replacements on OS support grounds, not performance," as Bowles phrases it. The migration pressure was compounded with the end of support for Windows 10, which triggered a rush of new purchases at the time.

Then there's the components crisis. Memory prices have surged due to the AI infrastructure buildout, which has carried seemingly permanent ramifications for the cost of devices and the memory powering those systems. Low-budget devices have been a major casualty, amid a general decline in PC shipments and a projected 20% year-on-year fall in sales, according to IDC analysis. Interestingly, PC manufacturers are struggling to maintain full product portfolios, the report highlighted.

Then there's the question of AI PCs, divided between the initial marketing push and newer systems that boast actually useful on-device capabilities. The expectation in the latter camp is that power users want to tap into the most powerful systems, like AI agents, without necessarily requiring a constant connection to a cloud server to churn through expensive and scarce tokens. Ole Bülow, director of business development at Egiss, poses the question: "Which capabilities will employees actually use during the device’s lifecycle?"

"This requires understanding applications and processing locations, rather than using an AI label as a universal refresh trigger," adds Bülow, whose experience lies in advising enterprises, including one instance supporting the transition of a 120,000-strong global enterprise from a three-year device refresh cycle to a four-year cycle.

How is the refresh cycle changing?

While experts can't agree on the 'death' of the conventional cycle, they all agree that it's both shortening and lengthening at the same time, depending on a few different factors, including the technology itself and the user.

"A hybrid lifecycle, supported by signals such as DEX, is recommended," Bülow continues. "A supported device used for predictable tasks may remain productive longer. Employees with increasing demands may require earlier replacement. The original device can then be redeployed to another user after secure recovery and preparation.

He continues: "An employee may require an upgrade before the device reaches end of life. Mini PCs or thin clients should not be assumed to be underpowered. Suitability depends on the specific work requirements. For thin clients, the business case should consider the remote computing environment, connectivity, licensing, support, and endpoint price.

That's certainly true across the equipment landscape that organizations of any size may conventionally seek to upgrade, from business smartphones to on-premises servers and everything in between. The challenge, as Rosholt sees it, is that much of this is now ad hoc rather than following a single schedule.

"IT leaders are looking more closely at how equipment is being used and whether an upgrade delivers a meaningful benefit," he explains.

"Employees with heavier workloads or a clear need for AI capabilities may receive newer machines sooner, while regular users can often keep their existing devices longer. Refurbished equipment is also becoming a more important part of the equation, particularly for users who don't require the latest hardware."

How can businesses navigate being pulled in multiple directions?

Greater flexibility seems like a broad solution for navigating this new, complex, mixed picture. Bowles says his company tiers hardware by role instead of buying the same laptops for everyone. That includes cheaper thin clients or refurbished units for lighter roles and newer machines for finance or development teams.

Bülow also seeks to understand how an employee is likely to use the device based on data and what the existing device estate can support in terms of functionality.

"Combine performance and employee experience data with repair history, battery condition, warranty and security requirements," he advises, adding: "Determine whether poor performance requires a new device, a repair or a software adjustment. Define clear criteria for retaining, repairing, redeploying or replacing equipment. Assign responsibility for reviewing exceptions."

How manufacturers are responding to the new market conditions may also inform what businesses do. The OEMs are pushing to replace PCs with newer models, but enterprises are being far more disciplined about how much they spend due to the components crisis.

One of the fixes they're employing is actually trimming configurations of memory to, at the very least, maintain price levels. It may well be, Rosholt says, that organizations now choose to take them up on the offer of financing options and leasing.

Long gone are the days of fixed refresh cycles for every device across an organization. Many factors are temporarily affecting the decisions that businesses make on a day-to-day basis, including price volatility and Windows support. But still, the increased complexity across the suite of enterprise hardware is also a sign of technological progress and diversity in the systems that end users will rely on as AI becomes more useful and prevalent.

Huawei and Qualcomm have announced a multi-year patent license deal aimed at supporting research in areas such as 5G, AI, and networking.

The deal between the duo will include cross licensing for both companies’ patent portfolios, which span a range of technology domains. It also marks the first 5G patent licensing agreement between the two firms.

In addition, Qualcomm has agreed to purchase Huawei US patents relating to areas such as AI and networking. This reportedly includes licensing for the Chinese company’s LogicFolding chipmaking technique.

While the deal remains subject to regulatory approval, Huawei said the agreement highlights a shared commitment to IP rights and “licensing practices consistent with fair, reasonable, and non-discriminatory principles”.

"Huawei's decades of sustained investment in fundamental R&D have driven innovation and progress in mobile communications and other technology fields,” said Huawei’s chief intellectual property officer, Alan Fan.

“Huawei's broad contributions to the 4G/5G standards, such as the near-physical-limit signal transmission technology using polar codes, have established Huawei's leadership in the mobile communications industry, and continuously transformed the way people communicate and live.”

Huawei made its first IP license payment to Qualcomm in 2001, the company noted, while its first IP license income came from Motorola in 2011. Since then, the total cumulative revenue from IP licensing has reached $6.9 billion.

To date, the company has signed upwards of 270 patent licensing agreements with companies in the United States, Japan, South Korea, and Europe.

John Han, EVP and GM of Qualcomm Technology Licensing, said the move reflects a recognition of Huawei’s “continued innovation and intellectual property in 5G and other technology fields”.

"Qualcomm has invested in foundational wireless technologies that have enabled successive generations of mobile innovation and earned broad recognition across the global wireless industry,” Han commented.

The announcement by the duo builds on a similar licensing agreement signed between Huawei and HP in August this year. At the time, the Chinese tech giant said it agreed to a multi-year deal for Wi-Fi patents.

FOLLOW US ON SOCIAL MEDIA

AWS CEO Matt Garman has published a lengthy blog post suggesting that protests against data centers are misguided, and often prompted by hostile nations.

According to Garman, the US economy, national security, and ordinary people stand to benefit immensely from AI, and thus from data centers.

With more than 100 data center moratoriums being considered across the country, Garman said making the wrong choices could put the country behind, potentially irreparably.

"With any change, there will be important questions raised, but there will also be misinformation and outright lies, and in the age of social media and 24/7 news, myths take hold faster than ever before," he wrote.

"In fact, this build out is so important geopolitically that there are widespread reports of various countries intentionally seeding misinformation in the US about data centers to trick us into slowing down."

As a sweetener to encourage more communities to approve data center plans, Garman announced plans to allocate $1 billion over the next five years under a program called Built Together.

How it's spent would depend on the communities themselves, but the money would likely go towards education, job training, energy affordability, and water and energy conservation.

Like Microsoft, Garman noted that the hyperscaler has also stopped using nondisclosure agreements (NDAs) when seeking approval for new data centers from government agencies, another long-standing source of contention.

Instead, he said the company is providing information about its plans as early as possible and listening to concerns.

Garman shrugs off data center concerns

Much of the post goes to trying to counter those concerns. Data centers actually use very little water relative to other industries, he said, with US golf courses alone using around 200 times as much water as all of Amazon’s data centers put together.

Nor are they driving up energy prices or emitting huge amounts of pollution via their large backup diesel generators, he claimed.

Data centers also offer significant advantages for local communities, according to Garman.

"Data centers bring considerable tax revenue, positive economic impact, large numbers of good jobs, education programs and other benefits to the communities where they’re built," he said.

"Additionally, thousands of construction jobs are created to build data centers. These jobs benefit families in these communities and are a boon for local small businesses,” Garman added.

“The average wages paid far exceed local and national medians, and the job numbers associated with these projects are typically very significant compared to the size of these communities."

AWS has a hard battle to fight, however, with surveys showing that the US public is extremely wary about the proliferation of data centers. In a poll last month by the University of Massachusetts Amherst, just 11% of Americans said they would support the construction of an AI data center in their local community, with 65% against.

Garman's comments on job creation also run counter to recent research in the UK, which showed that only a quarter of the roles predicted in upcoming projects will actually materialize.

Analysis from environmental think tank Verdant noted that data center projects will only create 10,400 workers. That marks a stark contrast to figures touted by TechUK, which predicts around 40,000 overall.

“The remarkable thing is that we can’t find a constituency that actually wants these data centers,” said Raymond La Raja, professor of political science at UMass Amherst and co-director of the poll.

"The politics of AI may be somewhat polarized, but resistance to putting its infrastructure next door crosses the partisan divide.”

FOLLOW US ON SOCIAL MEDIA

AI is going to have to earn its keep by inspiring $4 trillion of as yet unknown applications and services, with annual spending on AI infrastructure reaching $1.5 trillion in the next five years.

That's according to Bain & Company's Technology Report, which notes that the size and cost of AI data centers is doubling every 12 to 16 months.

Last year, Meta's 600MW Prometheus cost about $24 billion to build, and data centers currently in the works at 1-2GW will cost between $40 billion and $80 billion to build — climbing to $200 billion for a 9GW data center in 2030, the analyst firm predicted.

That suggests annual spending on AI infrastructure will top $1.5 trillion by 2031, Bain predicted, including new data centers and compute, as well as upgrades to existing GPUs, memory, and networking equipment.

Doubling costs

That's double what's expected to be spent this year, with Bain predicting $780 billion in capital expenditure across the five key hyperscalers, though that does include investments beyond AI. IDC predicts AI infrastructure spending of $497 billion this year, while Gartner puts it at more than $1 trillion, including a wider range of equipment.

What they do agree on is significant growth: Bain has the figure doubling to $1.5 trillion by 2031, IDC predicts $1.2 trillion by 2030, and Garter sees total AI spending doubling in just two years.

That raises a challenge: can AI earn enough to cover these costs? Bain argues yes – but with some caveats.

"The infrastructure is being built ahead of the demand curve, and funding it sustainably will require adding approximately 1% to the annual global GDP growth rate," Bain's analysts say. "The question is whether the applications arrive in time to pay for it."

Paying for AI

First, the analysts at Bain argue that CapEx for cloud providers is usually about a quarter of industry revenue, meaning AI will need to earn $6 trillion in revenue by 2031 to maintain that margin.

By 2031, consumer AI products could contribute between $200 billion to $400 billion by 2031 via subscriptions and advertising. "Enterprise adoption could contribute another $1 trillion to $1.4 trillion in gains to providers alone as AI delivers meaningful productivity gains to enterprises across software development, sales, marketing, customer service, and IT operations," Bain's analyst predict.

That leaves about $4.2 trillion of revenue missing. Bain figures this will come from four new categories, starting with search and advertising at $100 billion to $200 billion, with ads integrated into chatbots as traditional online search is disrupted.

Missing $4 trillion

That leaves $4 trillion. Bain says $400 billion could come from autonomous operations of cars, drones, and robotaxis, as well as logistics, while physical AI could add $900 billion via simulations, digital twins, and the deployment of robotics – including humanoid ones.

That leaves another $2.7 trillion to be made up in new product development, anything from drug discovery to materials sciences, Bain suggests.

The point isn't that AI is expensive, but that companies need to do more with it than merely shave a few percentage points from their bottom lines.

"The economics required to generate ROI from AI infrastructure are demanding trillions in new revenue, not just cost savings," Bain said. "The industry needs a wave of application innovation comparable with what mobile and cloud unlocked, not just productivity gains on existing workflows."

Cannibalising tech

While the tech industry is seeking revenue for AI, that same technology is eating away at margins in other departments.

Bain noted that software revenue growth has "moderated significantly" from 20% annually in 2022 to half that last year. On the flip side, Bain predicted that agentic AI could prove to be a $100 billion opportunity for SaaS providers.

Beyond costs and revenue, Bain also warned in its technology report that the $5 trillion data center rollout still faces constraints from power, chips, skilled labor, and permits.

"Individual projects are finding workarounds, but they can’t be replicated at scale, and some are controversial. Colossus (from xAI, now part of SpaceXAI) was built in under a year by repurposing a factory and leasing dozens of gas turbines without federal clean air permits, drawing a lawsuit," Bain said.

The analysts added: "Meeting the global ambition instead will require system-level fixes outside the usual free-market playbook. A grand challenge on this scale demands grand solutions, and executives planning on business as usual will get blindsided."

Proposed capital expenditure from big tech companies into AI infrastructure has been astronomical. The likes of Meta, Google, and Microsoft, among other hyperscalers, are leading the charge on emptying their coffers to spend on data centers not just in the United States, but across the world. Among the beneficiaries of this spending are neoclouds, gobbling this cash up as part of their ongoing infrastructure buildout, alongside borrowing and share issuances.

But all this infrastructure spending is not simply numbers on a screen or figures to trawl through on an Excel spreadsheet. It's required spending, experts say, for the AI industry to sufficiently build the systems we need in the future – whether that might be smarter models, agentic clusters or simply to provide more capacity as more and more users start plugging into a new wave of AI-enabled productivity tools. And, at the other end of the dollars spent, are people who live in areas affected massively by the presence of these ever-humming, environmentally-questionable monuments of silicon and steel.

That's why, in many parts of the US and elsewhere, severe public backlash has spilled into regulatory action – which ranges from more consultation with local communities to a moratorium on construction. Is there anything that businesses can do to move past this movement?

Explaining the big AI backlash

Amid a proposed piece of federal legislation that seeks a national moratorium on building or upgrading AI data centers until safety is mandated by law, government at the state level has issued several banning orders against new data center construction. This includes New York State, Texas, Maryland, and Pennsylvania, among others. So where does this growing sentiment against AI come from, realistically? Much of it has to do with the impact on power and the environment.

"There is a trend right now of going from local bans and moratoriums to statewide ones," explains Arif Gasilov, partner at energy and environmental consultancy Gasilov Group. "New York enacted the first statewide moratorium in July, and in just one week in September, the governors of New Hampshire, Connecticut, and Oregon all announced pauses or refusals.

"In addition, moratoriums that are countywide still copy each other; more than half of Maryland's 24 counties now have one. In addition, the costs of data centers in local counties, such as electricity bill increases and whatnot, show up for everyone, while tax revenue shows up only to the local government and jobs only to the local skilled electricians, and so on. Overall, as such, the overall negative sentiment gets spread across more people."

This is a sentiment that Nicholas Nadeau, co-founder and CTO of Onix, echoes, telling ITPro that "most of the concerns are real".

"Underneath the noise and water complaints is an ownership question. People already paid for this industry in three currencies: their work that became training data, their wages that never moved with the productivity gains, and their savings that index funds route in at valuations insiders set. Data centers add a fourth, and it is the visible one: their grid, their water, their land."

The rising cost of electricity is the biggest factor that's impacting public mood, as well as the impact on water usage in arid areas – with evaporative cooling using the same supply that residents are under restrictions for. For example, a Meta facility in Newton County, Georgia, uses 10% of the entire county's supply. When you live nearby, there's a need to rely on diesel-based backup generators, traffic due to the construction, and a reduction in the local housing supply too, among other knock-on effects.

Dan Ye, CEO of CollegeNode and UC Merced lecturer (and former Johns Hopkins University lecturer) in AI, blames some, but not all, tech companies. "A handful of bad actors have done real damage here – xAI's Colossus data center in Memphis, Tennessee, is the most cited example, generating significant noise and discharge complaints – and the media coverage that follows fuels fear among residents elsewhere who are simply weighing whether to welcome a data center into their neighborhood."

He adds that a data center done right has "minimal impact" on neighbors. "These are engineering and siting choices, not inherent features of the technology – which is why it's worth being careful not to throw the baby out with the bathwater," he explains. "Data centers are also a significant source of local revenue."

What does this movement mean for AI progress?

Frontier training could slow down significantly if there is a sustained shortage of computing capacity, Ye says, while limiting how AI services are deployed and keeping usage costs higher than they'd be with more available infrastructure. But temporary pauses may not make much of a difference at all in Nadeau's eyes.

"Arvind Narayanan ran the arithmetic on Hard Fork podcast last month and found a one-year state moratorium delays AI progress by hours, not months, because efficiency at a given capability level improves several times a year while new capacity adds roughly a quarter more, and blocked sites get rebuilt elsewhere anyway," he says.

There's also a distinction to make between calls for a slowdown in the infrastructure and calls for a slowdown on the safety and security front. This is something that Ye characterizes as two pressures pulling in different directions. If political pressure persists, he says, tech companies will push data centers offshore to places like Canada, given its abundant hydro power, or even China, which is enduring a renewables boom presently.

The pausing for safety rhetoric, meanwhile, is just cover, according to Gasilov, "because it made them look responsible while the real delay, as per the data, is grid queues". Indeed, there's a massive backlog when it comes to the power generation and infrastructure, with many sites not expected to go live until 2027 or 2028. These sites will largely build their own behind-the-meter gas generation plants to power them, sidestepping both grid connection delays and protests, but may result in much higher rates of air pollution and premature deaths, he warns.

Can the industry do anything to ease concerns and get back on track?

Almost half of the states in the US have approved large load tariffs, which make data centers pay their own grid costs for upgrades to the grid – and more than ten states have filed moratorium bills. Gasilov says that the opposition argument wins out at county planning meetings, but won't necessarily solve the problems with bills, for instance.

Companies in this space can do several things to ease concerns and smooth the resumption or acceleration of this massive infrastructure buildout, the experts tell ITPro. Ye persuades site developers to bake community acceptance and deliverable power into the investment case from day one and not treat this as an afterthought. As he puts it: "Evaluate sites against actual grid capacity, water availability, residential proximity, and how much development is already planned nearby. A cheap parcel can turn very expensive once you factor in years of interconnection queues and approval delays."

As for businesses that buy AI services, he says a reasonable practical response is to understand a supplier's capacity constraints and lock in a service commitment that matches up with your organization's particular needs. He advises that critical workloads should remain portable, if possible, and they should invest in smaller models that can run either on-site or even on-device rather than chucking everything into the cloud. Nadeau puts it more bluntly, saying: "Stop naively buying compute by the token." He adds that businesses should measure cost and energy per completed task, then size the model to the job, while pushing AI inference as close to the edge as possible – cutting down on latency and compute costs while also limiting privacy liability in the process.

Ultimately, however, the buck stops with the infrastructure developers and city planners. Microsoft, for example, set a very useful benchmark with its commitment in January 2026 to cover power costs and water use, Ye says, adding: "The strongest reassurance is a project residents can actually hold accountable. If the economics only work by shifting infrastructure costs or environmental burdens onto neighbors, the business model itself needs to change."

NetApp has unveiled a file system it claims is the “fastest storage on the planet” in a bid to tackle AI factory bottlenecks.

Dubbed ‘NetApp Novus’, the new system is designed primarily for organizations running AI factories, such as Neoclouds and GPU as a service providers. According to the firm, the launch comes in direct response to long-standing storage-related constraints, sparked by rampant AI advances.

Speaking to assembled press at NetApp Insight, Kurian noted that traditional storage architectures were never designed to contend with the huge compute demands of AI factories.

“We see clients in the largest AI factories or the enterprise struggle with how to work with the modern data and compute infrastructures,” he said.

“Most GPUs are not fed data at the rates that they need to be fed. They have a high degree of parallelism and concurrency, which makes traditional storage architectures not capable of addressing those requirements.”

Put simply, traditional storage architectures can’t keep pace with the scale and performance requirements of AI factories.

Bottlenecks, increased infrastructure costs, and reduced GPU utilization are all a direct consequence, and that has a downstream impact on broader AI innovation.

Figures touted by the company at its annual Insight conference this year show that, under traditional architectures, GPU utilization can drop below 30% when AI factories can’t feed enough data to GPUs.

With Novus, NetApp is tackling this challenge head on. The system will support zettabyte scale and is built to exceed throughput of 100Tbps, allowing AI factories to continuously feed thousands of GPUs at any given time.

Kurian went to great lengths to emphasize the sheer scale and performance afforded by Novus during the opening keynote.

“Imagine if you could unify all of the world’s data center capacity that shipped last year - 2,000 exabytes of data into a single namespace, the first of its kind,” he said.

“The first zettabyte-scale file system on the planet. Let that sink in”

Capitalizing on AI factories

The move by NetApp comes as the scale of AI factories skyrockets. While previous clusters typically housed around 10,000 GPUs, new builds on the horizon are targeting 50,000 and upwards.

In October last year, for example, Nvidia and Samsung unveiled plans for a new AI factory with 50,000 Nvidia GPUs. NetApp's own figures noted that a cluster of this size would require 100Tbps of cumulative bandwidth.

With this in mind, NetApp appears confident that it’s capitalizing on a highly lucrative market.

Recent analysis from Deloitte projects adoption to accelerate across a range of industries. In the energy sector, for example, adoption rates are expected to reach 82% by 2028, marking a significant increase compared to 2025 levels.

Elsewhere, financial services is another key growth area with regard to AI factory adoption, surging from just 24% in 2025 to 63% by 2028.

FOLLOW US ON SOCIAL MEDIA

For more than a decade, fiber-to-the-home (FTTH) has been defined by a single contest - speed. Each new generation of Passive Optical Network (PON) technology arrived promising more bandwidth and more headroom, and operators competed accordingly.

But that contest is quietly losing its meaning. There is no obvious "killer app" driving exponential demand, and even bandwidth-hungry use cases - 16K video among them - are increasingly absorbed by better compression.

Consumption is still rising, but the rate of growth has slowed sharply. Remote work, 4K streaming and cloud gaming are now mainstream, yet widely deployed 10 Gigabit Symmetrical Passive Optical Network (XGS-PON) networks already carry far more capacity than most households will ever use. Average peak usage sits at around 5 Mbps, spiking only when a major title like Fortnite or Call of Duty pushes out a large update. Raw speed, in other words, no longer separates one network from another.

That shift is changing the question operators ask. It is no longer "how fast is your fiber?" but "how effectively can you share it?" - and for the channel, that change opens a market that used to be closed.

The limits of competing on speed

The old model of differentiating through ever-higher tiers runs into three walls.

The first is diminishing returns. Once a customer reaches 300–500 Mbps, the jump to 1 or 2 Gbps rarely changes their day-to-day experience, so speed upgrades deliver little retail payoff.

The second is rising deployment costs. Building parallel fiber networks is capital-intensive and duplicative. Overbuilding may satisfy regulators, but it rarely improves how well fiber is actually used.

The third is opportunity cost. Operators locked into their own infrastructure spend their time and budget on civil works rather than on the cloud, content, smart home and enterprise services that customers increasingly value.

With these pressures mounting, the industry is looking for a more sustainable way forward.

What FANS changes

Fixed Access Network Sharing (FANS) offers a way out. It lets multiple operators, wholesale and retail, share the same physical fiber access network. At the same time, each keeps full control of its own service logic, quality of service, traffic management, and customer experience. Retail providers operate over virtualized, isolated slices of shared infrastructure rather than building parallel networks of their own. Wholesale providers, meanwhile, monetize the same fiber many times over, sharply improving how hard each asset works.

Three things make this viable now. Bandwidth has matured, moving the value from throughput to availability, reliability, and quality of experience. Virtualization and software-defined access have reached the scale needed to support programmable sharing that was impractical a decade ago. And, finally, regulators in many markets now actively favor open access to widen coverage and cut duplication.

Why the channel should care

The most interesting part of FANS is who it lets into the market. A model that echoes the way mobile virtual network operators (MVNOs) reshaped mobile - but with far more flexibility - means utilities, mobile-only operators, niche ISPs and other brands can enter FTTH without laying a single strand of fiber.

That lowers the barrier to entry dramatically. It also lowers cost per subscriber, which makes business cases stand up in rural and hard-to-reach areas where standalone builds would never pay back. New entrants get a route to market; wholesalers get new tenants.

It also changes what a wholesaler sells. Instead of dark fiber or basic bitstream, a wholesale provider can offer a service-ready platform on which partners launch differentiated services almost immediately. That is a more valuable proposition than per-line pricing, and it reframes the wholesaler as a platform operator rather than a pipe.

Innovating without waiting for the next PON cycle

Because the access layer is shared and programmable, innovation can happen at the service layer rather than through expensive hardware refreshes. Providers can introduce smart home bundles, home security, SME connectivity, or edge-cloud experiences without waiting for 50G or 100G deployment cycles. The retail experience can evolve even while the underlying access network stays stable, which decouples competitive differentiation from the PON roadmap.

The payoff is better economics across the board: higher return on every home passed, lower operational complexity as each operator manages its own slice independently, and faster time to market for new services.

A reset for the fiber era

None of this happens on its own. Multi-tenant operation at scale depends on a full stack - operational support systems (OSS), software-defined networking (SDN)- based management and control, virtualized broadband network gateway (BNG), and multitenant-capable optical line terminals - working together to deliver true wholesale flexibility. But the strategic direction is clear.

The old playbook - compete on speed, build parallel networks, chase the next PON generation - is giving way to something more collaborative and more efficient. Treating fiber as a shared, service-ready platform rather than a siloed asset lets operators grow faster with less capital, reach more customers, and compete on service instead of infrastructure.

Beyond speed lies shared value. For the channel, that is where the next phase of FTTH opportunity is being built.

Akamai Cloud and Anthropic have announced a partnership extension that will see the AI developer commit to $11.6 billion in compute spending over the next seven years.

According to Akamai, the multi-year commitment will enable Anthropic to draw upon its distributed AI infrastructure to meet “accelerating CPU workload demands”.

"Anthropic is advancing the AI revolution, and we are thrilled they chose Akamai’s capabilities for building and operating AI infrastructure at scale,” said Akamai co-founder and CEO Tom Leighton.

“Our expanding global footprint, combined with our years of experience serving the world’s largest enterprises, positions us to be the infrastructure provider for secure and responsible AI applications and workloads.”

The announcement builds on a previous deal struck between the two firms earlier this year and highlights a “growing demand” for infrastructure services, according to Akamai.

At the time, the firm said this would include more than $2.8 billion in Cloud Infrastructure Services (CIS) commitments.

As part of the move, Akamai has issued a warrant for Anthropic for the purchase of non-voting convertible Series B stock. In a statement, Akamai Cloud said it projects around $5.5 billion in capital expenditures related to the agreement.

Rapid infrastructure investment

The move marks the latest in a string of infrastructure-related deals for Anthropic over the last two years.

With the firm ramping up development timelines and expanding its customer base, Anthropic has penned a number of multi-billion-dollar deals for compute capacity.

In May, for example, the company announced a deal with SpaceX to use all of the compute capacity at the Colossus 1 data center in Tennessee, hosted by xAI.

This gave the company access to more than 300 megawatts of new compute capacity, as well as roughly 220,000 Nvidia GPUs.

That deal came hot on the heels of a similar partnership expansion with Broadcom and Google to use “multiple gigawatts” of the latter’s tensor processing unit (TPU) chips.

In February, Anthropic raised $30 billion as part of a Series G funding round to support frontier model development and infrastructure expansion.

Led by GIC and Coatue, the funding round brought the AI developer’s valuation to $380 billion at the time, marking a near-doubling in the space of a year.

FOLLOW US ON SOCIAL MEDIA

Huawei is, I was told last week at Connect 2026, a company of engineers run by engineers. Many of the top executives in the company have come up from an engineering background and it feels apparent in the highly technical announcements made at the conference.

Beneath the technological developments and breakthroughs, though, is a portfolio that many enterprise vendors could only dream of. Chips, networking, servers, and storage, Huawei has a finger in every pie (and that’s before we get anywhere near devices and, yes, cars).

Landing a whale of a client with DeepSeek

Huawei pulled off what, at first blush, appears to be something of a coup when it comes to big customers this month. DeepSeek, the open weight model that smashed onto the generative AI scene in 2025, announced it was swapping Nvidia AI chips for Huawei Ascend in its infrastructure.

There’s no doubt that this is a big deal – DeepSeek changed how people understood generative AI models work and how much they cost to run – but there’s another aspect at play.

Due to the US’s ongoing trade restrictions, there’s no fresh supply of advanced chips from Nvidia or AMD going into China, so DeepSeek had no choice but to turn to another supplier. With the US players off the board, the alternatives become quite limited.

Nevertheless, this offers an opportunity for Huawei to show what its data centre hardware can do. It may not find a market for Ascend, OceanStor, Atlas and the rest in the USA but there’s a big wide world out there full of markets that will be interested in a proven alternative.

In fact, there were many customer spokespeople from Asia and Europe speaking at the breakout sessions held throughout the two day conference in Shanghai, including French supermarket Auchan and Turkish investment bank İşbank Türkiye.

The battle of the SuperPods

Without wishing to drive the Nvidia comparisons too hard, it’s worth also looking at Huawei’s pride-of-place AI infrastructure, the SuperPod.

SuperPods are extremely large infrastructure platforms made up of multiple components acting together as a single system. They are designed for demanding AI and high performance computing (HPC) workloads.

When you enter SuperPod into your search engine of choice, you will find one other name making this type of technology and that’s Nvidia. This hasn’t gone unnoticed and many commenters have suggested that Huawei is gunning for Nvidia’s AI throne, both with SuperPods and its latest upcoming Ascend 970 and 980 chips.

Even Nvidia’s CEO, Jensen Huang, has remarked on Huawei’s growing presence in the AI market, saying in 2025 that Huawei’s “presence in AI is growing every single year”. He added that enterprise tech companies elsewhere shouldn’t assume the company won’t be a factor in the AI race.

Huawei isn’t going to take over the world overnight. I noticed that it wasn’t named in the technical specifications of LineShine, for example, and trade restrictions in the US keep it out of a large market. Once you get through the dense technical detail, though, what’s underneath is a capable company that has a real breadth of expertise, deep technical knowledge and a huge commitment to R&D.

Alibaba Cloud is planning a big infrastructure push, hoping to meet the increasing need for data sovereignty with scalable, locally hosted AI compute.

The Chinese cloud giant is creating its first cloud regions in Türkiye, Finland, and the Netherlands, and is increasing its data center footprint in Malaysia, Germany, the United Arab Emirates, France, and Hong Kong.

By 2032, the company said it's aiming for a total global data center capacity of more than 20GW.

The announcements follow a commitment last year to invest around $52 billion in cloud computing and AI infrastructure over three years.

The expansion, announced at the 2026 Apsara Conference, will boost the firm's ability to provide localized cloud and AI services to customers across key markets.

Alibaba Cloud currently operates 107 availability zones in 31 regions around the world.

“As Alibaba Cloud has advanced from AI-native infrastructure to an Agent-Native Cloud architecture, AI is also moving rapidly from experimentation to real-world deployment, creating demand for both scalable infrastructure and more accessible AI capabilities,” said Dr Feifei Li, chief technology officer and president of international business of Alibaba Cloud Intelligence.

“We are expanding our global cloud footprint to bring computing resources closer to customers and partners, while continuing to evolve our AI technologies to help enterprises deploy models, optimize multi-model workloads and create new AI-powered applications."

Alibaba Cloud ramps up AI features

The company has also launched three new AI solutions to help businesses build and scale production-grade applications.

Smart Studio is designed to help businesses quickly build dedicated Model-as-a-Service (MaaS) platforms, with everything needed to package, price, meter, and bill model APIs under their own brand.

Boasting an “advanced inference architecture”, the firm said this will deliver up to five-times the throughput of standard open source frameworks on the same hardware.

Smart Fusion, meanwhile, is an intelligent engine designed to help users and businesses achieve state-of-the-art model performance at low cost through a unified standard model API.

This supports multi-model collaboration within one task, picking the best-suited available models to work together by analysing real-time dimensions. By abstracting multi-model deployment into a single endpoint, Alibaba said it can help reduce token spending by half.

New Qwen model and chips are coming

Elsewhere, the firm's next-generation model, Qwen 4, is currently in training, with the company also announcing its roadmap for the upcoming Qwen 4.5 and Qwen 5 model series, projected to scale up to 5 to 10 trillion parameters.

Alibaba Cloud also introduced the Zhenwu V900, its next-generation AI chip, which it said delivers three times the performance of its predecessor, the Zhenwu M890 released earlier this year.

The processor has 216GB of memory and 1,200GB/s of inter-chip bandwidth, with support for data formats including FP8 and FP4. It's scheduled for mass production and commercial release in the first quarter of next year.

FOLLOW US ON SOCIAL MEDIA

Huawei unveiled a flurry of enterprise hardware announcements at its annual Connect conference in Shanghai, China, underpinned by a steady current of agentic AI.

Speaking on stage at the Shanghai Exhibition Center during the opening keynote, Huawei rotating chairman David Wang unveiled the Atlas 960E SuperPoD and OceanStor M900 ranges.

According to Huawei, Atlas 960E SuperPoD is the industry's first near-packaged objects (NPO)-based SuperPoD, capable of massively accelerating training and inference for 10-trillion-parameter foundation models.

The star of the show, however, was UnifiedBus – a new interconnect architecture that the company claims will help reduce data bottlenecks in cluster computing.

Huawei pointed out that a 10 trillion-parameter model requires massive amounts of data for training and claimed that, under a traditional architecture, it significantly exceeds the capacity of a single accelerator.

Huawei targets agentic compute gains

Agents, which can have hundreds of interactions per hour with AI models, add to this problem. Interactions between CPUs and NPUs, Huawei said, generate KV cache that’s also greater than memory capacity, which in turn drives up demand for data storage across compute systems.

In his opening keynote Wang said: “The generation and consumption of intelligence will continue … growing by orders of magnitude, reaching infinite degrees of depth and breadth. These changes will place far higher demands on the scale, performance, and reliability of AI infrastructure. Only by building a powerful AI infrastructure can we lay a rock-solid foundation for this intelligent world.”

Wang explained the problem as it stands. “In a conventional server architecture, inter-server latency is an order of magnitude more than that of intra-server – namely 100s of nanoseconds vs 10s of microseconds,” he said. “Therefore, optimizing server architecture for communication overhead is a critical challenge for AI infrastructure.”

“Huawei’s SuperPoD design is equally connecting all components within a SuperPoD via a unified protocol, supporting unified memory access across physical servers and adopting a copper near optical four interconnect architecture,” he added.

“This enables near linear bandwidth and latency across tens-of-thousands of chips, allowing them to operate as one single computer and effectively boost the MFU. To build this complex system, the first thing you need is the interconnect protocol that equally connects all components within the cluster. This protocol is UnifiedBus.”

UnifiedBus is a new system architecture that, according to Huawei, solves these bottleneck issues with four key features:

• A unified protocol and memory semantics
• Heterogenous compute collaboration
• Tiered storage with global pooling
• Optoelectronic interconnect and flexible networking

A unified protocol

According to Yang Chaobin, CEO of Huawei’s ICT business unit, as UnifiedBus brings together ten protocols in one umbrella, round trip time (RTT) latency is reduced from seven milliseconds to two milliseconds, while interconnect bandwidth is increased from 100 Gbits/s to 1 Tbit/s.

Heterogenous compute collaboration

CPUs, NPUs, memory and storage are all directly interconnected under UnifiedBus, which Huawei said enables decentralized, peer-to-peer access between them. It also enables flexible NPU-CPU mixing.

Tiered hardware acceleration, meanwhile, enables Attention-FFN disaggregation.

Tiered storage with global pooling

UnifiedBus uses hybrid-media resource pooling, which enables double data rate memory (DDRM) to be used as an alternative memory source for NPUs.

According to Huawei, this reduces latency for services such as search and recommendation, and doubles the performance of vector retrieval for large, complex data loads.

It also, the company said, reduces per-NPU high bandwidth memory (HBM) capacity requirements for the training of 10 trillion-parameter models and improves cluster model FLOPs utilisation (MFU).

Optoelectronic interconnect and flexible networking

Finally, thanks to its networking features, UnifiedBus acts as a “global ultra-high bandwidth, ultra-low latency ‘data highway’” Huawei said, which the company claimed enables more flexible compute scaling.

“Huawei is committed to innovating at the system level, building a computing-foundation product portfolio, and going open source and open system,” said Yang, concluding his keynote.

“We will continue working with customers, partners, and developers to help this ecosystem thrive and offer the world a new option for compute."

FOLLOW US ON SOCIAL MEDIA

Sovereign cloud provider Civo has announced plans for 40 edge data centers in the UK, with the first to be sited in Hertfordshire.

The Hertfordshire site is currently being fitted out and will open next year at 8 MW, with plans to expand this to 38 MW.

As for the rest, the company has secured five further locations totalling 150 MW, with sites in Lancashire, Greater Manchester, London, and Oxfordshire, and the rest planned over the next three years.

With the rising use of AI agents inference workloads need to happen near the device making the request. Civo cited this trend as its justification for a distributed national network, rather than a single campus concentrated in one region.

Research from the firm earlier this year found that digital sovereignty is now viewed as a “strategic priority” by 73% of firms, but that only 15% have successfully migrated to a domestic provider.

"Britain has the talent. What it needs is infrastructure to match it. Policy alone will not deliver that. Compute, cooling and density have to exist physically, in this country, close to the organizations using them,” said Mark Boost, CEO of Civo.

"Our distributed edge network gives UK organizations sovereign, high-density infrastructure where the processing actually needs to happen, and gives international organisations a secure, compliant UK home for their services and data.”

Civo targets Vera Rubin gains

Civo will be among the first cloud providers in the UK to offer the Nvidia Vera Rubin platform, the successor to Blackwell, with up to five times the inference performance.

The Vera Rubin NVL72 system carries 288GB of HBM4 memory per GPU and brings together 72 Rubin GPUs and 36 Vera CPUs in a single rack.

Civo is taking orders for Vera Rubin now, alongside existing capacity on Nvidia B300, B200, H100 and H200 instances, with early-access pricing for Vera Rubin starting at $10.00 per hour.

All this runs on Civo’s AI stack, developed in-house to be open and interoperable, giving customers sovereignty over their data alongside UK compliance and security.

Notably, the sites are being built for direct-to-chip liquid cooling, with waste heat recovered and fed into district heating networks serving homes, factories and greenhouses local to each site.

FOLLOW US ON SOCIAL MEDIA

In September, the UK’s air traffic control system NATS suffered not one, but two outages.

In the first, a reported software glitch in NATS’ main systems caused the cancellation of over 2,000 flights. The defect happened “in the space of a millisecond”, according to an initial NATS investigation. This, in turn, caused corrupted data and forced NATS to limit UK air traffic for most of September 8th.

The second incident, a “connectivity issue” at its Prestwick centre in Scotland, was more minor, affecting some 150 flights. But at the time of writing, its full impact was still being assessed. The organization said air traffic control was back up to “full capacity” by the evening of September 21st.

Impactful beyond initial issue

Airlines, though, warned of knock-on effects. Ryanair said 140 of its flights were delayed, affecting 25,000 passengers. “Passengers are once again suffering lengthy delays because NATS has failed to deliver either a reliable system or an effective back-up,” said Neal McMahon, the airline’s chief operating officer, in a statement. Ryanair called on NATS CEO Martin Rolfe to resign, or for the government to sack him.

The failures are only the most recent to affect NATS. Back in 2023, the air traffic control service experienced another IT shutdown over the Bank Holiday weekend. Then, 700,000 passengers were delayed. A further, smaller incident forced the cancellation of 150 flights in 2025.

As far as we know, none of the incidents at NATS were caused by a cyber attack or other malicious action. Instead, it appears that all the incidents were the result of IT failures.

The organization’s initial report into the September 8th outage said that corrupted data forced NATS to restart the National Airspace System (NAS) and reload data. To do this, they had to restrict all of the UK’s airspace. The report also stated that the incident was not connected to the 2023 outage.

NATS claims it has invested more than £1bn in technology over the last 10 years. Airlines and their passengers will be questioning whether that was money well spent. But NATS's issues illustrate how fragile modern IT systems can be, and why organizations need to plan for failure.

“In its 2023 failure incident, NATS had a backup. It failed the same way as the primary, at the same moment, because it ran the same logic on the same input,” Ivan Milenkovic, vice president for risk technology EMEA at Qualys, a security vendor, told ITPro.

“After 2023, the CAA’s independent review produced 34 recommendations across NATS, the regulator, airports, airlines and government. The review was formally completed in June 2026 with all recommendations closed,” he explains. Less than three months later it failed again, in the same flight processing area. NATS says it was a different sub-system this time. Even so, I’d expect the new review to ask why “closed” recommendations didn't prevent another day like this.”

Far-reaching consequences

Clearly, air traffic control is critical to national infrastructure (CNI), and operators cannot take chances with safety. In the case of the NATS outages, the organization will have had no choice but to close airspace and to divert flights.

But disruption on this scale has consequences far beyond the inconvenience to individual passengers, or even to airlines and airport operators. They also cause a significant economic impact.

The 2023 NATS outage is calculated to have cost airlines £65m, and costs to passengers, airports, tour operators, insurers and others brought the total to £75m to £100m, according to the Civil Aviation Authority’s independent review. Understandably, governments want to see organizations, especially those running CNI, improve their resilience and recovery planning.

“It wasn’t a recent software update that caused the incident, rather a defect that had been present for some time, but the specific course of events hadn’t occurred to trigger it,” James Watts, managing director at Databarracks, told ITPro.

“You prevent this type of issue through testing software before rolling out into production. Once an incident occurs, the challenge is recovering systems, maintaining operations and minimising the impact. Fixing the defect comes later.”

As Watts says, once NATS found they had a problem, they moved quickly to implement their major incident plan. “The priority in transport and aviation must always first be to safety. NATS’ incident processes here look to be strong, and it operated according to its incident and crisis management planning,” he notes. But there are still lessons enterprises can learn.

The first is not about technology, but communications. NATS’ systems were offline for a relatively short time, but the impact was still significant. Even though NATS has come in for plenty of criticism, there are other organizations that have handled IT outages far less well. So CIOs, CISOs and heads of operations need to ensure their communications plans are as well-drilled as their business recovery processes.

More than just a tick-box exercise

Organizations also need to keep up with changing laws and regulations, even though compliance on its own does not necessarily improve resilience. Enterprises could face a growing regulatory burden while also having to redesign systems to be robust during an incident and easier to restore afterwards.

“Regulation cannot be a substitute for effective organizational and operational practices which seek to anticipate and simulate risks and take anticipatory steps to mitigate them,” James Morris, director of CSPR, a think-tank focused on business resilience. Organizations across critical infrastructure need to embed a “culture of prevention”, he argues.

In the near future, laws such as the EU’s Cyber Resilience Act, DORA and the UK’s Cyber Security and Resilience Bill will set minimum standards in some areas. But this might not be enough to make the economy better able to withstand accidental outages, let alone hostile acts.

“There is still a question over how exactly “critical national infrastructure” will be defined,” Chris Dimitriadis, chief global strategy officer at ISACA, told ITPro.

“Cyber attacks on major retailers such as M&S and Jaguar Land Rover have highlighted the wide-reaching impact of organizations that are not currently classed as critical national infrastructure. Regulation is moving in the right direction, but there is still a clear job for organizations to ensure that they are compliant.”

Businesses should review their business continuity and cyber incident plans now, and not wait for regulators to force their hands. This means reviewing skills, practices, and protocols, Dimitriadis says.

Otherwise, there could be far more than a missed flight at stake.

Dealing with shifting sands is par for the course for cybersecurity practitioners, but what's happening now is different. Indeed, it’s more of a sea change that upends conventional wisdom and business-as-usual approaches to safeguarding an organization. The impact of AI in cybersecurity changes not just the nature of the threats and attack modalities but also adds new challenges, such as dealing with massive spikes in volume.

These challenges are why many across the cybersecurity industry see value in building and propagating a pre-emptive cybersecurity architecture that deviates from a conventionally static and reactive approach. After all, getting ahead of any threat is better than letting it fester – responding only when the damage has taken place. Such an approach may seem like common sense, so why hasn't this notion become more widespread among practitioners?

What is pre-emptive security – and how does it differ from conventional cybersecurity models?

The traditional detect and respond cybersecurity framework has served businesses well for as long as they've tapped into cyberspace. But changes primarily driven by AI are exposing frailties in this model. For example, recent Crowdstrike research shows a plunge to just 29 minutes for the time it takes for an attacker to move from initial compromise to laterally breaking out to other systems. Garner, meanwhile, forecasts there will be a 300% spike in the number of documented vulnerabilities between 2025 and 2030. There's also the question of practitioner burnout, with conventional frameworks relying on overworked humans to investigate security alerts raised after, not before, an intrusion has been detected – and shut off a threat as a matter of urgency.

So what's the alternative? "In simple terms, pre-emptive security architecture means putting security controls directly in the path of an attack so that the attack gets blocked, diverted or contained before any serious damage has happened," explains Mudita Khurana, staff security engineer at Airbnb and cybersecurity expert with a decade of experience in application security, security tooling, vulnerability management and AI security.

"The idea is more architectural than a particular security product or control (thus the word architecture in the term itself)," she continues, adding: "It is a way of connecting security controls across different layers (like users, applications, systems, data, etc.) so that even if one layer is compromised, the blast radius remains limited because of other existing controls."

This is an environment in which attacks are far more likely to "fail on contact" or never really get a grip on systems, Steven Coppola, security engineer at BARR Advisory, tells ITPro. That is, instead of an architecture that relies on somebody noticing the attack while it's happening.

"I love analogies — for pre-emptive security architectures, I use a house," he says. "Detect and respond is akin to a good alarm and a fast call to the police department. A pre-emptive strategy is thinking ahead about where the doors will go, what they are made of, and whether the valuables need to be in the house at all."

What does pre-emptive security entail?

Framing pre-emptive security as a list of components or a single blueprint misses the mark, experts tell ITPro. Nor is this architecture a product that you can buy. Gartner's managing vice president Carl Manion distilled the trend into three elements, including 'deceive', 'disrupt', and 'deny' – and also warned product leaders that they may face losses in market share if they fail to embrace this security architecture in the services they provide to organizations.

His first thesis is to deny hackers entry with a preemptive shield, including advanced obfuscation technologies. Next, organizations should deceive bad actors using decoys, misdirection, and illusion to throw them off the scent. Finally, organizations should anticipate emerging threats and prepare to disrupt these rather than waiting for the attack to materialize and engaging in damage control after the fact.

"I believe that there are many other security controls that can support these three areas that Gartner has set out," adds Khurana.

"For example, zero trust access controls and separation between systems can limit the reach of an attack. Secure development checks can prevent vulnerable code from being deployed in the first place, and encryption can protect stored data so even if someone steals files, they can't read them without the encryption key."

She also noted that confidential computing – in which data is processed inside a separate hardware-protected area with only approved code running within it – can add another layer of defense.

Coppola adds: "Confidential computing closes the gap in the middle, which is data in use. The workload runs inside a hardware-protected enclave — essentially a walled-off section of the processor—so the data stays encrypted even while it is being processed. Even the cloud provider or someone with admin rights on the host cannot read it."

Even if the attacker reaches the underlying server or cloud infrastructure, there is no immediate access to the data, which should be protected. However, it must also be combined with proper access controls and sufficient application security.

How can businesses implement pre-emptive security?

Chris Bailey, head of innovation at systems integrator Jigsaw24, tells ITPro that any plans to implement pre-emptive security should be shaped around a clear idea of what your organization is protecting, how sensitive the data is, and the internal policies.

"Start by identifying the data and systems you’re trying to protect. From there, map out where the vulnerabilities lie, how people access the environment, and what the worst-case scenario would look like," he advises. "Once that is clear, it is important to establish clear policies around access, roles and devices. Ask who genuinely needs access to a particular system or data, what they need to be able to do, and whether that access can be restricted further."

At this stage, Khurana explains that it's important first to understand how an attacker may reach the services or data sources, a process similar to threat modeling, and then where the attacks can be denied, deceived, or disrupted. "This is where the team takes the threats and paths from the previous activity and understands where the controls can sit. If a path is found [to be] unnecessary, it should be removed. For legitimate paths, the principle of least privilege should be adopted."

After putting controls into place, organizations should test them under realistic conditions to make sure they work well, she adds. This may involve penetration testing – which Bailey adds is important because it gives businesses the chance to identify weaknesses before an attacker does – as well as continuous auditing and monitoring.

Organizations should also carry out attack exercises to see if the controls set in place actually stop the intended attack path, and that, importantly, things can be quickly rolled back if the attacks disrupt important business processes.

It's also worth considering escalating the approach. If pre-emptive security works for one service, teams may feel more confident in applying this to more critical services.

While implementing a new approach to safeguarding your organization, it's also important to remember not to simply ditch or replace old techniques. Coppola tells ITPro that pre-emptive controls can make alerts more meaningful – but don't replace them. Then, when it comes to measuring attack paths, he says: "The question at the end of the quarter should be: what can an attacker with stolen credentials reach now versus 90 days ago? If the answer isn’t shrinking, the program in place is not working, no matter how many products were deployed."

The Scottish government has agreed to tighten up the rules on new hyperscale data centers, but has stopped short of halting their development altogether.

MSPs voted down a motion brought forward by the Scottish Greens to introduce a complete moratorium, but said the government would only approve applications for new data centers with a capacity of over 50MW if they have undergone a full Environmental Impact Assessment (EIA).

The new rules come into force today.

"There are considerable benefits for all parties where the need for EIA is identified and planned for at the earliest possible stage, when there is greatest scope to integrate mitigation measures to minimize impacts from the outset in designing green data centers which support sustainable infrastructure growth," said chief planner Fiona Simpson in a letter announcing the decision.

"Embedding the consideration of environmental factors from the outset can also help minimize the risk of delay at later stages and can support a smoother passage of proposals through the planning process."

Meanwhile, until national planning guidance has been formalized – probably early next year – no decisions on planning for data center applications will be made.

The Scottish Greens had been calling for a full moratorium on new data centers, and MSP Patrick Harvie has expressed his disappointment at the decision.

"This debate was a chance for MSPs to reassure local communities and to put the national interest ahead of the interests of big business, but they have done the opposite," he said. "Instead they have chosen to muddy the waters and thrown up more questions than answers."

There are currently dozens of active planning applications for new data centers across the country, with decisions due for at least two by the end of this year. One, in Fife, is expected to be the second largest in the world.

The country's a desirable location, thanks to its leadership in renewable energy and the existence of former industrial sites with access to existing electricity infrastructure.

However, the Greens claimed that if just the first 24 of these active applications are approved, they would consume one-and-a-half times as much electricity as the whole of Scotland currently uses at peak times – between 4.5 GW and 6GW.

Meanwhile, current cooling techniques require large amounts of water, and people living near large data centers report high levels of noise.

But, said public finance minister Hannah Mary Goodlad, "We must balance the economic and employment interests in developing data centres with national energy, climate and community wealth building ambitions, which are vital to our future prosperity, as well as the potential impact on local communities."

The Green Party has similar ambitions to halt data center development across the whole of the UK. However, last week, Prime Minister Andy Burnham ruled out that idea, saying that data centers pulled in other investment and that AI Growth Zones were an important vehicle for economic growth.

Two years ago, the government classified data centers as critical national infrastructure, on the same footing as water, energy, and emergency services systems, and allowing some projects to be fast-tracked.

FOLLOW US ON SOCIAL MEDIA

Space-poor SMBs will love Qnap's TS-432XeU as it's one of the smallest rackmount NAS appliances around. Its standard-width 1U chassis measures a mere 292mm deep, so it'll slip easily into a small wall box or a standard two-post data cabinet

Stepping up as a lower-cost alternative to Qnap's TS-435XeU, it employs the same super-slim chassis but drops the two internal M.2 NVMe SSD slots and supplants the 2.2GHz quad-core Marvell CN913X CPU with a 1.7GHz quad-core Annapurna Alpine chip. Networking is slightly diminished too, as it sports the same dual 2.5GbE multi-Gig ports but halves the 10GbE SFP+ port count to one.

The appliance presents four hot-swap LFF/SFF bays at the front, and it's possible to expand capacity outside the box. The rear USB-A 3.2 Gen 1 ports can be used to add up to two of Qnap's tower or rack expansion units, and we suggest the TR-004U model as this 4-bay 1U rack shelf is the same depth as the appliance.

Qnap TS-432XeU review: Build quality

The TS-435XeU is well built with a solid steel chassis and sturdy metal drive carriers. The appliance includes 4GB of DDR4 SODIMM memory, which can be expanded to 16GB – half that of the TS-435XeU.

As you'd expect with such a small chassis, there isn't a lot of room inside, but the memory slot is located on the top of the motherboard and easily accessible. As with the TS-435XeU, there's only one slot, so any upgrades will require the base memory to be ditched.

Three cooling fans are arranged across the rear panel, with one drawing air over the CPU's heatsink via a small plastic shroud. During testing, we found the appliance to be whisper-quiet, making it a great choice for small offices that value their tranquility.

Qnap TS-432XeU review: Deployment and limits

For testing, we fitted four 4TB Seagate IronWolf NAS-specific HDDs and used Qnap's Qfinder utility to discover the appliance on our lab network. It provided quick access to the web browser setup wizard, where we left it to load the latest QTS 5.2 OS.

Note that Qnap's QuTS OS is off the menu, as this is only supported on Intel- and AMD-powered appliances and not ARM-based ones. QTS is a more nimble OS that doesn't have such a high demand for memory and is designed for entry-level NAS appliances.

QuTS provides superior data protection features such as ZFS copy-on-write, end-to-end checksums, and WORM (write once read many) policies for NAS shares. However, its data deduplication feature will only be enabled when you have at least 16GB of memory.

There are also software limitations, as Qnap's ARM-based appliances can't run high-profile apps such as the Virtualization Station and Hyper Data Protector. That said, we counted 79 apps available for this appliance, which include key data protection apps.

Qnap TS-432XeU review: QTS data protection

We used the Storage & Snapshots app to create a RAID5 array which also offers facilities for running manual and scheduled snapshots of NAS shares and iSCSI LUNs. QTS can manage up to 256 snapshots, with a limit of 64 per volume or LUN and Qnap's snapshot technology supports standard EXT4 file systems, so it doesn't require you to choose between this and BTRFS when creating volumes.

The licence-free Hybrid Backup Sync 3 (HBS 3) app is designed to protect your NAS data, helps you create a 3-2-1 backup strategy in only four steps, and uses Storage Spaces to define multiple backup destinations. These can be local repositories on the appliance, remote locations such another Qnap NAS and Rsync, FTP or CIFS/SMB servers, and cloud storage with Qnap supporting no less than forty providers, including its own myQNAPcloud service.

HBS 3 can run data integrity checks when backups have completed to ensure they can be restored, and file version management provides extra protection against ransomware. HBS 3 also offers one-way and two-way syncing services between NAS source folders and destinations such as remote Rsync, FTP and CIFS/SMB servers or any supported cloud provider.

QTS can boost performance for cloud storage as the HybridMount app uses standard file protocols to mount storage from providers such as Amazon S3, Dropbox, OneDrive and Google Drive as NAS shares and assign a local cache to them. You get perpetual licences for two providers included with extra ones costing £15 per year.

Alternatively, you can use the VJBOD Cloud app to present cloud storage as local block-based NAS or iSCSI volumes. One perpetual provider licence is included, with more also costing £15 per year.

Qnap TS-432XeU review: 2.5GbE and 10GbE performance

With the appliance hooked up over 2.5GbE to our Zyxel XMG1930-30HP multi-Gig switch, we mapped a share to a Dell Windows Server 2025 host on the 10GbE LAN. Performance is good with Iometer reporting raw sequential read and write speeds of 2.3Gbits/sec and 2.2Gbits/sec.

Real world speeds were in the same ball park, with 25GB file copies between the server and appliance averaging read and write speeds of 2.3Gbits/sec and 2.1Gbits/sec. Our backup copy test using a 22.4GB folder with 10,500 small files delivered 1.1Gbits/sec.

We moved to 10GbE by connecting the appliance using a DAC (direct access cable) to the switch's 10GbE SFP+ port. Running Iometer on the share again returned sequential read and write rates of 8.6Gbits/sec and 5.6Gbits/sec. Our real world large file copy delivered sustained read and write speeds of 3.8Gbits/sec and 2.9Gbits/sec while our backup test averaged 1.7Gbits/sec.

Over 2.5GbE, the TS-432XeU matches the TS-435XeU, but Iometer NAS read and write speeds over 10GbE are down by 4.6% and 22%, respectively. Even so, that's nothing to be sniffed at and will easily be enough for the target market of small businesses.

Qnap TS-432XeU review: Is it worth it?

The TS-432XeU is an affordable choice for small businesses with it listed on LambdaTek for £617 excluding VAT. The ARM CPU does have some limitations but Qnap's QTS OS delivers plenty of data protection apps, performance over 2.5GbE and 10GbE is quite respectable, and its compact dimensions make it a great choice where office space is at a premium.

Qnap TS-432XeU specifications

As AI initiatives scale, managed service providers are under pressure to help customers deploy new capabilities quickly while enhancing quality and controlling costs. But behind the scenes, engineers are running into a quiet bottleneck—power delivery at a reasonable cost.

Modern AI compute is increasingly limited by how quickly and efficiently engineers can deliver power to processors, accelerators and memory. For Managed Service Providers (MSPs), that makes power delivery more than an engineering issue. It becomes a strategic consideration that affects capacity planning, operating costs, cooling requirements and the ability to scale customer environments reliably. The challenge is not having enough watts—it’s getting those watts to the right place, with minimal latency, without excess heat, all while managing operational costs to keep revenue in check.

MSPs that overcome the AI power problem will be the ones that treat power not as an afterthought, but as a foundational pillar of performance, efficiency and scalability.

Why power density matters to MSP service delivery

A decade ago, a 200 W GPU was considered massive. Today, single AI accelerators regularly exceed 600 W, and full server boards can draw 5–10 kW.

That power needs to be delivered across a variety of distances, to multiple voltage domains, and respond in nanoseconds as workloads shift. Every bit of distance, impedance or delay creates loss, droop, and heat, hindering scalability and delivery service.

In short: as computing density grows, power density becomes the new productivity limiter.

This shift has also changed how engineers think about system-level design. Instead of optimizing only compute capability, architects must now consider the electrical and thermal realities that shape what is actually achievable.

High-performance silicon is only as good as the infrastructure that supports it, and many organizations are discovering that traditional power-delivery networks were never designed for bursty, AI-driven workloads. For customers and MSPs alike, these workloads demand not just more power, but more responsiveness and stability than previous generations ever required.

Leveraging power innovation to build customer trust

One emerging approach to this power delivery challenge is moving voltage regulation closer to—or even beneath—the silicon itself. Integrated voltage regulation (IVR) technologies are designed to deliver cleaner, faster, and more localized power that scales with chip demands.

For MSPs, these advances can support more predictable performance, better energy efficiency, and infrastructure designs that are easier to scale across customer deployments. Some of the positive outcomes associated with these advanced IVR solutions include:

• Higher current density compared with traditional discrete regulators, enabling denser AI infrastructure and serving modern AI workloads.
• Efficiency improvements by reducing long power-delivery paths and associated transmission losses.
• Low thermal resistance to dissipate power losses in a constrained environment, supporting more reliable performance and reduced cooling requirements.

These improvements are not incremental gains—they enable entire system architectures that were previously constrained by power delivery. This means a more profitable, effective, and efficient infrastructure that MSPs and customers can trust.

Building delivery-driven AI infrastructure

With IVRs and other localized power-delivery technologies, GPUs, CPUs, and custom AI accelerators can sustain higher operating frequencies under dynamic workloads by reducing impedance and minimizing voltage droop. This translates to:

• Data centers can deliver more compute per watt and reduce cooling overhead, improving the economics of managed AI services.
• Edge and distributed deployments can operate more efficiently, helping MSPs support AI use cases closer to customer environments.
• Compact, modular designs can simplify scaling and make it easier to standardize AI infrastructure across multiple customer deployments.

Instead of designing around power constraints, system architects can now design with power as an enabler of AI services.

Looking ahead

As more businesses and MSPs move towards AI to automate workflows and streamline services, power delivery for scaling AI must evolve too. The next wave of competitive advantage will not come only from smarter models or faster chips; it will also come from power architectures that help providers deliver performance, efficiency and control costs-- at scale.

For MSPs, the opportunity is clear: now is the time to reassess whether your AI infrastructure strategy can support the performance, efficiency and scalability your customers will expect next.

Providers that invest early in smarter power-delivery architectures will be better positioned to differentiate their services, protect margins and lead customers confidently into the next phase of AI adoption.

Google is set to expand its footprint in Finland as part of a €13 billion infrastructure investment in data centers focused on the towns of Hamina, Kajaani, Muhos and Vaala.

During the initial construction phase, which will run through 2027 and 2028, the expansion is expected to contribute €3.6 billion to Finland's GDP and support 37,000 jobs nationwide.

Once operational, Google said the facilities are projected to support 7,000 jobs annually, spanning on-site facility and technical roles, as well as roles in downstream businesses.

“This is Google’s largest single investment in Europe and a testament to Finland’s leadership in responsibly building AI infrastructure – demonstrating what is possible through strong partnerships to create long-term community value and world-class environmental sustainability,” the company said.

Running parallel to this, Google will also invest in energy infrastructure designed to support data center build-outs. T

his includes sa life extension power purchase agreement for the Loviisa nuclear power plant with Nordic energy provider Fortum, as well as upgrades to grid capacity and renewable energy.

Google said it will support two onshore wind projects, with development led by Valorem and Suomen Hyötytuuli. This will bring the total amount of onshore wind energy contracted by the firm to 629 MW.

Google builds on long-standing ties

Google already has a strong presence in Finland, having opened its Hamina data center 17 years ago. The location has since been transformed into a “digital data hub”, according to Google, and forms part of the company’s global network of 43 cloud regions.

Several organizations, including Nokia, Elisa, and the City of Hamina rely on the data center to support operations.

Google said the Hamina site and Finland more broadly serve as a “model for responsible data center development”.

Indeed, the Hamina data center’s liquid cooling system uses sea water sourced from the Bay of Finland. It also has an offsite heat recovery facility, which is used to warm nearby households and businesses.

Finland in the spotlight

Finland has rapidly grown to become one of Europe’s leading data center infrastructure hubs. Research from Savills in August found the Finnish data center market has grown by 57% since 2024, making it one of the fastest-growing regions both in Europe and globally.

Google isn’t the only tech giant capitalizing on the Finnish market boom. A host of companies including Microsoft and TikTok have made significant investment pledges.

Microsoft, for example, signed a preliminary agreement to acquire 190 hectares of land in Vaasa and Mustasaari in June this year for data center development. Microsoft is also developing its first data center region in Southern Finland.

A key factor behind Microsoft’s expansion lies in the region’s existing industrial capabilities, Microsoft said at the time. The land it acquired is located in the GigaVaasa industrial zone.

TikTok, meanwhile, announced plans last year to invest €1 billion in its Kouvola data center, which is expected to have a capacity of 150MW.

FOLLOW US ON SOCIAL MEDIA

Qualcomm has signed a major deal with Amazon, under which it will supply customized chips and networking hardware for Amazon data centers.

The 'multi-generation' collaboration covers customized silicon for Amazon's large-scale AI data centers, bringing together the hyperscaler’s AI infrastructure and Qualcomm's power-efficient processing, silicon design, and system-level integration.

The deal will also cover high-performance optical connectivity solutions specifically designed to support the growing scale and bandwidth demands of AI infrastructure, with speeds extending up to 1.6T and eventually beyond.

This work will be based on Qualcomm Technologies’ SerDes and optical DSP technologies.

"As AI demand accelerates, data center infrastructure will require advances in both computing and connectivity to deliver greater performance with more efficiency,” said Cristiano Amon, president and CEO of Qualcomm.

“Qualcomm is pleased to work with AWS on customized silicon and connectivity solutions, bringing decades of leadership in advanced processing and power-efficient compute, to deliver breakthrough performance and enable the next generation of AI infrastructure.”

As part of the deal, Qualcomm plans to expand its use of AWS AI infrastructure, including Amazon Bedrock, for its own electronic design automation (EDA) workloads, hoping to cut the time of chip design cycles.

The deal forms part of a broader announcement that will see Qualcomm issue warrants to acquire 25 million shares in Amazon for a total investment of $4 billion.

The issuance of some shares is contingent on Amazon purchasing up to $60 billion worth of Qualcomm products.

“This collaboration with Qualcomm Technologies builds on a strong foundation of partnership and reflects our shared commitment to pushing the boundaries of what’s possible,” said Prasad Kalyanaraman, vice president at AWS.

“By working together on customized silicon and advanced connectivity, we’re delivering more performant, efficient, and cost-effective infrastructure for our customers.”

Qualcomm eyes market gains

Qualcomm has hopes of challenging Nvidia's dominant position in the AI data center market. The latter currently holds between 80% and 90%, according to figures from Silicon Analysts, although the figure is predicted to fall to 75% this year. AMD is in second place.

Earlier this summer, Qualcomm updated its Dragonfly A1000 CPUs, Qualcomm High Bandwidth Compute (HBC), and its AI300 inference accelerator. All are engineered to maximize performance per watt and token throughput at lower total cost of ownership (TCO), the company said.

The company recently named Meta as its first customer, with production expected to begin in 2028 and to continue as Meta's server fleet expands. It has also, along with Nvidia and AMD, has signed a memorandum to supply the EU’s AI gigafactories.

Optical networking, meanwhile, is a major element of Amazon’s data center strategy, with the firm buying billions of dollars worth of hardware from fiber-optic cable manufacturer Corning earlier this year.

FOLLOW US ON SOCIAL MEDIA

A new direct-to-device satellite service for businesses has been launched by O2, with the network provider saying it will extend connectivity beyond the reach of its traditional network.

O2 Satellite for Business, as the service is called, is the first service of its kind in the UK, the company claimed. Jo Bertram, CEO of O2 Business, said the product aims to bolster support for organizations working in remote locations or with distributed workforces on the move.

“Businesses increasingly rely on always-on connectivity, but work doesn’t stop where mobile coverage ends,” she said.

“O2 Satellite extends connectivity beyond the reach of traditional mobile networks, helping organizations stay connected in more places through the smartphones they already use every day.”

According to O2, the service will automatically engage when customers are no longer connected to its mobile network and doesn’t require any additional satellite hardware or applications.

Services supported through the offering include messaging, maps, weather, and location sharing through apps such as WhatsApp, Messenger, Google Maps, and Apple Maps.

The service is available on Apple, Google, and Samsung smartphones and O2 said the new network will extend its landmass coverage from 89% to 95% of the UK.

Reliable connectivity

The move by O2 aims to bolster connectivity for the significant number of UK businesses operating in rural or remote areas.

One in five businesses in England are based in rural areas. O2 said reliable connectivity is often a challenge for these firms, which has a direct impact on productivity.

The technology that underpins O2 Satellite for Business has already been used in real-world conditions, the company said.

During wildfires in Derbyshire in late June, it provided 100 satellite-enabled phones to Derbyshire Fire and Rescue Service to support operations in areas without traditional mobile coverage.

“The phones connected directly to satellites, helping teams on the ground stay in contact with colleagues, share updates, and access important information across the affected area,” O2 said.

FOLLOW US ON SOCIAL MEDIA

In recent years, there have been tremendous technology developments and industry shifts. The rapid adoption of generative AI has been a significant disruptor, forcing both vendors and the channel to re-evaluate their offerings and take a more agile stance to expand market reach and boost operational efficiency.

New business opportunities are being driven by trends such as AI storage integration, cloud sovereignty, and the demand for proactive cybersecurity measures. However, given this rapid pace of change, with some partners still focused on legacy resale and non-proactive, break-fix models, are resellers really in a prime position for success?

The biggest trap channel partners could fall into is failing to recognise new opportunities and respond with agility. With AI-driven workflows, regulations, and shifting buyer expectations, those who stick to old approaches risk being sidelined and seeing their margins shrink.

Key industry and technology shifts

Major trends reshaping the channel landscape include the integration of AI with storage infrastructures, the rise of sovereign clouds for data compliance, and the development of advanced data protection platforms.

Another area we see skyrocketing is the edge-cloud-hybrid ecosystem. Organizations are seeking to operate across edge, on-premises, and public/hybrid cloud with seamless data mobility and unified management. With multi-cloud and hybrid architectures becoming the norm, especially for backup and disaster recovery (DR), scalability and cost control remain paramount issues for organizations.

Equally apparent is the need for consistent security and compliance measures across these environments, along with ensuring flexibility to prevent reliance on a single vendor. Data sovereignty, which requires that data be governed by the laws and regulations of the country where it is stored or processed, marks a shift away from generic cloud-first strategies towards more localized management, hybrid-cloud arrangements, and thorough sovereignty evaluations.

With organizations struggling to balance increasing data volumes against rising infrastructure costs, there is also growth in smart multi-tiered storage. This includes storage analytics and tools for real-time monitoring and automated management to enable visibility across different tiers. With the capability to assign "hot" or frequently accessed data to high-speed, premium storage and "cold" or archival data to more affordable options, organizations can avoid the costly mistake of over-provisioning with high-performance drives.

Meanwhile, legacy hardware sales and on-prem-only solutions will continue their sharp decline, falling short for partners who rely on them. The shift to cloud and as-a-service models, combined with economic pressures, has made this a low-margin area.

Channel opportunities for business growth

As global data volumes near 200 zettabytes and AI-native technologies become increasingly sophisticated, customers are demanding comprehensive services that extend beyond traditional project-based hardware or standard cybersecurity offerings.

Customers are seeking partners who can provide AI-ready, sovereign, and resilient data protection that works seamlessly across on-prem, edge, and cloud. With this in mind, key growth areas for the channel are:

AI-driven services and automation

Cyber-resilient data storage solutions and integrated, AI-powered security platforms are driving significant growth opportunities for channel partners. These companies can tap into new opportunities by providing services focused on AI integration that extend beyond just implementing AI tools. These offerings may include automating IT operations, AI monitoring, and intelligence in data protection.

The recent geopolitical shift created by the U.S. government’s unprecedented export-control directive - which restricts global access to the most powerful AI models developed by Anthropic (the prominent US-based AI research and safety company) - has created immediate channel opportunities. If Value-added Resellers (VARs) and Managed Service Providers (MSPs) pivot quickly, they can offer services such as sovereign AI consulting (providing data residency, technological autonomy, and jurisdictional control), regional cloud hosting with sovereign cloud space, as well as advisory services on cross-border regulatory compliance, such as offering risk mitigation audits.

Hybrid-managed offerings

Customers are increasingly adopting multi-tiered architectures that integrate edge computing, on-premises systems, and public cloud solutions. This shift calls for seamless data mobility, unified management, and real-time analytics to ensure efficiency. For channel partners, this trend opens new opportunities for designing, managing, and securing these distributed, complex environments.

The ubiquity of cloud has highlighted a challenge: the importance of operational visibility. Managing data across multiple environments without a unified dashboard has introduced unnecessary complexity and increased the risk of security blind spots. The lack of visibility across edge-cloud-hybrid ecosystems presents an opportunity for partners to act as trusted advisors while also delivering unified management and monitoring across fragmented infrastructures.

Cybersecurity and compliance-as-a-service

Cybersecurity continues to be the foremost growth driver for the channel. Providers are already broadening their offerings to counter the rising frequency of cyberattacks, introducing services such as managed detection and response (MDR), ransomware protection strategies, and compliance solutions. With the emergence of AI-driven threats, there is a growing demand for integrated cyber recovery solutions, presenting lucrative opportunities for partners capable of delivering robust managed security services.

Many opportunities will come from customers who are ready to adopt AI-powered infrastructures and hybrid governance models (the framework that enables managing operations across both physical offices and remote endpoints) in response to data sovereignty pressures. As organizations incorporate AI into their operations, they must address increased security vulnerabilities while ensuring their workloads and data adhere to relevant jurisdictional and privacy laws. Therefore, partners who can offer high-value consulting, data auditing, automation, and managed services, as well as architecting hybrid or sovereign cloud environments, will see the most growth.

Potential challenges

Obstacles to capitalizing on opportunities in the AI era include a lack of sufficient expertise (the talent gap) and longer sales cycles. The channel could struggle to find people who can manage hybrid ecosystems, navigate compliance, and handle the complexity of AI-driven environments. This is an area where vendors can really add value, especially in helping partners bridge that gap with hands-on training, workshops, and webinars.

Another hurdle is that most customers, especially outside the enterprise segment, will face tighter budgets and more cautious spending, which could slow adoption and delay new projects.

There could also be setbacks resulting from vendor complexity, too many tools, programs, and licensing models, which could further slow down the momentum if not simplified early in the year.

The path to success

To embrace these new opportunities, the channel must anticipate changes among customers, vendors, and markets. Buying technology is no longer what customers are looking for; they want solutions that deliver resilience, compliance, and sustainability. The channel must move from being generalists to specialists, guiding and supporting customers and accelerating the shift to service-led models in areas like AI and cybersecurity (such as delivering Backup-as-a-Service and cyber recovery).

Meanwhile, vendors must support this shift by making things easier: offering flexible licensing, providing specialised training and integration tools, and driving closer collaboration with partners. Vendors will also be expected to simplify hybrid and sovereign architectures, provide automation tools, and jointly develop go-to-market initiatives that help partners scale their operations.

The current industry transformation presents a huge opportunity for channel businesses to shift from just reselling to delivering high-value, integrated solutions. As the rise of AI-native and autonomous storage creates a demand for consulting and managed services, clients will most definitely need expert guidance to implement and manage these complex systems.

More HPE networking gear is going into Oracle data centers globally, in an expansion of the companies' existing collaboration.

The expanded deal includes a multi-year deployment of HPE Juniper Networking PTX and MX routing platforms, and QFX and EX switching platforms across Oracle data centers around the world, along with multi-year networking support services.

"Together, we are advancing a high-performance networking architecture that helps Oracle address the traffic management, power, space, and operational requirements of increasingly large AI environments,” said Rami Rahim, executive vice president, president and general manager, networking, HPE.

As things stand, HPE Juniper Networking routing and switching platforms are used to support key elements of Oracle Cloud Infrastructure’s (OCI) data center and edge networks. MX and PTX routers support portions of the edge network, while QFX switches power multiple layers of the regional data center fabric.

However, scale, performance, and availability are becoming increasingly important as OCI expands its AI superclusters. Large GPU clusters place significant demands on bandwidth, latency, congestion management, and fault recovery, HPE said.

Meanwhile, Oracle’s regional and edge networks need to scale to connect clusters with services, customers, and external networks worldwide.

The latest QFX switches will provide the high-density connectivity, dynamic load balancing, and advanced congestion management required for large RoCEv2 (RDMA over Converged Ethernet version 2) AI backend networks.

A common HPE Juniper Networking foundation across these domains helps simplify engineering and operations while enabling each environment to be optimized for its distinct requirements, Oracle said..

“Building AI infrastructure at Oracle’s scale requires an integrated ecosystem where every layer of the stack works together,” said Mahesh Thiagarajan, executive vice president, Oracle Cloud Infrastructure.

“HPE Juniper Networking has been an important part of our cloud architecture and now plays a key role in our AI clusters, regional data centers, and edge networks with the scale, performance, and reliability our customers expect.”

HPE & Oracle eye network visibility

To help OCI keep on top of performance and resiliency as its AI infrastructure scales, the two firms are also collaborating on intelligent telemetry, aimed at delivering deeper visibility into device and fabric behavior.

This should speed up the detection of queue buildup, packet loss, traffic imbalances, and component degradation, to help OCI resolve issues faster and minimize their impact on GPU utilization and network availability.

“Every microsecond matters to AI models and agents,” said Praveen Jain, senior vice president and general manager, data center and AI networking, HPE.

“Through high-performance networking and our work with Oracle on deeper fabric visibility, we are helping OCI detect bottlenecks faster, maintain network availability, and improve the efficiency of AI infrastructure.”

As part of the agreement, HPE has issued Oracle warrants to purchase shares of HPE common stock.

FOLLOW US ON SOCIAL MEDIA

IBM is opening up its most complex quantum processor to researchers, claiming it can solve quantum advantage-level problems ten times faster than its current best model and with no loss in accuracy.

With 120 programmable qubits and a new high-speed qubit reset architecture, Nighthawk r2 is capable of executing over 100,000 circuits per second – 25 times the circuit throughput of IBM’s existing Quantum Heron fleet.

It's already demonstrated accurate observable estimation on circuits containing more than 7,500 gates, according to the firm.

The increase in speed comes thanks to Nighthawk r2's ability to address a longstanding bottleneck in quantum computing: reducing the time needed for qubits to 'reset' between computations.

Previous generations of IBM Quantum processors, including Heron, have managed the reset through a technique powered by dynamic circuits and called conditional reset.

The processor measures the state of a qubit and, if it's found to be in the |1⟩ state and applies a bit flip gate – specifically a π-pulse – that flips the qubit back to the ground state.

However, this technique is limited by the accuracy of measurement, and can't reset qubits that have leaked outside the computational state, meaning a delay of hundreds of microseconds between circuit executions.

Nighthawk r2, instead, uses a dissipative reset gadget, in which each programmable qubit is linked through a high-dynamic-range tunable coupler to a cold environment that draws it back to its ground state on demand.

Activating that coupler pulls a qubit’s effective T1 – the measure of how long it holds its energy – from a median of about 200 microseconds down to roughly 25 nanoseconds, enabling significantly faster high-quality reset.

This active qubit reset leaves neighboring qubits undisturbed while reducing idle time between circuit runs to as little as a single microsecond.

IBM hails Nighthawk compute gains

In practice, IBM said these gains are most noticeable in the large-scale, repetitive workloads that are common in the research community.

"Nighthawk r2’s improvements in speed and quality are already enabling exciting demonstrations of its quantum computational capability," said the firm.

"For example, following recent demonstrations of quantum advantage through trusted quantum computation, researchers used Nighthawk r2 to perform the doped Clifford sampling advantage experiments originally developed in a collaboration between UChicago and IBM."

IBM also cited other advantages of the new processor, claiming it maintains the scale of its r1 predecessor while delivering targeted improvements in quality.

Because the new high-speed reset capability actively cools the qubit to its ground state, it cuts initialization error across the device by about 25 times, making for more accurate results.

At the same time, it keeps up with the gate fidelity of Heron, even while operating at much higher speeds, and performs individual qubit reset without degrading the performance of the qubit’s neighbors.

"Whether you’re scaling application research, exploring the power of dynamic circuits, or advancing the latest quantum error correction research, this new system will help you do more quantum computing, faster," promised IBM.

"Ultimately, the most important thing about a quantum computer isn’t the number of qubits or components; it’s the amount of useful computation the system can deliver."

FOLLOW US ON SOCIAL MEDIA

French infrastructure provider Bull has been selected to lead development of the Lumi-AI supercomputer in a €387.8 million ($449.7 million) deal.

The procurement contract from the EuroHPC Joint Undertaking (EuroHPC JU) will see the next-generation supercomputer hosted at CSC’s data center in Kajaani, Finland.

Under the terms of the deal, the new system will be hosted by the Lumi AI Factory consortium. The consortium launched in April 2025 with the aim of bolstering European HPC and AI infrastructure capabilities.

Anders Jensen, executive director of EuroHPC, said this marks the sixth AI Factory procurement contract since the launch of the scheme.

“Together, these systems will form a powerful European AI infrastructure ecosystem, enabling SMEs and startups to access world-class supercomputing resources, unlock the full potential of AI, and drive innovation, growth, and technological leadership across Europe,” Jensen commented.

Funding for the system will be shared equally between EuroHPC and the Lumi AI Factory consortium.

Under the hood of the Lumi-AI supercomputer

The Lumi-AI supercomputer will be based on the BullSequana XH3500 architecture and powered by AMD Instinct MI430X GPUs and 6th Gen AMD Epyc 256-core processors.

The BullSequana XH3500 platform is designed specifically for “advanced AI workloads”, the company said in a statement.

All told, EuroHPC said the system will deliver a “tenfold increase” in AI capacity and nearly double the high-performance computing (HPC) capability of the existing Lumi supercomputer, also hosted at the Kajaani site.

IBM will also contribute to development of the AI supercomputer, providing storage solutions based on IBM Storage Scale. Nokia, meanwhile, will lead on networking capabilities, leveraging Bull’s BXI interconnect technology for supercomputers and AI HPC workloads.

According to the Lumi consortium, the inclusion of additional compute and storage resources will play a key role in “maximizing” the overall performance and cross-functionality of the system.

“A key differentiator is its multi-tenant architecture, designed to meet the needs of diverse AI communities, combined with extensive API-based access to support modern development environments,” the company said.

Thomas Zacharia, AMD’s SVP for global public sector and strategic partnerships, said the inclusion of Instinct GPUs and Epyc processors marks an “important step” in the chipmaker’s existing relationship with EuroHPC and the Lumi AI Factory consortium..

The existing Lumi supercomputer already features AMD MI250x GPUs and AMD Epyc Milan CPUs.

“Powered by AMD Instinct MI430X GPUs and 6th Gen AMD EPYC CPUs, [Lumi-AI] will strengthen Europe’s sovereign AI capacity, accelerate scientific discovery and give researchers, startups and industry the scale and flexibility to shape Europe’s AI future,” Zacharia commented.

Bull’s first outing

Bull’s selection in the project marks the first major deal for the company after it was renationalized by the French government in March 2026. as part of a €404 million ($406.4 million)deal with Atos Group.

The motive behind the deal primarily centered around protecting French national and technological sovereignty.

At the time, Roland Lescure, French minister for the economy, said the acquisition marked a “decisive milestone for France’s technological sovereignty”.

FOLLOW US ON SOCIAL MEDIA

Anthropic has reportedly signed a $45 billion cloud computing deal with Nscale as the AI developer aims to scale compute resources.

Sources at the companies told CNBC the six-year deal will see Anthropic access 460MW of Nvidia's next-gen chips at the Monarch Compute Campus, starting at the end of next year.

The West Virginia site is expected to begin operations at the end of next, and will host Nvidia’s Vera Rubin chips.

Anthropic is yet to respond to a request for comment from ITPro. A spokesperson for Nscale said they could not confirm details from the reports.

Anthropic has signed similar deals recently aimed at scaling compute capacity, including a $5 billion deal with AMD. The company also struck a deal with SpaceX to cover up to 300MW of compute capacity. The AI developer also has an existing deal with Oracle.

London-based data center startup Nscale was spun out of an Australian bitcoin mining company in 2024, and was earlier this year valued at $14.6bn following a $2bn fundraising round. An IPO is widely expected to land imminently.

Anthropic isn’t alone in striking infrastructure deals, either. Major players including OpenAI have also invested incredible sums to build or rent compute capacity, helping to spark record results for cloud computing companies.

Last month, Google Cloud posted revenue growth of 82% — its best quarterly growth yet.

Monarch compute

In January, the American Intelligence and Power Corporation (AIPCorp) unveiled its plans to build the 2,250-acre Monarch Compute Campus in Mason County, West Virginia, dubbing it the first state-certified microgrid-based AI project.

The site boasts up to 2GW of power generated on campus by natural gas scaling to 8GW. The first phase is set to open at the end of 2027, and it is expected to become one of the largest AI infrastructure projects yet.

Two months after the unveiling, Nscale announced it had acquired AIPCorp, including the Monarch Compute Campus — and signed Microsoft as its first customer for 1.35GW of AI compute on Nvidia's next-gen Vera Rubin GPUs.

While the data center is drawing big-name AI customers, locals have complained about construction disruption causing damage to local streets and causing floods, with one report saying 60% of homeowners in nearby areas sold and moved rather than live near the campus.

Advocacy groups have raised concerns about water use, though Nscale has pledged there will be no impact on municipal water supply or residential users. As the campus is off-grid, local electricity prices or supply won't be impacted, but environmental groups have warned about the climate impact of the use of natural gas for generation.

A report from EY estimates the data centre will make up more than a third of Mason County's economy once operational.

Nscale has multiple sites planned — including Essex in the UK, and Texas — and one operational data center in Norway, with another set to open in the country imminently.

Beyond that, Nscale has several partner-run data centers in Europe and North America.

FOLLOW US ON SOCIAL MEDIA

Nvidia CEO Jensen Huang claims the AI industry is in the midst of a “golden age” after the company recorded huge revenue gains in its recent quarterly earnings report.

Revenue surged to $96.2 billion in the second quarter, marking a 106% increase compared to this time last year. The chipmaker said it expects revenue to surge to $108 billion by the end of the third quarter.

In a statement, Huang said the revenue gains reflect the growing impact of AI worldwide, with the technology having now reached an “inflection point”.

“It’s doing useful work. Its tokens are productive and profitable. Now, compute is revenue,” he said.

“And demand is accelerating. This time last year, one lab alone was driving the buildout; today, we have a golden age of new AI labs and startups, multiple frontier labs scaling in parallel, a thriving open model ecosystem and physical AI coming online — with strong momentum across the US and around the world.”

Huang noted that AI infrastructure build-outs have played a key role in recent revenue gains – an area in which the company also recorded significant growth.

Data center revenue across the second quarter increased to $89 billion, up 18% from the previous quarter and marking a 117% surge year-over-year. Notably, this earnings report marks the first since the Nvidia Vera Rubin rackscale platform announcement in May, which the company said is now “ramping into full production”.

Vera Rubin racks are now up and running at a host of major industry players, including Google Cloud, Microsoft, Oracle, CoreWeave, and Nebius.

Nvidia supply issues still a cause for concern

Forrester principal analyst Naveen Chhabra echoed Huang’s comments regarding AI infrastructure build-outs. However, for enterprise technology buyers, a combination of industry trends could dampen excitement - particularly supply constraints.

“The most important message is that AI infrastructure demand is still accelerating, but supply constraints, ecosystem lock-in, and ROI pressure are becoming the dominant strategic issues,” he said.

Chhabra said the Nvidia results suggest that the AI infrastructure market is “entering a new phase where enterprises must focus less on acquiring GPUs and more on extracting measurable business value from AI investments”.

Huang acknowledged that supply constraints are a lingering concern. The company expects revenue growth of 70% for its 2028 year, and while Huang said demand is “much greater” than that, it’s limited on the volume of products it can roll out.

Chhabra said supply issues show that multi-cloud and hybrid sourcing strategies have become “increasingly important to avoid dependence on a single provider” – especially for larger enterprises.

“Large AI training clusters and high-end inference deployments may continue to face availability constraints,” he said.

Heightened competition

While Nvidia has been center stage since the get-go in the generative AI race three years ago, enterprises are shifting their focus to alternatives in a bid to diversify infrastructure hardware.

A host of major industry players are also developing their own chips, in part to reduce reliance. In an earnings call, Huang brushed off concerns on this front, adding that he has “100% confidence that our technology will continue to be extraordinary for them”.

Alongside the earnings results, Nvidia revealed AWS plans to acquire another two million GPUs to bolster data center capacity. The announcement comes less than six months after the hyperscaler unveiled plans to deploy one million GPUs across data centers.

Chhabra noted that while enterprises are tweaking strategies, Nvidia’s broader ecosystem still holds significant weight and is a key differentiator for the chipmaker.

“Accelerators options are emerging. However, Nvidia's advantage is not simply silicon but the entire platform, software stack, and ecosystem,” he said.

“Nvidia supports "every model" and every deployment environment. The market is unlikely to become "Nvidia vs. one competitor." Instead, buyers will operate in a heterogeneous environment with Nvidia, hyperscaler chips, and specialized accelerators."

FOLLOW US ON SOCIAL MEDIA

SMBs seeking a powerful single-socket rack server that'll fit in the smallest of spaces will find Broadberry is offering a tempting proposition. Presented as an ultra-short-depth 1U rack system, the CyberServe EPYC EP1-102A packs a lot into its compact dimensions.

Targeting SMBs and workloads such as edge computing and web hosting, Broadberry offers it with a good choice of AMD EPYC 4004/5 CPUs, and it can handle up to 192GB of fast DDR5 memory. It has another trick up its sleeve as its single PCIe slot supports an Nvidia Quadro RTX A1000 or A400 GPU card for lightweight AI and rendering duties.

Remote server monitoring comes as standard, as the CyberServe sports an integrated AST2600 BMC (baseboard management controller) with a dedicated Gigabit port. You can also manage and monitor multiple servers with the Asus Control Center Express (ACCE) utility providing a centralised dashboard.

Broadberry CyberServe EPYC EP1-102A review: Build and design

The CyberServe showcases the Asus ExpertCenter Pro ER100A B6 1U rack platform. The chassis is solidly constructed, and removing the lid reveals a smartly designed and tidy interior affording easy access for upgrades and maintenance.

Broadberry supplied our system with a meaty 16-core 3GHz AMD EPYC 4545P topped off with a chunky passive heatsink. Flanked on one side by four DIMM slots, it came with one occupied by a 32GB DDR5 UDIMM module.

Cooling is handled efficiently by a central bank of three dual-rotor fans in front of the motherboard with a plastic air shroud directing air over the CPU's heatsink. We noted that the shroud is secured in place with four screws – many similar short-depth Supermicro rack servers we've seen leave their flimsy air shroud loose, with only the lid keeping it in place.

Another cooling fan sits to the left of the chassis and is angled in to push air over the PCIe expansion slot. All told, we found the server ran very quietly, making it a good choice for small office environments.

The rear panel presents dual 2.5GbE multi-Gig ports, and if you feel the need for more speed, Broadberry offers 10GbE, 25GbE and 100GbE PCIe adapters. You're spoilt for choice if you want to connect a local monitor, as the server offers VGA ports fore and aft, while the rear 10GbE 3.2 Gen 2 port supports DP Alt-Mode for piping output to an HDMI-enabled monitor.

The front panel also has a handy 2-character Q-code LED display. This provides a running status tally as the server boots up and warns if any errors are detected – the manual provides a full list of Q-codes and their meaning.

Broadberry CyberServe EPYC EP1-102A review: Storage features

Standard storage features are fairly simple, with the base system offering two hot-swap SFF bays at the front with the backplane cabled directly to the motherboard's embedded SATA ports. Drives are easily added by snapping them into the tool-free carriers, and it's worth noting that unlike the blue-chips, Broadberry allows you to install your own choice of storage devices without invalidating the warranty.

Further storage expansion is possible as the server has an extra internal tray behind the front air grill that accepts two fixed SFF devices. Power cables are already provided, and with spare SATA and SlimSAS ports on the motherboard right next to the tray, you can cable up SATA or U.2 NVMe drives.

The system was supplied with a 2TB Micron 3500 M.2 NVMe SSD installed in the single motherboard slot beneath the expansion bay. A nice touch is that the SSD is completely covered with a large aluminium heatsink, which also gets the benefit of the expansion bay cooling fan.

Broadberry offers a range of optional Broadcom SATA/SAS RAID adapters, but the AMD chipset does have embedded RAID capabilities, which come into play when you change the BIOS SATA settings from AHCI to RAID. Arrays can be created and managed using AMD's RAIDXpert2 utility, which is available for Windows Server 2022/2025. We dropped a couple of WD Red SATA SSDs in the front bays and had no problems using RAIDXpert2 to create a software-managed mirrored array.

Broadberry CyberServe EPYC EP1-102A review: Server management

The server's embedded AST2600 IPMI controller and dedicated Gigabit port offer a tidy web interface packed with plenty of information on system health and critical components. It isn't as feature-rich as Dell's iDRAC or HPE's iLO controllers, but its dashboard provides an overview of all the latest events, and you can pull up graphs of hardware sensor data, define thresholds for each one, and link them up with email alerts if any are breached.

Firmware upgrade tools are provided along with remote power controls, and the controller's built-in firewall provides options to create IP address and port-based protection rules. Unlike Dell and HPE, the IPMI controller includes full OS remote control and virtual media services as standard features and not as chargeable licence upgrades.

The ACCE software is available for Windows and Linux hosts and is used to manage multiple servers from a single console. A licence code is included on a card in the server's accessory box, and we loaded ACCE on a Windows Server 2025 host.

ACCE focuses primarily on managing Windows and Linux endpoints and offers a remote agent deployment tool. Systems with agents appear in the main dashboard page where you can view their hardware sensors and perform a variety of tasks such as running BIOS and driver updates and system inventory.

The management controller section scans for BMC devices, and after discovering the CyberServe's AST2600, it presented a list of tasks including pulling up hardware sensor details, updating the BMC's firmware, and loading event logs. As far as we can see, BMC monitoring doesn't need a licence, which is only required for each agent deployment.

Broadberry CyberServe EPYC EP1-102A review: Is it worth it?

With a price tag for the system on review of £3,656 excluding VAT, Broadberry's CyberServe EPYC EP1-102A is offering a very affordable alternative to the blue chips. It comes up a little short in the storage department, but it is available with a good choice of EPYC CPUs, has a big memory capacity, supports an optional GPU, and its short-depth chassis can go places most other servers can't.

Broadberry CyberServe EPYC EP1-102A specifications

Semiconductor revenue is set to almost double this year as the AI sector continues to reshape the market – and drive up prices for in-demand components.

New figures from Gartner project global semiconductor revenue growth to top 92% this year, reaching $1.6 trillion over last year's $809 trillion. In 2027, that’s expected to climb further to $1.9 trillion, the company said.

"The semiconductor industry is entering a fundamentally new phase of growth," said Ben Lee, Director Analyst at Gartner. "The pace of industry expansion is accelerating dramatically, driven by sustained investment in AI infrastructure and a stronger-than-anticipated memory pricing cycle."

That's good news for those manufacturing semiconductors, but the rise of AI – and the surge in demand that it's sparked for components – has led to price rises across a range of industries.

In the cloud market, a host of providers have issued price hike warnings. Meanwhile, consumer electronics are being hit hard by rising hardware costs. Apple recently warned that prices will rise this year, for example, while PC shipments have hit record lows.

Memory and beyond

Semiconductor gains are largely driven by memory revenue, with prices for the key component skyrocketing because of AI demand. Some predictions suggest that AI will consume a fifth of all DRAM capacity in 2026.

That shortage is sending prices skyrocketing and already impacting businesses' plans to upgrade tech.

This trend is expected to continue, according to Gartner. The firm predicted DRAM revenue would increase 246.6% this year, with NAND flash revenue up by 371.9% – and that’s despite additional capacity being made available.

Memory now makes up 54% of total semiconductor revenue, a huge leap from 27% last year. Revenue for memory will top $837bn this year and $1trn next year, Gartner predicts.

"AI infrastructure has fundamentally changed the dynamics of the memory market,” said Shrish Pant, Director Analyst at Gartner.

“While pricing expansion is accelerating growth in 2026, continued AI infrastructure deployments, higher memory content per AI server and sustained demand for high-bandwidth memory (HBM) will support memory revenue growth through 2027 and beyond."

While semiconductor revenue growth is dominated by memory price rises and demand, Gartner noted that the AI roll-out is benefiting all aspects of the industry, as AI data centers need CPUs, networking, power management, and more.

"AI is expanding the semiconductor opportunity across the entire infrastructure stack rather than benefiting a single device category," said Lee.

Gartner said semiconductor revenue growth excluding memory would grow from $589 billion in 2025 to $718 billion in 2026, an increase of 21.9% – and reach $864 billion in 2027.

AI data centers

Notably, the mass buildout of data centers is changing how semiconductors are used and where they end up, according to Gartner.

"As AI infrastructure scales, it is not only reshaping investment across the semiconductor ecosystem but also redefining the industry’s application mix," said Lee, saying Gartner sees a "a structural shift in where semiconductor value is created and how demand is evolving across the industry.”

Gartner predicted that the AI data center ecosystem would make up over a third (36.5%) of semiconductor revenue this year, climbing to more than half (53%) by 2030.

FOLLOW US ON SOCIAL MEDIA

AMD has boosted rack efficiency four-fold since 2024, and the company claims it’s on track to upping it by 20 times by the end of the decade.

The chipmaker revealed it recently exceeded its 2026 target of a three-fold improvement in rack-scale energy efficiency for AI training and inference.

Two 2030 AMD racks are now expected to deliver the same compute as roughly 570 racks did in 2024, enabling a 20x reduction in use-phase electricity and a 28x reduction in carbon intensity.

Alternatively, efficiency gains are expected to enable 20-times more compute, measured in floating point operations per second (FLOPs) per watt, using the same amount of energy.

AMD said energy improved energy efficiency delivers a range of environmental benefits, but also in helping to improve AI performance without increasing power, improving total cost of ownership, and helping customers scale faster.

“The next wave of AI efficiency will depend on tighter co-optimization across compute silicon, memory, interconnects, software and rack-scale system design," said Sam Naffziger, senior vice president and corporate fellow at AMD.

"Our estimated 4x improvement through 2026 reflects the strength of our approach and the progress AMD is making across the full system, putting us ahead of our projected pace toward the 2030 goal.”

The improvements in performance come down to compute capability, memory bandwidth, and interconnect bandwidth, according to AMD.

Advanced process technology and architectural improvements help increase floating-point compute performance per watt, advanced architectures and memory integration improve bandwidth, and high-speed interconnects and tighter integration increase the network bandwidths.

High-bandwidth memory, larger caches, and tighter integration between memory and compute can reduce unnecessary data movement, which uses energy. Meanwhile, high-speed scale-up interconnects help larger systems reduce bottlenecks and share data more efficiently.

AMD’s ROCm software driving gains

Elsewhere, AMD claimed its ROCm software stack and open standards help developers and organizations deploy AI workloads more efficiently on AMD platforms, with open ecosystems also supporting broader optimization across AI and high-performance computing (HPC) deployments.

"The projected rack-scale efficiency gains AMD has set forth can enable AI and HPC customers to achieve data center-level gains in power and cooling infrastructure, which can be limiting factors to increasing compute performance," said the firm.

"Better energy efficiency can help reduce operational electricity use, improve total cost of ownership, and support more sustainable AI and HPC growth."

AMD has also published its 2025-26 Corporate Responsibility Report, detailing its progress across product efficiency, operations, supply chain responsibility, and community impact.

It reduced operational greenhouse gas emissions by 30% from 2020 to 2025, sourced 58% of global AMD electricity from renewable sources and audited 99% of its manufacturing supplier factories for responsible business practices.

"We aim to cut operational greenhouse gas (GHG) emissions in half from 2020 to 2030 as we expand our use of renewable energy. In 2025, 100% of AMD manufacturing suppliers maintained public GHG reduction targets and 82% sourced renewable energy, said AMD chair and chief executive officer Lisa Su.

"The AI era is still in its early stages. Working closely with industry, government and academia, we will keep building the high-performance, energy-efficient compute foundation it requires."

FOLLOW US ON SOCIAL MEDIA

AI cloud firm Nebius is leasing high-density capacity at Vantage Data Centers' Newport campus, marking the first commercial capacity commitment in the South Wales AI Growth Zone.

Under the agreement, Nebius will lease Nvidia-powered capacity at Nebius's CWL1 facility to support AI training, inference, agentic AI and enterprise AI workloads, and serving enterprises, researchers, start-ups and public sector organizations.

“This is an important milestone for South Wales and for the UK’s AI infrastructure ambitions,” said David Howson, president, EMEA, Vantage Data Centers.

“Nebius is scaling AI cloud infrastructure in the UK at a time when demand for domestic compute capacity continues to accelerate. Vantage’s Newport campus provides the scale, connectivity and operational excellence needed to support that growth, and we’re proud to help establish South Wales as a leading destination for AI investment.”

One of Europe’s largest data center campuses, Nebius's CWL1 site opened in 2010. It runs on 100% certified renewable energy, with its latest facilities designed to minimize water consumption through a closed-loop cooling system that recirculates water rather than evaporating it.

The facility also uses air-cooled chillers and free cooling when ambient conditions permit.

The increased capacity forms part of Nebius’ broader UK expansion and commitment to scale domestic AI compute capacity.

In June, the company announced around £1.7 billion of committed capacity build-out across four UK sites, including the delivery of more than 1GW of AI-ready capacity across Newport, Bridgend, and Bro Tathan.

“The UK is one of the places where AI is being built, deployed and adopted by startups, enterprises and the public sector,” said Gary Tierney, general manager for EMEA at Nebius.

“Vantage’s Newport campus gives us a strong foundation to expand access to NVIDIA-powered AI infrastructure in the UK and support customers building the next generation of AI applications.”

AI Growth Zones take shape

The buildout is based on Nvidia's latest full-stack, end-to-end DSX AI factory reference design, and supports the UK government’s AI Opportunities Action Plan.

South Wales was designated as one of the UK's UK AI Growth Zones in November last year. At the time, the government said it could create more than 5,000 new jobs for local communities.

It's also home to the UK’s largest cluster of semiconductor businesses.

The AI Growth Zone stretches along the M4 corridor from Newport to Bridgend has the potential to harness over 1GW by early 2030s. 

“Today’s announcement marks an important milestone for the South Wales AI Growth Zone and demonstrates Wales’ growing reputation as a strategic location for the infrastructure that will power the AI economy,” said Adam Price, Welsh cabinet minister for enterprise, connectivity and energy.

"Nebius and Vantage’s investment will help unlock new economic opportunities, support highly skilled jobs and reinforce Wales’ position as a leading destination for advanced digital investment."

The move by Nebius comes hot on the heels of AI Growth Zone announcements elsewhere across the UK this week. The Lanarkshire growth zone in Scotland, for example, secured £300 million in investment to ramp up data center development, with Dell Technologies pledging its support for the site.

FOLLOW US ON SOCIAL MEDIA

Businesses of all sizes depend on a mix of technologies to access and move data.

Wired networks provide the critical data backbones and connections to data centers.

But more and more enterprise data is carried by alternative networks. Some traffic never even touches the conventional, corporate LAN but goes directly from a mobile user, an IoT device, or a branch office to a cloud provider or a SaaS application.

Cellular connections through 4G and now 5G routers, and satellite links are already common as backups to fixed lines, for events and emergencies, for hard-to-reach locations, and even for large sites, including warehouses, sports venues and transport hubs.

But alternative, wide-area network technology is gaining ground for routine use, rather than just edge or temporary use cases.

Businesses want to make more use of data-rich applications, from video and voice to AI, with high-speed, high-quality connections across their operations.

Then there is the growing interest in operational resilience. A single network connection is a bottleneck and a potential single point of failure.

Increasingly, this means looking at more than one network. And technologies previously viewed as only suitable for edge cases are now moving to the core of the enterprise network.

“The traffic patterns are becoming very bursty, asymmetric, and highly latency sensitive,” Hardik Ajmera, a vice president at Extreme Networks, told ITPro. It no longer makes sense, he says, to route all network traffic through the corporate office or data center.

But how do enterprises manage the inevitable complexity this brings?

Ever since the first groundbreaking microchip was created in the late 1950s, the race has been on to make these components smaller, more powerful, and faster.

In 2026, the quest for dominance in the AI sector is driving investment and innovation in the processor sector: AI may be software, but the power and speed of these chips will determine the success of the AI platforms.

AMD, Intel, and Nvidia are all leading producers of GPUs and CPUs for artificial intelligence, but coming up behind this trio is Amazon Web Services (AWS) and its Trainium series of chips. We explain what AWS Trainium is and how the technology industry is adopting these chips to deliver AI models at scale and speed.

What is AWS Trainium?

AWS Trainium is the name of a series of computer chips designed to power both the training and the running of artificial intelligence. The name ‘Trainium’ is simply a play on words: machine learning models need to be trained, while the ‘-ium’ part evokes chemical elements, such as titanium.

According to Amazon, AWS Trainium has been custom-built “to do one thing and one thing only: process massive amounts of data to help advance the development of generative AI”. In fact, Amazon describes AWS Trainium as “a powerhouse of brute computational force,” meaning every microscopic particle of the chip has been refined for the maximum amount of processing capability.

Amazon says a single AWS Trainium chip can complete trillions of calculations in just one second, and it’s this power the company hopes will see its chips at the core of AI infrastructure in the years ahead.

With all the major tech giants competing to place their specific chips at the heart of AI, it’s easy to see why this matters so much; a 2025 report from UN Trade and Development predicted the AI market would be worth $4.8 trillion by 2033.

When was AWS Trainium launched?

AWS Trainium was first announced in late 2020 with a roadmap to arrive in the first half of 2021. The AWS Trainium family is now up to Trainium3 (Trn3), unveiled at AWS’s re:Invent expo in December 2025. Compared with Trainium2, it doubles the compute performance and has 1.5x the memory capacity.

AWS Trainium chips are made by Annapurna Labs, an Israeli-based microelectronics company acquired by Amazon Web Services in 2015 for around $350 million.

Who uses AWS Trainium?

Direct Trainium customers are typically organizations that are involved in the training of large language models and other frontier AI.

Anthropic’s Claude, for example, has been trained – and runs – extensively on AWS Trainium chips. Additionally, Anthropic’s Project Rainier – one of the world’s largest AI compute clusters – is built on more than a million AWS Trainium2 chips.

Other AWS Trainium customers at present include Databricks, HCL, Hugging Face, PyTorch, and Ricoh.

What are the benefits of AWS Trainium?

One of the key selling points of Trainium is its reported efficiency.

Since the launch of AWS Trainium3, Amazon says customers have reduced their AI training costs by up to 50% as well as massively lowering inference latency, which is the time taken between sending a request to an AI model and receiving an answer. Energy efficiency has also been increased by 40%.

AWS Trainium3 UltraServers consist of 144 Trainium3 chips, and Amazon says these cut the time taken to train AI models from months to weeks, making previously impractical or too expensive projects now a reality.

How does AWS Trainium differ from AWS Graviton?

Both AWS Trainium and AWS Graviton are types of computer chips from Amazon Web Services. However, while AWS Trainium is used for the training and running of AI models, AWS Graviton handles the day-to-day cloud computing needs that keep the internet running.

It’s also the power behind the rollout of agentic AI systems by companies worldwide and is used to manage the workload of these agents to respond automatically to queries from users. AWS Graviton customers include Arm, Crowdtrike, HubSpot, Pinterest, and Snap.

What comes next for AWS Trainium?

Amazon is already working on AWS Trainium4, which it says will represent a “foundational leap” forward in processing power, speed, and efficiency for modern AI workloads.

This could include offering a 6x performance upgrade over AWS Trainium3. These chips are expected to arrive sometime in 2027 and will have unprecedented interoperability with Nvidia NVLink Fusion high-speed chip interconnect technology.

With data centers dramatically increasing pressure on the electricity network, the UK's National Grid has announced plans to improve power provision in North West London.

Five new data centers – equivalent to 1GW of demand – will be connected in Hertfordshire to the existing 400kV Elstree GIS substation, with a new 132kV Letchmore Heath GIS substation planned for next door.

Also in the pipeline are upgrades to Sundon substation in Bedfordshire, the reconductoring the 35km Sundon–Elstree overhead line route, and an extension to the existing Elstree gas insulated switchgear (GIS) substation.

Meanwhile, work on the route between Sundon and Elstree will see over 200km of overhead line upgraded, with 60km of new cable infrastructure also to be installed within the existing Elstree–St John's Wood tunnel.

Additional works include:

• The installation of a new 400kV cable circuit in the existing Elstree–St John’s Wood tunnel
• Upgrades to the Elstree–St John’s Wood tunnel and headhouses
• The reinforcement of the existing St John's Wood substation

“The North West London Upgrade is a nationally significant investment that will strengthen the electricity network, unlock new connections, support regional economic growth and help keep long term energy costs down for consumers," said Joe Senior, project director at National Grid.

“Working closely with our delivery partners and supply chain, we're bringing together substations, cables and tunnels, and overhead line reinforcement in a single program to deliver the infrastructure needed to power future homes, businesses and digital industries as electricity demand rises.”

In line with government guidance, communities near the new Letchmore Heath substation are set for £530,000 in funding for projects that improve the social, economic, and environmental wellbeing of the community.

National Grid eyes capacity gains

According to National Grid, these upgrades will increase network capacity and strengthen the resilience of electricity supplies across the region, helping to keep energy bills down.

It should also play a role in supporting economic growth across North West London as electricity demand rises – enabling new data center connections without jeopardizing power supply to homes and businesses.

As part of these efforts, National Grid is carrying out other ongoing work across the capital, including London Power Tunnels 2 (LPT2) and the North London Reinforcement (NLR) scheme.

Earlier this year, National Grid revealed that the queue for connections to the transmission network had grown by 460%, in large part because speculative applications were swamping the pipeline and delaying connections for strategically important projects.

The result, according to the National Grid, is that some waiting times have surged to up to 15 years.

Because of this, the government announced plans to prioritize AI data centers, AI Growth Zones, and industrial sites when it comes to new connections.

FOLLOW US ON SOCIAL MEDIA

Voice remains one of the most effective channels for customer relationships. For customers, having access to a human on the other end of the line, especially in times of confusion or crisis, is increasingly important in an era driven by automation and AI adoption.

This is exactly why so many Managed Service Providers (MSPs) are involved in running or supporting contact center infrastructure on behalf of their customers. However, despite the rewards, it naturally doesn’t come without some risks.

In April 2026, the UK government published the findings of its 2025/26 Cyber Security Breaches Survey, which highlighted some sobering statistics. According to the data, nearly half of all UK businesses (43%) experienced a cyber security breach or attack in the past year.

Phishing remained the dominant attack vector, cited by 38% of businesses, and rated as the most disruptive incident type by 69% of those affected. However, of those phishing incidents, a growing share is no longer delivered by email. Rather, they’re arriving by phone call, and thanks to AI, they are increasingly convincing.

The evolving threat

Voice phishing, or “vishing”, is not a new phenomenon. Bad actors and criminals have been spoofing voices and using fraudulent calls since the popularization of the telephone. What is new are the features that make voice phishing more convincing.

The proliferation of AI-generated voice deepfaking, caller ID spoofing tools, and lack of traceability have contributed to a new wave of fraudulent calls that appear more real than ever before.

Many organizations have spent significant resources securing their email environments, including spam filters, user training, and sandboxing. While this is a sensible investment, fraudsters have noticed and are adapting, which may explain why voice phishing has become an increasingly favored tactic for some bad actors.

Why voice is being targeted

The basic mechanics of voice phishing have always relied on social engineering. Impersonating a trusted person and creating a sense of urgency help to manipulate victims into taking a malicious action, whether that be authorizing a payment, sharing credentials, or granting undue access to a system. While traditionally a mismatched voice was a potential giveaway to fraud, AI deepfakes have made this process much more convincing.

Voice cloning tools that once required real technical knowledge can now be replicated by most people with a commercial PC. It’s now easier than ever to create a convincing replica of a known voice from a short audio sample, for example, from a CEO's recorded video message, a support call that was unknowingly captured, or a social media clip.

If a cloned voice calls a finance team member and asks them to process an urgent payment, citing relevant details such as a live deal, a regulatory deadline, or a supplier relationship, very few red flags are present for the human on the receiving end of the call.

Number spoofing compounds the issue. Fraudsters may be able to secure the capability to present any caller ID they choose, making a call appear to originate from a trusted source. The result of mixing emerging tech with good due diligence is a call that looks and sounds real.

MSPs can carry disproportionate risk

MSPs that support or run contact center infrastructure are in the middle of communications flows for multiple organizations and partners simultaneously. Critically, this also makes them the first port of call for customers when something technical goes wrong, which can offer an easy alibi for fraudsters looking to exploit them.

There are two distinct risks for MSPs: one operational, one reputational. The operational risk is straightforward: IT help desks control credentials, systems, and sensitive customer data, making them a target for malicious actors. The reputational risk, however, has a wider-reaching impact. If a customer is defrauded through infrastructure managed by an MSP, customers may find the MSP liable. Even if the MSP bears no technical responsibility, they are the supplier in the middle.

There is also a less visible third risk: the channel as a route of compromise. Fraudsters who have already gained access, whether through an earlier breach, open-source intelligence, or a compromised supplier, will probe the path that offers the least resistance. If an MSP's call handling has not been safeguarded with fraud in mind, that becomes the path of least resistance.

Practical steps to take right now

The tools to address this problem exist today, but gaps remain in general adoption and awareness. Firstly, MSPs should be implementing proper caller authentication where it is available. For example, in some markets, protocols exist to authenticate calls before they’re answered, such as the “STIR/SHAKEN” protocols used in the USA. These act as a digital "passport" for telephone calls, ensuring the number a call handler sees on their caller ID is legitimate and matches the actual person or business calling. These act as a strong defense against number spoofing.

Next, establish comprehensive verification workflows. Agents should never rely on voice recognition alone. Callbacks to verified numbers, challenge questions, and two-factor authentication must be embedded as standard security procedures. Call handlers must verify who is requesting sensitive information and challenge even those who appear familiar.

Additionally, staff should be trained to recognize the specific warning signs of voice phishing. Most security training currently focuses on email. Staff need scenario-based training covering AI-generated voice hallmarks, such as unnatural pauses and scripted rigidity, alongside classic social engineering tactics: creating a sense of urgency and requesting to bypass usual processes.

Finally, participate in industry data-sharing. No single provider sees the full picture. In the UK, mechanisms like the Home Office's Telecoms Fraud Sector Charter seek to improve cross-sector visibility of active campaigns. MSPs engaged in these networks get earlier warnings to ensure they’re not operating with only a partial view.

Voice is a channel worth championing

Voice is a communication channel with many strengths, and the human touch can be the differentiator between satisfied and dissatisfied customers. For so many reasons, it’s a critical channel that’s worth championing, and worth safeguarding.

Voice fraud is the channel's problem. But the channel is very well placed to address it.

Steps such as appropriate due diligence through the supply chain, monitoring for suspicious traffic patterns, and validating caller IDs used by customers are all effective. It’s also crucial that MSPs and contact centers take the appropriate steps now to ensure that criminals can’t get in and customers remain confident.

AMD is planning to offer up to £5 million in support for a new lab in Oxford carrying out research into frontier AI.

The company has signed a memorandum of understanding to provide compute resources for researchers at the British Open-ended Learning and Discovery Lab (BOLD) during the lab's first eighteen months in operation.

BOLD is led by the University of Oxford, and brings together researchers affiliated with the University of Oxford, University College London, and Imperial College London.

They, together with AMD researchers, plan to explore initiatives spanning compute access, software optimization, research collaboration and open-source development.

Work could include experiments to explore model training techniques beyond backpropagation, discovery using multi-agent systems and adaptive physical AI.

The lab will be built around three research aims - new learning algorithms, human-centered AI and embodied systems such as robots - and will translate breakthroughs into open source tools and encourage commercial spin-outs.

“Through our memorandum of understanding with the BOLD lab, AMD is exploring bringing together advanced computing, open software and technical expertise to help researchers pursue ambitious new frontiers in AI," said Thomas Zacharia, senior vice president, global public sector and strategic partnerships at AMD.

"By expanding access to compute and supporting open academic research, our aim would be to help BOLD researchers accelerate experimentation and contribute to the UK’s leadership in AI innovation and scientific discovery.”

AMD to provide underpinning infrastructure

AMD plans to provide support through infrastructure that will be maintained by the company itself, along with cloud service providers, public-sector partnerships and hardware contributions.

Phase 1 of the research program will focus on broad exploratory experiments, with later phases following up select research directions and validating the most promising ideas.

The company said it will work with BOLD researchers across the AI technology stack, exploring novel ways to optimize AI workloads using its open software ecosystem.

The ecosystem envisaged by the MOU would be built on permissively licensed open-source projects such as AMD ROCm software. AMD would help BOLD researchers publish research software, benchmarks, and evaluation tools.

"BOLD aims to advance a more open, efficient and human-centered path for AI through collaboration between academia and industry. Working with AMD, the University of Oxford would combine world-leading research with computing technology to test ideas, learn quickly and translate breakthroughs into applications," said associate professor Jakob Foerster, of the University of Oxford's Department of Engineering Science.

"This support would help BOLD develop approaches to learning, human-AI collaboration and embodied intelligence while expanding open access to computing resources that benefit science, healthcare, education, transportation and public services globally.”

UK pledges continue

The MoU marks the latest in a number of UK-related pledges made by AMD in recent months.

In June, for example, the chipmaker announced plans to invest £2 billion in the UK to drive AI innovation and research. This also saw AMD pledge to expand access to compute resources for enterprises across the country.

Along with University College London-led research lab Science of Fundamental AI Research (SOFAIR), BOLD is to receive a share of £60 million in government funding over the next six years.

This funding forms part of the £1.6 billion UKRI AI Strategy, which aims to strengthen the UK’s leadership in AI over the next four years.

FOLLOW US ON SOCIAL MEDIA

Most small and mid-sized businesses (SMBs) are grappling with networks that have evolved one tool at a time, as new vendors were added to solve individual challenges as they emerged.

The result is a patchwork of solutions, from wireless access points and firewalls to monitoring tools, bolted together with little strategic planning. Each one has its own interface, licensing model, and operating expenses.

This is costing SMBs time, money, and resources. Too often, the IT infrastructure and management tools businesses rely on today were designed for enterprise environments, making them unnecessarily complex for smaller organizations. Disjointed systems and multiple dashboards make it harder to identify issues, and if a problem arises, which support service should a stretched IT team contact first? Add in the licensing cost for supporting multiple tools from multiple vendors, and the burden quickly grows.

If SMBs are to escape from the network tool maze, they need to think about how all of their systems fit together and work with managed service partners to move toward a simpler, more effective cloud-based IT infrastructure.

From fragmented tools to unified management

No longer reserved for the deep pockets of enterprises, unified, cloud-managed platforms are transforming network operations for SMBs. By bringing together networking, security, and monitoring into one interface, IT teams get a single pane of glass view across their entire environment.

For businesses, this means clearer visibility across devices such as routers and switches, insights into performance and power consumption, and the ability to identify and resolve issues before they impact operations.

Many Managed Service Providers (MSPs) are recommending cloud-based systems because they recognize the operational and business benefits they deliver. With less hardware to oversee, time-consuming on-site maintenance visits are reduced, and the path from purchase to deployment is fast and straightforward.

A more strategic role for MSPs

For MSPs, the shift to the cloud also has advantages.

Remote management and automated software updates save time and costs. The advantages of this can be purely practical – looking after one unified platform will always be easier than overseeing several fragmented ones.

Moving clients to the cloud also allows MSPs to shift from reactive support when something goes wrong to proactive monitoring, optimization, and vulnerability management. They can deliver technical expertise, helping customers to scale seamlessly to meet market demands, and act as an outsourced IT department that transforms IT from being a burden to a strategic advantage.

Helping SMBs adopt cloud-managed platforms delivers another significant benefit - security. As compliance requirements and legal regulations change and cyber threats grow exponentially every year, the built-in cyber frameworks of cloud platforms become essential.

Compliance certifications, automated updates, and integrated monitoring provide a level of protection that was once available primarily to enterprise customers, and without the associated overhead.

Simplifying the path forward

If SMBs are to simplify their networks, they need to move away from tool sprawl and towards a unified platform with security built-in, not bolted on. The right solution will not only deliver improved visibility and simpler operations, but also the features businesses need without unnecessary complexity. This also makes implementation easier as the company grows.

MSPs can help identify the right solution for the customer, but vendors must demonstrate how their platforms work in real-world conditions.

And while even the best unified platform cannot do everything, integrating specialist tools should be seamless through open APIs and interoperability frameworks.

The future of networking is about using fewer tools to achieve more through simpler architecture, centralized control, and systems designed for how SMBs and their partners actually operate.

Secure access service edge (SASE) industry revenues are expected to surge to $24 billion by 2030 as enterprises eye rapid networking modernization.

SASE enables enterprises to integrate networking and security within a single unified framework. Simply put, it allows IT teams better control over access and identity.

Analysis from Dell’Oro Group projects the industry will grow at a 15% compound annual growth rate (CAGR) over the next four years, near-doubling compared to 2025 rates.

The consultancy noted this growth reflects “two complementary sources of strength”. In particular, SD-WAN and security service edge (SSE) modernization efforts are expected to accelerate over the next four years.

On the SD-WAN front, enterprises are ramping up security and lifecycle-led updates while managed service adoption is also growing. This is expected to drive revenues by up to 14% annually by 2030, Dell’Oro Group noted.

Meanwhile, SSE-related growth rests on two converging areas, with firms eyeing stronger digital experience management (DEM) and data security gains.

Mauricio Sanchez, senior director for enterprise security and networking at Dell’Oro Group, said a sharpened focus on these areas “expand(s) the SSE opportunity” by creating new monetization opportunities.

The consultancy expects revenues to exceed $13 billion in this domain by 2030.

“SD-WAN has regained strength because branch refreshes increasingly combine connectivity, security exposure, software subscriptions, and managed services in a common purchase decision,” Sanchez commented.

SASE in the spotlight

Several security providers have rolled out SASE products over the past few months.

In July, Zscaler announced a partnership with Schwarz Digits to launch a zero trust SASE service designed to combat AI-driven threats, bolster cyber resilience, and support data sovereignty efforts for European customers.

Zscaler’s announcement came hot on the heels of a similar move into SASE by Cloudflare. In June, the company unveiled the new Cloudflare One Design Partner designation for managed service providers.

This initiative will help organizations modernize legacy security environments and accelerate adoption of SASE architectures, Cloudflare said at the time.

Dell’Oro Group’s forecast aligns closely with previous research from Gartner. In 2025, the analyst firm projected a CAGR of 26% in SASE across five years, with industry revenue expected to reach $28.5 billion by 2028.

FOLLOW US ON SOCIAL MEDIA

Data center construction in Europe and the Middle East is forging ahead, with the region accounting for five of the ten fastest-scaling global data center markets.

According to new research from Savills, the US is – unsurprisingly – the biggest market in terms of capacity, at around 50GW of live IT power, and has grown by 19% since 2024.

China comes next in terms of existing capacity, with Japan, the UK, Germany, Ireland, and the Netherlands also boasting large and mature markets.

But while they show major absolute scale, they've experienced more modest recent growth thanks to increasing grid access constraints.

Some EMEA markets, meanwhile, are scaling at a rapid pace. The data center markets in Finland and Saudi Arabia have grown by 57% since 2024, for example.

Denmark (38%), France (28%), and Portugal (23%) are all recording significant growth.

Cities are getting in on the action

According to Savills, a host of European cities are capitalizing on the market boom and positioning themselves as go-to locations. Porto, Marseille and Malaga are offering comparatively low-cost power by European standards, Savills noted.

These cites also offer more permissive planning regimes and less competing demand than some of the region’s more established hubs.

Porto combines renewable power with strong connectivity, as well as more room to secure land and capacity.

This pattern extends across Europe, researchers said, with several AI projects being delivered further away from the major metros in locations where a faster route to significant power can be secured.

"For AI-driven data center growth, the Nordics, led by Finland, Denmark and Norway, offer one of the clearest delivery propositions globally, combining low-cost, clean power, cooler operating conditions and fewer physical constraints," said Rupert Duckworth, associate director of Savills' EMEA data center advisory.

"The region has increasingly become a focus for large-scale, AI-oriented development – with significant new campuses now progressing."

Asia markets are growing rapidly

According to Savills, it's Malaysia that's scaled at the highest rate over the last two years, recording growth rates of 132%. This is partly due to its proximity to data processing-hungry but land- and power-constrained Singapore.

Indonesia has grown by 66%, with India and Australia also having grown rapidly to hit a meaningful scale.

"The relative growth of countries such as Malaysia and Indonesia in just a couple of years has been extraordinary, with another tranche of countries not far behind," said Paul Tostevin, head of Savills World Research.

"However, these nations must now look to balance the issues which are sometimes constraining growth in the more mature markets: namely grid access, energy costs, land and water availability and environmental impact.”

Power constraints holding back growth

Power is the biggest factor holding back construction, Savills noted. In particular, grid access, where connection lead times can be far longer than build time.

As a result, developers are increasingly going where grid capacity is available, rather than simply following demand.

"Demand for data centres is not in doubt. But the harder question is where they can actually be built. Answering this will require close collaboration between operators, utilities, governments and local communities, particularly where AI workloads place new pressures on grids and resources," the researchers conclude.

"Climate risk compounds the challenge: higher temperatures, water stress and extreme weather will increasingly shape where capacity can be delivered and run reliably, and communities and governments will be asking whether development will exacerbate existing issues."

FOLLOW US ON SOCIAL MEDIA

One of our biggest gripes about network monitoring software products is their convoluted and confusing licensing schemes. Some charge by sensor, monitored element, volume, or network port; others do it by node or device, and there's also the extra cost of optional add-on modules.

The SolarWinds Observability suites cut through the confusion as you can choose from SaaS cloud-hosted or on-premises versions, and both are licensed purely by the node. We review the SolarWinds Observability Self-Hosted (SWOSH) version which offers three simple subscription plans.

With a monthly cost of around £6 per node, the SWOSH Essentials edition teams up SolarWinds Network Performance Monitor (NPM) module with the Log Analyzer, Server & Application Monitor (SAM), IP Address Manager (IPAM), VoIP & Network Quality Manager (VNQM) and User Device Tracker (UDT). The Advanced edition is well worth considering as for £10 per node/month, this adds the Network Configuration Manager (NCM), NetFlow Traffic Monitor (NTA), Virtualization Manager (VMAN) plus Server Configuration Manager (SCM) and also introduces the new ML-powered anomaly alerting service.

The Premier edition pushes the monthly node price to £13 and augments Advanced with storage and web performance monitoring and an incident response service. It also offers value-based licensing where you can monitor elements such as containers and on-premises virtual machines (VMs) without consuming a full node license.

SolarWinds Observability Self-Hosted review: Easy deployment

SWOSH offers a remarkable range of monitoring components, and all are neatly integrated into the Orion web console. This provides a single pane of glass, with the console's 'My Dashboard' tab presenting a drop-down menu for access to each module and its associated metrics, along with three other menus for alerts and activity, reports, and all settings.

We deployed SWOSH on a Windows Server 2025 Hyper-V VM in the lab with a single routine handling the entire process and taking 40 minutes to download and install all components. As we were evaluating SWOSH, we used the included SQL Server Express database, but for production environments, you'll need to source a full version of SQL Server 2016 or later.

Also, pay particular attention to the minimum hardware requirements. We started our VM with eight Xeon Scalable Gold vCPUs and 16GB of virtual memory but had to increase it to 32GB as we were receiving low memory alerts for the host system.

SolarWinds Observability Self-Hosted review: Voyage of discovery

The Discovery Central page offers wizard-assisted routines to discover cloud entities and network devices. We started with a scan of our lab network and ran through adding our network subnet and providing credentials for SNMP, Windows WMI, and VMware, after which SWOSH reported back in 10 minutes with a complete list of all identified systems.

Cloud monitoring services are provided for AWS, Microsoft Azure and Google Cloud Platform. We chose AWS and found the discovery process simple, as after providing our access and secret keys, SWOSH presented a dashboard of all our resources, including cloud instances and S3 buckets.

Discovery Central also provides a section for configuring the new AI agent. This uses the Platform Connect service to link up with the SolarWinds AIOps cloud service using a choice of three global data locations.

The agent presents a right-hand side menu where you can ask natural language questions about anything on your network or how to achieve something. It's currently in SW1 preview and focuses on the SolarWinds knowledge base documentation and the current state of your network. It impressed us nonetheless, as it came back with detailed answers to questions about, for example, the status of our VMware estate, the most critical nodes and patterns in alert summaries.

SolarWinds Observability Self-Hosted review: The Orion console

The home page presents a bold summary with at-a-glance views of your entire network. Widgets are included for areas such as node, group and application status, active alerts by severity, Top 10 views, hardware status plus anomaly-based alerts, and dashboards can be customised by creating new widgets for specific interests and dragging them to the desired location.

From NPM, you can pull up detailed summaries of all monitored nodes, each assigned a colour-coded status icon, see active alerts, a network Top 10 plus routing insights, and use NetPath to monitor external web locations and see hop-by-hop maps with latency and packet loss details for each step. NTA is another easy one to use as it automatically picked up the NetFlow feeds from our WatchGuard firewall appliance and provided detailed views of detected applications, the top conversations and busiest endpoints.

Accessed from the Home menu, the PerfStack Analysis tool can be used to identify the root cause of network problems by comparing data collected from multiple devices in one dashboard. You'll find the Quality of Experience (QoE) service here too, which uses sensors linked to switch mirror ports to identify, categorise and analyse traffic for over 1,500 applications.

The discovery process automatically identifies running apps, with the SAM module providing heaps of templates for monitoring server and app performance, although some are old enough to be in a museum. For MS365 monitoring, we were pleased to see SolarWinds has replaced its clunky PowerShell queries with API poller templates which are much easier to use.

SolarWinds Observability Self-Hosted review: Hardware health, VMAN, SCM and anomalies

If SWOSH detects a supported server during discovery, it'll assign a hardware health sensor resource that provides details on system temperatures, power consumption, and cooling fan status. It worked for our HPE ProLiant iLO-equipped servers as SolarWinds no longer needs HPE's System Insight Manager (SIM), but Dell fans must continue using the pensionable OpenManage Server Administrator (OMSA) as native support for iDRAC controllers probably won't be implemented until the 2027.2 release.

The Advanced version is ideal for businesses big on virtualisation as it includes the VMAN module. This augments NPM's basic availability checks with full host monitoring, reports on CPU, memory plus datastore usage, views of VM resource use and capacity planning reports.

SCM's hardware and software inventory were found to be accurate, and we used its baselines to keep track of server and application changes on our Windows hosts. The new anomaly alerting service allows you to create custom alerts that go beyond the static thresholds used by NPM. They're simple to create as you provide a name, assign one of five alert levels, apply it to a node, interface or VM, and add conditions such as average CPU load with a threshold of your choosing.

SolarWinds Observability Self-Hosted review: Is it worth it?

SolarWinds Observability Self-Hosted is good value as its cost per node is significantly lower than many competing products and made all the more affordable thanks to the sheer number of monitoring modules included. There are a lot of features to play with so it'll take some time to get it configured to your requirements but the suite is simple to install and easily managed from a single web console.

It's been a busy year for WatchGuard; along with a family of table-top Firebox appliances, it also launched a range of five rack-mount models. Not content to rest on its laurels, WatchGuard now delivers yet another new family of rack-mount appliances with a laser-sharp focus on high performance.

The range comprises three models, with the Firebox M4850 on review representing the entry point and targeting mid-sized businesses yearning for a major speed upgrade without being faced with complex enterprise level prices. All three appliances sport an Intel Granite Rapids-D HCC Xeon 6 SoC (system on chip), with the M4850 endowed with a 12-core model – the M5850 gets 20 cores while the top-dog M6850 receives 24 cores.

Add in 32GB of fast DDR5 memory, and you have some impressive performance claims. The M4850 boasts a high 66.5Gbps raw firewall throughput, 19.1Gbps for full UTM scans with antivirus (AV), the intrusion prevention service (IPS), and application controls all activated, and 8.8Gbps for full scan HTTPS content inspection.

The appliance offers a good range of high-speed network ports, and WatchGuard claims an industry first for a firewall platform. All three Fireboxes have an OCP 3.0 expansion slot with optional modules including a dual-port QSFP28 100GbE card.

WatchGuard Firebox M4850 review: Hardware features

The M4850 presents eight RJ45 Gigabit, four fibre Gigabit SFP, six fibre 10GbE SFP+, and a pair of 25GbE SFP28. The M5850 ups the 25GbE port count to four while the M6850 increases the 10GbE SFP+ port to ten and adds a pair of 100GbE QSFP28.

The rear OCP 3.0 port is the Gen5 x8 variety, and WatchGuard offers a good range of expansion options. Along with the dual 100GbE card, you can choose from quad 2.5GbE or 25GbE SFP28 modules.

Power is handled by dual 600W hot-plug PSUs while cooling is looked after by four (slightly noisy) hot-plug fan modules. An RJ45 port is provided for direct access to the appliance's CLI, with two USB ports available for system backups, restores and support diagnostics.

WatchGuard Firebox M4850 review: Subscription simplification

No need to worry about complex subscriptions and expensive optional add-ons, as WatchGuard offers only two subscriptions. All base Firebox models with a support contract come as standard with firewall, software-defined WAN (SD-WAN), and clientless VPN access portal services enabled.

Choose the Basic Security Suite (BSS) subscription, and you get gateway antivirus (GAV), web filtering, antispam, HTTPS inspection, IPS, application controls, network discovery, and WatchGuard's RED (reputation-enabled defence) cloud-based URL filtering. Moving up to the Total Security Suite (TSS) subscription enables every available security service, including the advanced persistent threat (APT) blocker with cloud sandboxing and ThreatSync XDR for collection, correlation and automated responses for threat events from all Fireboxes.

There's even more as TSS adds WatchGuard's DNSWatch service, which monitors client DNS requests and blocks them from accessing known malicious domains. The IntelligentAV anti-malware service augments the GAV with the Cylance AI-based engine scanning files such as Office documents, Windows portable executables, and PDFs.

You can manage the appliance in standalone mode, but both subscriptions include access to the WatchGuard Cloud portal with the latter extending log and report data retention to 1 year and 30 days respectively – features some other vendors charge extra for. TSS also enabled WatchGuard's Gold support, which provides 24/7 access to technicians, live calls for critical issues, a one-hour SLA for high-level issues, unlimited support cases and advanced hardware replacement.

WatchGuard Firebox M4850 review: Cloud management and Rai

Deploying the M4850 for cloud management is simple as we registered it with our WatchGuard support account and from the WatchGuard Cloud portal, allocated it to our site and chose the cloud management and monitoring option. A wizard helps with basic network configuration and all you need to do next is power the appliance up and leave it to connect to the cloud portal, after which it disables local management access.

The portal can remotely manage all compliant WatchGuard fireboxes, Wi-Fi access points, AuthPoint, EPDR, FireCloud and AI-powered ThreatSync + NDR services. A smart new feature is WatchGuard's Rai (Red AI), which correlates security feeds and telemetry from every cloud-managed product, regardless of their location, and presents a highly informative dashboard with AI-generated daily summaries of all activity.

Primarily aimed at MSPs, Rai is designed to cut through alert smokescreens and streamline problem resolution by automatically investigating and responding to security incidents. Rai impressed us during testing as we have three Fireboxes, EPDR and FireCloud running in the lab, and the dashboard told us everything we needed to know about their security posture, the number of incidents and all automated actions.

WatchGuard Firebox M4850 review: Security features

The M4850 makes the most of its port-fest as it defaults to mixed routing mode, allowing you to segment your network by configuring each port as separate interfaces, each with its own IP address and DHCP services, and designate them as external, trusted, optional, or custom. Other options are drop-in mode, which creates one logical network for all interfaces, and bridge mode if you want to place the appliance between an existing LAN and gateway.

The TSS subscription enables a wealth of security features, but the portal makes light work of configuration as all key services are presented in one page. From the content scanning section, you can enable both AV scanners, the APT blocker and spamBlocker for applying anti-spam policies to inbound SMTP, IMAP and POP3 traffic.

The network blocking section covers botnet detection, IPS, port and site blocks, detection of Tor (The Onion Router) exit points, and default packet handling. WatchGuard's WebBlocker and application control services are both accessed from the content filtering section, with the former offering 166 URL categories that can be allowed, blocked, or set to display a warning page to users.

App controls are equally impressive with nearly 1,300 predefined app and protocol signatures provided, and both services can be applied to a firewall rule by assigning a single action policy. VPN features are abundant with the M4850 supporting 5,000 branch offices and 10,000 mobile VPNs.

WatchGuard Firebox M4850 review: Is it worth it?

Yes. WatchGuard Online lists the Firebox M4850 with a 1-year TSS subscription for £22,610 excluding VAT, rising to £41,646 for a 3-year subscription. This compares very favourably with the competition in this space and WatchGuard's simplified licensing cuts through any subscription confusion.

The M4850 offers plenty of ports for your money with native 25GbE support and the OCP 3.0 expansion slot adding extra levels of flexibility and valuable future-proofing. The Intel Xeon 6 SoC delivers an impressive performance boost, security services are impeccable, and WatchGuard's Rai will appeal to busy MSPs.

WatchGuard Firebox M4850 specifications

At the heart of all of our technology in the 21^st century – from smartphones to computers to servers to manufacturing machines – lies the humble microchip.

These tiny components of integrated circuits could be described as the brain through which everything happens, whether that’s processing data or storing information.

Over the decades, the main goal when designing such chips has always been to make them smaller, more powerful, and faster than ever – and there are many different types created for all sorts of different devices and roles.

For example, logic chips such as central processing units (CUPs), graphics processing units (GPUs), and neural processing units (NPUs) are the key to making something work or happen. Separately, there are also memory chips that store data, for instance flash or random access memory (RAM).

All of these chip functions can also be combined to form systems-on-a-chip (SoCs), a single microchip featuring processing, graphics, and memory.

What on earth then is an ASIC?

The application-specific integrated circuit (ASIC) is a microchip that’s custom-designed and optimised to do a single designated task absolutely perfectly.

This enables it to achieve the sort of speed, power, and efficiency that isn’t possible when a microchip is designed to handle many tasks. Peter Groos, senior AE Group director, at Cadence Design Systems, explains: “They’re optimized for particular, predictable workloads, enabling higher performance and greater energy efficiency when that task is well-defined.”

Who makes ASICs?

ASICs are designed by companies such as Broadcom, Nvidia,Qualcomm, and many others and manufactured by foundries such as TSMC and Samsung.

“These companies collaborate across the value chain, with designers creating chip architectures and foundries fabricating them at advanced process nodes,” Groos adds.

What types of ASICs are in use right now?

The AWS Trainium family of chips from Amazon is an example of an ASIC. These have been specifically designed for training and running AI models and platforms.

Other examples of where an ASIC can be used are in:

• 5G network signal processing and packet routing in telecommunications
• Advanced driver assistance systems (ADAS) in motor vehicles
• Bitcoin Mining or for the creation of other cryptocurrencies
• Face detection or voice recognition in smartphones and other smart devices
• Hearing aids or medical imaging within the healthcare sector

What else is interesting about an ASIC?

ASICs are built on the idea of stripping away everything unnecessary to ensure they are fully focused on executing the single task at hand, however complex it is. But they can also be built to be SoCs, with everything in the one place that’s needed for the one job.

ASICs will often be found in modern data centers and industries where compact and energy-efficient computing is critical. According to BAE Systems, you will also find special radiation-hardened ASICs in nuclear power plants, spacecraft, and inside satellites in orbit.

Where do ASICs sit within AI?

“As AI systems scale, ASICs are becoming increasingly important for efficiency, particularly in data centres and edge deployments,” says Gross.

“The role of ASICs in the AI ecosystem is based on their ability to naturally complement the GPUs that are so critical to AI infrastructure.”

This allows ASICs to play a key role in optimising network traffic and workload distribution within systems, Gross further explains, adding: “ASICs are used because they deliver superior speed and power efficiency compared to general-purpose chips when handling predictable, repetitive workloads.

“In AI, this makes them particularly valuable for tasks where performance-per-watt and cost efficiency are critical, such as large-scale inference and data movement.”

What else is important about ASICs?

There are three different types of ASIC, each with its own clear role: full custom ASIC, semi-custom ASIC, and programmable ASIC.

As the names suggest, the first is built from the ground up, the second will use pre-designed parts, and the third can be programmed for a defined use once it is manufactured.

With such a tiny physical footprint, each ASIC has no wasted action or space onboard; everything serves a defined purpose.

Google's tensor processing units (TPUs) – a type of ASIC – are said by the company to push computing performance by up to 15-30x, while their energy efficiency can rise from 30-80x versus a general-purpose processor of the same level.

Why have ASICs become so important?

Amazon’s entry into ASICs with AWS Trainium shows how vital they are to the future of computing, technology, and AI. This is a battle for supremacy in a multi-billion-dollar race, and the potential efficiencies and cost-savings ASICs can go on to deliver are enticing other big players to join in. Google and Meta, for example, are also investing in their own ASICs.

ASICs’ role in processing data at the edge also makes them vital for smart cities, smart appliances, and autonomous vehicles.

What is the future for ASICs?

A lot of investment is being pumped into ASIC design and manufacture right now. For example, in April 2026 NVIDIA invested $2 billion in Marvell, a company that counts Amazon, Microsoft, Google, and Meta among its customers.

This comes as the entire semiconductor industry is expected to reach $1.6 trillion in revenue by 2030, up from $775 billion in 2024, according to McKinsey.

Broadcom is another benefiting from this surge in interest. Its Custom ASICs are provided to the likes of Google, Meta, OpenAI, and Apple, and Q2 FY2026 revenue for AI semiconductors soared 143% to $10.80 billion year over year. CEO Hock Tan projects AI revenue exceeding $100 billion in 2027.

Amazon, which makes the AWS Inferentia chip family (another custom-made ASIC), is also buoyant about the future of its chips business. Amazon CEO Andy Jassy said in late April 2026 that this part of the company saw 40% quarter-to-quarter growth in Q1 2026.

He also revealed that while its AWS Trainium4 chips won’t roll off the production line until 2027, much of its available inventory is now already reserved. “As best as we can tell, our custom silicon business is now one of the top three data center chip businesses in the world – and the speed at which we’ve gotten here is extraordinary,” he says.

Engineer and tech entrepreneur Marco Riedesser, however, sees another change coming over the horizon, pointing to first-generation hardwired chips from companies like Taalas. These chips put the AI model into the silicon.

“We are moving from a world where we run AI on computers to a world where the chip is the AI,” he explains.

“By removing the need to move data back and forth from memory, ‘hardwired AI’ can run medium-sized models at speeds 10x–50x faster than traditional GPUs.”

Riedesser describes how instead of loading a model like Llama 3 into a chip's memory, you instead “etch the model’s logic” into the metal of the chip itself.

He predicts that “while AWS and Nvidia are building chips to run software, companies like Taalas are fundamentally changing the game”.

Even if data center compute demand falls short of current projections, it still won't be easy for power suppliers to keep up, according to analysis from McKinsey.

With AI infrastructure projects continuing at a rapid pace, the consultancy anticipates that around 75% of power demand growth across the US in the next decade will come from data centers.

Based on current expansion rates, this will require the "equivalent of almost 30 GW of power per year”.

"Whether that growth materializes depends on whether demand for AI compute will last. Several indicators suggest that it will,” the consultancy noted.

In the short term, McKinsey a shortfall is likely in the US. But future demand is hard to predict, with parts of the AI ecosystem exhibiting bubble-like characteristics.

More than half of 2025's new unicorns are AI-focused, including several multibillion-dollar firms without commercialized products. Meanwhile, some AI-native players and infrastructure intermediaries are pursuing aggressive expansion, despite high leverage or negative cash flow.

Broader macroeconomic and geopolitical factors such as tighter capital markets, elevated power prices, or supply chain disruptions could also dampen demand and slow the pace of data center build-outs.

Concerns over compute capacity and infrastructure run counter to current trends in the industry, McKinsey noted, particularly in terms of concrete use-cases. Most organizations deploying AI are reporting good results, the consultancy noted, and there’s no sign of an imminent technological plateau when it comes to LLMs.

"Investments are increasingly tied to contracted capacity utilization and concentrated among scaled, investment-grade data center operators — factors that point to durable, committed growth rather than speculation," the firm said.

"Finally, demand is anchored by sophisticated, well-capitalized hyperscalers, whose capital spending has grown approximately 50 percent annually from 2022 to 2026 while maintaining strong margins and cash flow."

Infrastructure misalignment

McKinsey warned that at the current pace of growth, there's a misalignment between where stakeholders want to build data centers and where generation, energy demand, and grid infrastructure can quickly scale - raising the possibility of energy supply problems in parts of the US.

McKinsey's model suggests data center power demand could grow by around 27% annually between now and 2030, reaching 121 GW of data center IT demand.

However, it looks likely to grow more slowly after that, thanks to a potential saturation in demand for chips and slower AI adoption, as well as a shift from building new data centers to replacing today’s components with more energy-efficient, next-generation chips that add compute per unit of power.

McKinsey pointed out that the massive backlog in load interconnection pipeline requests isn't necessarily a sign of a demand bubble. Many applications are speculative and duplicative, with developers and hyperscalers submitting multiple requests to see which secures power first.

Meanwhile, early-stage developers are trying to secure interconnection rights before selling projects to well-capitalized parties that will do the building and install the IT hardware.

"Beyond 2030, our analyses point to a wide range of outcomes, but several structural dynamics are likely to hold," concluded the firm.

"In addition to the growing role of gas, we expect solar paired with battery storage to sustain strong momentum, given its cost competitiveness, modularity, and speed to deploy, though this is contingent to some degree on stable policy support, interconnection reform, and continued declining costs of storage."

FOLLOW US ON SOCIAL MEDIA

The European Commission is calling for tenders to establish up to seven AI gigafactories across the EU, as part of its efforts to secure European tech sovereignty.

The initiative will be supported by up to €10 billion in EU and national funding, making it the largest public-private partnership in the world for the development of trustworthy AI.

It's aimed at giving businesses of all sizes, academia, and public authorities access to infrastructure for training, inference, and fine-tuning advanced frontier AI models.

It's hoped that the facilities will also act as a catalyst for the European semiconductor ecosystem and stimulate domestic design of processors by providing a predictable, large-scale demand for high-end computing. It’s also hoped that, over time, this will evolve into the manufacture of European-designed AI processors in the EU.

The gigafactories, said the Commission, will combine advanced AI processors, software and cloud technology stacks, high-speed connectivity and energy-efficient data centers, dedicated secure cloud user access environments, and specialized secure AI-oriented support services.

The supporting infrastructure will be environmentally sustainable, particularly in terms of energy and water supply.

The plan was first revealed in February 2025 and an informal call for expression of interest drew 77 proposals, with potential sites in Germany, Italy, France, Poland, Czechia, Denmark, Finland, Greece, Portugal and Spain.

"The opening of the AI gigafactories call represents a milestone in our AI Continent ambitions," said Henna Virkkunen, EVP for tech sovereignty, security, and democracy at the European Commission.

"Access to the raw scale of computing power within AI gigafactories is a strategic necessity for Europe as AI development accelerates. I am pleased to see member states, industry and the Commission coming together to work on these facilities that are key to our technological sovereignty," she added.

The plan will involve two consecutive development phases and two lots, with participating member states matching EU funding.

The first lot will support up to four projects, each eligible for up to €100 million in EU funding in phase one and up to an additional €400 million per project in phase two. Successful projects will have to deploy at least as many of the most advanced AI processors as are currently installed in Europe's most powerful AI factory.

In phase two capacity must triple, with the second lot supporting up to three projects, each eligible for up to €200 million of EU funding in phase one and up to an additional €800 million per project in phase two.

To qualify, projects must feature up to twice as many of the most advanced AI processors as are currently installed in Europe's most powerful AI factory; in phase two, capacity must increase to at least four times that level.

The AI gigafactories can be set up on one site or across several locations in a single member state, but can also include cross-border developments through distributed AI computing facilities.

The call closes on 12 November 2026, and consortia or Special Purpose Vehicles bringing together companies, public entities, investors and other partners can apply.

Decisions are expected by early 2027, with construction starting later that year, and operations expected to start within 18 months.