Neoclouds Challenge the Hyperscalers in Big Bets on AI Infrastructure

AI infrastructure is booming like few other sectors of the tech industry: Hundreds of billions of dollars are currently flowing into data centers, GPUs, networking equipment, and energy capacity to power the training and operation of ever-larger AI models. For a long time, the hyperscalers Amazon, Microsoft, and Google dominated this business — and Meta, too, is building massive capacity of its own with multi-billion-dollar investments. But now a new generation of companies wants to profit from the boom: highly specialized providers that focus entirely on AI compute.

The term for them first appeared in analyst reports in late 2024 and has since become a fixture of tech vocabulary: “Neoclouds” refers to a new class of cloud providers that specialize almost exclusively in renting out GPU compute for AI workloads. Research firm SemiAnalysis, which shaped the framing with its “AI Neocloud Playbook,” defines them as “a new breed of cloud compute provider focused on offering GPU compute rental” — no CRM hosting services, no object storage empires, just GPUs, networking, and an invoice.

The numbers behind the trend are remarkable: According to Synergy Research Group, the neocloud sector generated more than $25 billion in revenue in 2025, with fourth-quarter revenues alone surging 223 percent year over year to $9 billion. By 2031, the market is projected to grow to nearly $400 billion at a compound annual growth rate of 58 percent.

What Sets Neoclouds Apart From Hyperscalers

The traditional hyperscalers — Amazon Web Services, Microsoft Azure, and Google Cloud — offer a broad portfolio of hundreds of services: virtual machines, databases, storage, networking, compliance certifications, and global redundancy across dozens of regions. Neoclouds forgo almost all of that. Their business model centers on a single bottleneck: access to Nvidia GPUs (and increasingly AMD accelerators) for training and inference of AI models.

This gives rise to several structural differences:

Speed over breadth. Neoclouds emerged because AI demand grew faster than the established providers could build capacity. AI labs and enterprise customers needed GPU clusters immediately — and specialized providers could deliver faster, in some cases with modular data centers that are operational in roughly six months instead of two to three years.

Take-or-pay contracts and debt. The typical playbook: sign multi-year offtake agreements with hyperscalers and AI labs, borrow against those contracts (often GPU-collateralized at 60 to 70 percent loan-to-value), and use the proceeds to finance the next stage of expansion. CoreWeave alone has disclosed roughly $8 billion in term loans and delayed-draw facilities.

Customers double as competitors. Curiously, the hyperscalers themselves rank among the neoclouds’ biggest customers. Microsoft has struck commitments worth around $60 billion with CoreWeave, Nebius, and Nscale — partly because such contracts are booked as operating expenses over their lifetime rather than weighing on the company’s own balance sheet as capex. Meta has likewise signed multi-billion-dollar deals with CoreWeave ($35.2 billion) and Nebius (up to $27 billion).

Power as the bottleneck. Competition is increasingly shifting from GPU procurement to energy supply. CoreWeave is targeting 1.7 gigawatts of active power by the end of 2026; Nebius is aiming for up to 1 gigawatt of connected capacity against more than 4 gigawatts of contracted power. Grid operators in the US are already warning that data center plans are straining their load forecasts.

The Key Players

CoreWeave is considered the reference case of the category. The US company went public on the Nasdaq in March 2025 at $40 per share and a valuation of around $23 billion, reached $5 billion in annual revenue in 2025 as the fastest cloud in history to hit that mark, and reported an order backlog (remaining performance obligations) of $99.4 billion as of March 2026. Its customers include OpenAI, Microsoft, Nvidia, Meta — and, since April 2026, Anthropic under a multi-year compute contract. Capex plans for 2026 run as high as $35 billion. The company is not profitable: Depreciation most recently consumed more than half of its revenue.

Nebius, which emerged from the international remnants of Russian tech giant Yandex and operates its flagship data center in Finland, grew revenue by 684 percent to $399 million in the first quarter of 2026 and swung adjusted EBITDA to a $129.5 million profit. Its backlog stands at $46 billion, including a Microsoft deal worth $17.4 to $19.4 billion. Unlike CoreWeave, which relies almost entirely on colocation, Nebius operates more than 75 percent of its contracted capacity in its own data centers. For European investors, the stock is regarded as one of the few liquid ways to invest in AI infrastructure through a European-domiciled, US-listed company. Together with CoreWeave, Nebius was added to the Nasdaq 100 at the end of June 2026.

Together AI pursues a different model: Founded in 2021 and based in San Francisco, the startup positions itself as a developer platform for open-source models, combining a token-based inference API (roughly 30 to 40 percent of revenue) with classic GPU rental. Together historically leased its capacity from providers such as CoreWeave and Lambda, but is increasingly building its own data centers (Maryland live since July 2025, Memphis in preparation). In February 2026, the company reached around $1 billion in annualized revenue according to Sacra — more than three times its rate in mid-2025. Together reportedly closed a funding round of roughly $1 billion at a $7.5 billion valuation, more than double the $3.3 billion from February 2025. Its investors include Nvidia, General Catalyst, Salesforce Ventures, and Aramco’s venture arm Prosperity7.

SpaceX is the most unusual newcomer to the category. Following its acquisition of Elon Musk’s AI company xAI in February 2026, the space giant now leases out compute capacity from its terrestrial data centers: In May, Anthropic secured the entire capacity of the Colossus 1 data center — reportedly for around $1.25 billion per month — with similar deals following with Google and Reflection AI. At its IPO on June 13, 2026 (ticker: SPCX), SpaceX raised $75 billion; the stock closed its first trading day at $160.95, implying a market capitalization of around $2.1 trillion. In parallel, the company is pushing ahead with its plan for orbital data centers via the AI1 satellite (150 kilowatts of peak compute, 70-meter wingspan) and an FCC filing for up to one million satellites — the first prototypes are slated to launch in early 2027. Whether space-based compute can become economically competitive is contested: SemiAnalysis projects that per-unit compute costs in orbit may not reach parity with terrestrial data centers until around 2040.

Behind them, a broad field of private providers is taking shape: Lambda raised more than $1.5 billion in late 2025 and is working toward an IPO. Crusoe, builder of OpenAI’s Stargate campus in Abilene, Texas, was valued at over $10 billion in a $1.375 billion round in October 2025. The UK’s Nscale — Europe’s largest homegrown neocloud provider, with Microsoft, OpenAI, and Nvidia on its investor list — reached a $14.6 billion valuation with a $2 billion Series C in March 2026 and is building, among other projects, Stargate Norway and a 66,000-GPU site in Portugal for Microsoft.

Valuations and Key Metrics at a Glance

As of July 1, 2026. Market capitalizations of listed companies are subject to strong fluctuations; private valuations are based on the most recent funding rounds or media reports. RPO = remaining performance obligations.

The Risks: Debt, Depreciation, Customer Concentration

As impressive as the growth figures are, the risks are just as real. None of the major neoclouds is profitable under GAAP — depreciation on the GPU fleets consumes roughly half of revenue at both CoreWeave and Nebius. Analysts also point to a “refinancing wall”: Tens of billions of dollars in predominantly GPU-collateralized loans from CoreWeave, Nebius, Lambda, Crusoe, and Applied Digital come due between 2026 and 2028 — within a narrow window and among a small, highly correlated circle of lenders.

Then there is customer concentration — a risk that became starkly visible on July 1, 2026: After Bloomberg reported that Meta is exploring its own cloud business under the name “Meta Compute” and could sell off excess AI capacity, shares of Nebius and CoreWeave temporarily plunged by around 15 percent, while IREN lost a good 6 percent. The biggest customer could thus become a competitor — a pattern SpaceX has already demonstrated by marketing its own Colossus capacity. Skeptics among venture investors quip that neoclouds are less technology companies than real estate and power market plays wrapped in an AI story.

Proponents counter: Meta’s willingness to build tens of gigawatts of AI capacity confirms just how massive compute demand remains. IDC analysts see the market heading toward a lasting bifurcation — hyperscalers for the broad cloud business, neoclouds as specialized, long-term AI infrastructure operators. Which of the two readings prevails should become clear in the coming quarters — namely, when it must be proven whether the record backlogs can actually be converted into profitable revenue.