Google Cloud is more confident than ever in its AI strength, from hardware to software

When you attend an international tech conference, you find that the week passes as a blur of announcements and hyped-up keynotes.

Google Cloud Next 2026 was no different, as the primary event at which the hyperscaler, which operates on a scale few companies can claim to, shouts about its biggest achievements and teases all the advancements yet to come.

But as a journalist rushing between all the news briefings, air-conditioned interviews, and corridors packed with 32,000 attendees, I also always walk away with a clearly-formed idea of the overall messaging the company is trying to sell.

In my pre-conference piece, I said that Google Cloud had two key questions to answer this year: why Gemini? And why Google Cloud?

The answer to both seems to be ‘in-house expertise with no lock-in’. Throughout Google Cloud Next, the hyperscaler emphasized not just its unique strengths in hardware, software, and infrastructure, but also its commitment to interoperability.

One of the event’s star moments was Google Cloud’s announcement of its eighth-generation TPUs, 8t and 8i, which offer a generational leap in compute and low latency for AI training and inference.

This is custom-built hardware intended to enable massive AI agent deployment, which traditional infrastructure is unfit to run. It’s a huge victory for Google Cloud’s in-house hardware teams, and what sets Google DeepMind apart from the likes of OpenAI or Anthropic.

Uniquely among AI developers, Google can say its chips were built specifically for Gemini, and vice versa, enabling optimization all the way through its product offerings.

“I'd say uniquely within Google, we're co-designing across this full infrastructure stack,” said Mark Lohmeyer, VP & GM, AI and Computing Infrastructure at Google Cloud, explaining that Google’s hardware teams worked closely with Google DeepMind to establish the performance baselines for its eighth-generation TPUs.

This is Google’s unique advantage: two years out from a product launch, it can gather internal teams around a table to figure out what needs to be done across hardware, infrastructure, and software to make a product possible. While AI competitors like OpenAI and Microsoft are waiting to secure orders from Nvidia, Google is already creating the frameworks.

But while the TPUs are an impressive boast, they’re also merely the driving force behind an ever-growing data infrastructure and software layer that Google Cloud seems determined to keep as open as possible.

As agentic AI becomes the key focus for customers and partners alike, Google Cloud appears to be banking on the magnetic appeal of its home-grown solutions to keep them in its ecosystem, without requiring them to go all-in on its platform to make the most of its AI offerings.

Meeting customer data where it resides

I’m loath to repeat the cliche of good AI requiring good data, but it’s true. Without optimized data resources, LLMs simply won’t deliver the value business leaders want them to, and agents only complicate this picture further.

Each agent deployed in an enterprise environment requires a blend of institutional knowledge and business data – which may or may not be held in Google Cloud – to accomplish its goals.

Google Cloud has addressed the former with Knowledge Catalog, its new universal context engine for enterprise users, which aggregates and enriches data for agents and allows them to search enterprise resources with precision.

“In the past, a file landing in your storage sat passive, waiting for a pipeline,” explained Karthik Narain, chief product and business officer at Google Cloud, during the opening keynote.

“Now, the second that image or PDF hits Google Cloud Storage, they are instantly tagged, enriched, and made agent ready. Zero manual data engineering. Powered natively by Gemini, the Knowledge Catalog goes even deeper, reading files autonomously, extracting entities, mapping relationships, and learning about your unique business semantics.

“When an agent hears net revenue or risk, it understands the exact meaning.”

When agents need to access data held in another cloud environment or third-party data source, the norm up to now has been to use connectors. However, these have to be manually created and are brittle as they’re dependent on APIs that businesses must maintain.

Yasmeen Ahmad, managing director, Product Management, Data & AI Cloud at Google Cloud, told ITPro that access to data is the “critical piece” for enterprise AI, particularly as customers look to scale AI agents.

“So the biggest thing we heard from customers is they want to be able to leverage all of Google's amazing AI capabilities that now we're releasing every week,” Ahmad told ITPro, explaining that the new Cross-cloud Lakehouse offering fills this need.

This is enabled by new technologies such as Cross-Cloud Interconnect, a new Google-managed, high-bandwidth network that connects Google Cloud data to another cloud service provider, Cross-cloud Lakehouse helps enterprises move data with much lower latency and more reliability than with data interconnectors.

Ahmad added that with the new Apache Iceberg standard, Cross-cloud Lakehouse conforms to a common data language that allows customers and partners to expand its access to a wide range of ecosystems.

“So that's enabling us to really connect data, not just across clouds, but across data platforms,” Ahmad said.

“So you heard with our cross Cross-cloud Lakehouse, we can connect into DataBricks through Unity Catalog, Snowflake, Polaris, Amazon S3, so we're able to create this unified data fabric, and then that means customers can run all of the amazing AI agentic capabilities without having to move data and be in the old world of connectors.”

Little said on sustainability

A glaring omission from Google Cloud Next 2026 was any real acknowledgement of the environmental impacts of expanding data centers and AI infrastructure.

Take this year’s TPU 8t and 8i announcement for example, which Google Cloud says offer twice the performance per watt of their predecessor. The same announcement mentioned that 8t is cooled with fourth generation liquid cooling – though when I asked Mark Lohmeyer, VP & GM, AI and Computing Infrastructure at Google Cloud, he was unable to provide exact details on what this entails.

We’ve also had precious little news on where TPU 8t and 8i will be deployed first, though there’s a clue in Anthropic’s recent announcement that it will work with Google and Broadcom to deploy “multiple gigawatts” of TPUs in 2027, primarily in the USA.

Looking to Jevon’s paradox, which holds that efficiency improvements actually cause an increase in a resource’s use, I’d predict that Google will see emissions rise, not fall, with TPU deployment.

It’s clear that with the energy efficiency improvements of the latest TPUs, Google is recognizing that customers need to be able to run AI agents at scale without raising their energy bills to ruinous levels. But there’s a gap between the efficiencies Google Cloud claims and the scale of the clusters it’s building.

We don’t actually need to speculate here, given that hyperscalers have begun to measure new data center projects in power capacity rather than chip numbers. A gigawatt of TPUs requires a gigawatt of firm power generation regardless of performance per watt, which in the US tends to mean gas power and increased emissions.

More detail on this would be welcome, especially as Google publicly calls for stepped up power generation to meet its AI goals.

A broad portfolio

Google Cloud obviously sees interoperability and customer choice as a key concern.

Google’s relationship with Anthropic is a prime example. On paper, these firms are competing, with Google DeepMind working to produce Gemini models that can outperform the likes of Claude Opus 4.7. In practice, of course, Google is a major investor in the AI startup and just days ago Anthropic revealed Google would invest up to $40 billion in the near future.

It seems to me that Google sees Anthropic as its most useful side bet, a frontier lab that builds capable models customers already rely on like Claude Code and that partly relies on TPU access for its continued success.

This is a win-win for Google, which either continues to innovate and deliver the best coding models via future Gemini releases or secures customers access to the best in class via Claude releases. Either way, it avoids competition complaints and continues to meet customer expectations.

Openness is also a major factor in Google’s $32 billion acquisition of Wiz, with whom it is now expanding cloud security across popular enterprise offerings. This includes AWS Agentcore, Cloudflare AI Security for Apps, Databricks, Darktrace, Microsoft Azure, Oracle Cloud, Salesforce Agents, SAP, Vercel… the list goes on.

For example, Google offers a wide range of Nvidia products to meet specific customer needs, including the new Vera Rubin NVL72 when it launches later this year.

“Nvidia is a phenomenal partner of ours, and our customers have choice, and that is the entire reason why we built the stack we did,” said Chris Sakalosky, VP, Strategic Industries at Google Cloud.

“It's built and it's integrated for the best price performance, because that stack is running literally every billion and 2 billion user service that Google provides, from YouTube, to ads, to search. So we bring that price performance capability to our enterprise customers, while at the same time, we recognize that there's more than just our TPUs in the marketplace, and there are more uses.”

More than anything else, Google Cloud owns its own destiny. Across hardware, infrastructure, and software it has ensured that it has little in the way of third-party dependencies and can choose its partnerships with care.

It’s easier said than done, but if Google Cloud can up the ante and begin to deliver the best models, it can enjoy a reputation for openness while also reaping the rewards of massive customer spend.

With Gemini 4 likely to be announced at its I/O event in May, and its newest TPUs allowing for truly competitive inference and much faster model development, the future looks bright indeed for Google Cloud.