Dawn of the neoclouds — the rise of a new class of hyperscaler is here. But can they dominate the AI era?

When Meta announced plans in July to sell excess compute, a cohort of new-era businesses known as 'neo-clouds' dropped in valuation like a stone. The market deemed this a threat to their business models at the time; their entire raison d'être is about providing the infrastructure to power today and tomorrow's AI workloads. But they're so embryonic that they're barely breathing in a world in which hyperscalers already dominate.

In hindsight, however, it further validated the thesis behind their business models given the shortage of both compute and energy – with Meta attempting to tap into the AI-centric future upon which these businesses have hinged their futures.

The executives leading these so-called neoclouds, which have emerged as big kingmakers in the last few years, say their growing footprint of data centers is essential to the AI buildout. But can they shake the stranglehold of the cloud giants?

What are neoclouds and how do they differ from hyperscalers?

The cloud computing boom in the early 21st century led to a period of domination for businesses such as Amazon's AWS, Google Cloud, and Microsoft's Azure, among others. These companies, to all intents and purposes, are the backbone of modern enterprises and startups alike. But as we've moved into the late 2020s, it's become clear that the next era of computing may rely on an entirely different model of cloud availability. That's where the neoclouds come in.

"A neocloud sells GPU capacity. That is the product, and the narrowness of it is deliberate," tech industry veteran and cofounder of Remote.it Ryo Koyama, tells ITPro.

"Reproducing what a hyperscaler has built, meaning the managed services, the global regions, the storage tiers, the identity systems, takes a decade and an enormous amount of capital. No new entrant can fund that and compete on price at the same time. So neoclouds concentrate on the one input hyperscalers currently cannot supply in the volume the market wants, which is accelerated compute. They are filling a real shortage, and that is genuine value today."

Generally, they provision mostly GPU-centric compute for AI workloads and HPC via a massive data center that they construct or colocate and, in many cases, also energize that land rather than plugging into the grid. But not all neoclouds are the same, according to Dave McCarthy, group vice president, enterprise infrastructure at IDC.

"Some are functioning solely as bare-metal GPU wholesalers,” he says. “The strategic players, on the other hand, are moving up the stack — building out the software, high-performance data storage, and orchestration layers required to address enterprise workloads."

Where did neoclouds originate?

Although there are many different AI chips out there, the long-term focus on GPUs and their importance in the AI training process has positioned one cohort of businesses incredibly above all others: bitcoin miners.

Key names in this space include Applied Digital, Cipher Digital, Core Scientific, Hut 8 Corp, IREN, and TeraWulf. There's also Bit Digital, CleanSpark, Keel Infrastructure, MARA Holdings, Riot Platforms, and so many others. Many of these names are looking to distance themselves from their Bitcoin-centric business models so that new clients can power the next generation of AI models.

"A significant share of this capacity was built by operators who learned dense power and cooling economics during the crypto mining cycle, and by colocation providers who already had the buildings, the substations, and the utility relationships," says Koyama.

"When demand for accelerated compute arrived, those were the only people in the market who already knew how to get eighty kilowatts into a rack and take the heat back out. That is a real skill, and it takes years to build."

Many of the names in this space pivoted from Bitcoin after realizing that their high-density infrastructure and hardware engineering skills perfectly matched the demands of the AI boom.

It's not quite on the scale of the shift that Allbirds performed in April – shifting from luxury shoemaking to AI data center provisioning – but certainly an opportune one given the infrastructure and embedded expertise of these entities.

"Others are entirely new venture-backed entities that capitalized on structural GPU shortages," adds McCarthy.

"But origins matter less than the destination. What started as a quick supply-chain fix must now mature; survival requires abandoning the crypto-miner mindset to deliver secure, multi-tenant enterprise reliability."

What role will these companies play in our AI future?

Our progress with AI faces many bottlenecks, with compute, infrastructure, and energy the major areas. These businesses absorbed a shortage that emerged in the last few years, when demand far outstripped supply and a huge degree of model training was based in these data centers.

But the demand has no signs of slowing – and, on both energy and infrastructure, there simply isn't enough capacity to support the ambitions of frontier AI labs. By 2030, global energy consumption by data centers earmarked to run AI could reach 945TWh, according to the International Energy Agency (IEA) — amounting to 3% of global energy supply.

Currently, this figure stands at up to 500TWh. This is why many of these businesses haven't even gotten started, with their plans dependent on projects that may not come to fruition for many years to come. That said, McCarthy sees them as essential to the future of AI.

"Right now, they are at the epicenter of the massive shift from AI model training to real-time inference scaling. These providers are deploying the high-density power environments that legacy data centers simply cannot support," he says.

"Moving forward, their true value will be in orchestrating heterogeneous compute. By combining custom silicon, specialized GPUs, and scale-out network-attached storage, neoclouds have the potential to provide the raw horsepower required for the agentic AI era."

There is, however, a massive question mark over their fundamental business models. How exactly are they funding this massive infrastructure and energy buildout? Some of it likely through the deals they sign with clients, and some of it likely coming from raising equity through the issuance of new shares. Oh, and debt – plenty of it, at that.

"The capital behind a lot of this capacity assumes a resale market for used accelerators that has never been tested at this scale," says Koyama. "It is an assumption, and it is load-bearing. Depreciation schedules on this hardware were written by people who needed the math to work.

If a generation of accelerators turns over faster than those schedules assume, the residual value underwriting the debt is not there. Nobody really knows what a three-year-old GPU is worth in a market where everyone is selling one at the same time, because that market has not happened yet."

McCarthy describes the bare-metal business model as "incredibly fragile," with many operating against forces such as depreciation and the circular financing models that have drawn fierce criticism from many, including the writer Ed Zitreon.

As such, any softening in demand translates to price volatility for existing customers. That's why he feels that "consolidation" in the industry is inevitable, with the strongest businesses that survive the turbulence holding the most resilient portfolios, with additional layers like software and orchestration that can address complex enterprise IT needs.

That said, Koyama is broadly optimistic about their future and capability to withstand immediate turbulence in the wider context of expanding AI usage.

"No one under 35 knows the world without the internet,” he says. “[No one] under 25 [is the same] without smartphones; Under 15, they will only understand the world with AI. GPU compute usage will only expand."