The numbers sound like a typo. Microsoft has committed approximately $60 billion to neoclouds. Meta has pledged $35.2 billion to CoreWeave alone, then added $21 billion more, then signed a separate deal with Nebius worth up to $27 billion. Add OpenAI and Anthropic’s undisclosed compute contracts, and you get total hyperscaler commitments to two companies that did not exist a decade ago that exceed $145 billion — roughly 90% of AWS’s trailing twelve-month revenue.
CoreWeave’s actual revenue in FY2026 is estimated at $12.6 billion. Nebius is expected to do $3.4 billion. These commitments are not matched by revenue; they are an order of magnitude larger than current sales. Something structurally unusual is happening in AI infrastructure financing, and if you are an engineer who relies on cloud GPU capacity — or who works at a company that depends on AI API providers that rely on this capacity — you should understand what it is.
The short version: Nvidia is investing in the companies that buy its GPUs, those companies are using that investment to buy more Nvidia GPUs, and the entire stack is being underwritten by hyperscaler contracts that allow Microsoft and Meta to shift capital expenditure off their balance sheets. I am going to walk through each layer of this structure, explain why it exists, and tell you what it means for the durability of the GPU buildout that every AI application currently depends on.
What a Neocloud Actually Is, and Why Hyperscalers Need One
CoreWeave and Nebius are the two most prominent public “neoclouds” — a category of infrastructure provider that sits between Nvidia and the hyperscalers. They are not chip designers. They are not hyperscalers. They are GPU cluster operators that specialize in acquiring the newest Nvidia hardware, deploying it fast, and offering it to large buyers under multi-year capacity contracts.
The core value proposition has two parts. First, speed of deployment. JLL, the commercial real estate firm that tracks data center construction, published analysis noting that neoclouds can deploy high-density GPU infrastructure within months, compared to multi-year builds for hyperscale data centers. CoreWeave claims it can provision Nvidia GB200 NVL72-based instances and make them generally available within two weeks of receiving hardware from Nvidia. That is not a marketing claim — CoreWeave was genuinely the first cloud provider to make those instances available at the start of 2026, and it was the first to have a Vera Rubin system running at the start of June. Speed of access to the newest Nvidia generation is the central product.
Second, utilization efficiency. The model FLOPs utilization (MFU) metric measures how much of a GPU’s theoretical compute capacity is actually being used for productive work. A GPU might report 100% utilization in the standard sense — meaning a kernel is always executing — while only using 30% to 40% of its available parallelism. This “MFU gap” of 30–40% is common in large clusters, and it represents expensive hardware sitting partially idle. CoreWeave layers Kubernetes-based scheduling through its CoreWeave Kubernetes Service (CKS), mixed training-inference cluster optimization through a service called SUNK, and high-speed model loading through CoreWeave Tensorizer. The claim is that these software layers close the MFU gap better than a hyperscaler’s internal tooling.
For Microsoft and Meta, working with a neocloud solves a real operational problem: they cannot build data center capacity fast enough through internal construction to keep pace with model training and inference demand. Contracting with CoreWeave or Nebius lets them access GPU capacity that is already racked, powered, and networked — on a timeline that their own facilities teams cannot match.
The Capital Structure Problem That Nobody Talks About in the Benchmark Posts
Here is the part of this story that most GPU benchmark write-ups skip. CoreWeave and Nebius have committed to deploying enormous amounts of infrastructure. CoreWeave has contracted 3.5 GW of power capacity and is targeting 1.7 GW of active power by end of 2026. Nebius has contracted 3.5 GW as well and is targeting 800 MW to 1 GW of connected power. The gap between contracted power and active power represents a construction pipeline that will cost tens of billions of dollars to complete.
Neither company has the cash flow of a hyperscaler. They are not AWS, which generated $39 billion in operating income in 2025. They are companies with rapidly growing revenue and soaring debt loads. The question of how they finance the gap between current cash flow and future capital requirements is not a footnote — it is the central structural question for anyone thinking about whether this infrastructure buildout is durable.
The answer is a layered structure that is, at minimum, unusual, and at maximum, circular in ways that create systemic fragility.
Layer One: GPU-Backed Debt
CoreWeave and Nebius both finance infrastructure buildouts using GPU clusters as collateral. Banks and private credit funds lend against the value of the Nvidia hardware — the argument being that Nvidia GPUs are liquid assets with a functioning secondary market and predictable depreciation curves. As long as Nvidia hardware retains value, the debt is secured.
This structure has a hidden dependency: it works only if the value of the collateral stays above the loan-to-value thresholds in the debt covenants. Nvidia H100 GPUs that cost $30,000–$40,000 each in 2023 have already dropped substantially in secondary market value as H200 and GB200 hardware entered production. If the pace of new Nvidia architecture releases accelerates — which Nvidia’s roadmap suggests it will — the depreciation curve on the collateral could steepen faster than the debt is retired.
This is not a theoretical risk. It is baked into the basic logic of GPU-backed lending. The lenders know this, which is why the debt terms typically include hyperscaler contracts as an additional underwriting factor. The actual credit quality of a CoreWeave loan is a function of both the GPU value and the strength of the Microsoft or Meta contract backing the revenue stream that services the debt.
Layer Two: Hyperscaler Contracts as Off-Balance-Sheet Capex
When Microsoft commits $60 billion across CoreWeave, Nebius, and other neocloud providers, it does not appear as capital expenditure on Microsoft’s balance sheet in the same way a data center it owns would. The commitments are operating expense contracts — pay-as-you-go compute agreements at scale, structured as capacity reservations rather than asset purchases.
This accounting treatment is partly why hyperscalers like these deals. Microsoft, Meta, and Google are all facing increasing scrutiny over capex trajectories that have reached $60–80 billion per year across each company. Shifting infrastructure spending to opex via neocloud contracts does not eliminate the economic cost, but it changes how it appears in financial statements and, importantly, what happens if demand softens: an unused capacity reservation is a contractual liability, but it is a different kind of balance sheet risk than owning a half-built data center.
For CoreWeave and Nebius, these contracts are the essential underwriting for their entire capital structure. Without long-term hyperscaler commitments, the GPU-backed debt does not get done, the equity raises do not close at current valuations, and the construction pipeline stops. The hyperscaler contracts are not just customer agreements — they are the collateral behind the collateral.
Layer Three: Nvidia’s Direct Investment
This is where the structure becomes genuinely circular. Nvidia has invested $2 billion each in CoreWeave and Nebius. Under these investment agreements, CoreWeave and Nebius have each committed to deploying more than 5 GW of data center capacity using Nvidia hardware by 2030.
Let me be direct about what this means: Nvidia is investing in companies that have contractually committed to buy Nvidia’s hardware, using money that in part comes from Nvidia. The investment makes CoreWeave and Nebius more creditworthy, which makes their GPU-backed debt easier to issue, which funds the purchase of more Nvidia GPUs, which shows up as Nvidia data center segment revenue, which supports Nvidia’s stock price, which is part of why Nvidia can make these investments in the first place.
I am not claiming this is illegal or even necessarily unusual by the standards of how tech ecosystems are built. Qualcomm has similar equity relationships with device manufacturers. Intel has historically done strategic investments in companies that accelerate x86 adoption. The structure is not unprecedented. But it is important to be clear-eyed about what it is: a set of interlocking financial relationships that make it difficult to determine whether the demand for GPU compute is “real” end-user demand or, to some degree, a demand signal that Nvidia itself is helping to create by financing the infrastructure layer that generates that demand.
What “Real” Demand Looks Like vs. What This Looks Like
I have run inference workloads on CoreWeave clusters, and the hardware quality and networking topology are genuinely excellent. The InfiniBand fabric, the NVLink configurations in the larger pod sizes, the actual measured latency and bandwidth between GPU nodes — these are not marketing claims. CoreWeave’s infrastructure engineers know what they are doing, and the product is real.
The question is not whether the infrastructure is real. The question is whether the utilization that justifies the entire capital structure is real at the claimed levels. And here the picture gets murkier.
CoreWeave’s argument about MFU gap advantages is credible in the context of large, specialized training workloads for customers like OpenAI or Anthropic who are running the same model architecture continuously at extreme scale. For those workloads, the customization that CKS and SUNK enable probably does translate to meaningful efficiency gains. Whether those gains persist across the broader customer base — the mid-size enterprises, startups, and research institutions that fill in the utilization gaps between major hyperscaler contracts — is less clear from publicly available data.
Hyperscalers’ own statements about why they use neoclouds are revealing. When Microsoft executives describe the neocloud relationship, they consistently emphasize two factors: speed of access to the newest GPUs, and the ability to scale AI capacity without the multi-year lag of internal data center construction. Neither of these is primarily a story about utilization efficiency. They are stories about access and timeline. The utilization argument is a secondary factor.
If the primary driver of neocloud demand is access to new Nvidia hardware on a fast timeline, then the neocloud business model is durable only as long as Nvidia is releasing new hardware that hyperscalers urgently need and cannot get fast enough through their own internal builds. That is currently true. Whether it remains true as hyperscaler data center construction timelines shrink — Google and Microsoft are both investing heavily in reducing internal construction speed — is the fundamental long-term demand question.
The Vera Rubin Moment: Why Speed to Market Is the Actual Moat
CoreWeave’s deployment of Vera Rubin systems in June 2026 is the clearest data point about where the neocloud moat actually lives. Vera Rubin (NVL576) is Nvidia’s newest architecture, announced at GTC 2025 and beginning to ship to lead customers in 2026. CoreWeave had live customer workloads running on Vera Rubin before any hyperscaler made the same hardware publicly available.
For an AI lab doing frontier model training, being able to access a new GPU architecture two to four months before it is available on AWS or Azure is not a marginal advantage. Training runs at the frontier take weeks to months. If a lab can start its next training run on Vera Rubin hardware in June and its competitor has to wait until August because it is sourcing through a hyperscaler, that is a meaningful competitive gap during a period when foundation model performance is still improving at a pace where a few months of training time can change the capability ranking.
This is the actual moat. Not the Kubernetes scheduling. Not the MFU gap. The moat is that Nvidia gives preferential hardware allocation to CoreWeave and Nebius, and in return those companies have made $2 billion equity investments that have come from Nvidia’s balance sheet. The circular financing is not just a risk factor — it is also the mechanism by which the core product advantage is maintained.
This creates a dependency structure that is worth being explicit about if you are an engineer making infrastructure decisions:
- If Nvidia continues to release new hardware generations on its current cadence and continues to prioritize CoreWeave/Nebius in allocation, the moat remains intact.
- If Nvidia’s hardware cadence slows, or if AMD MI450 or custom silicon from Google or Amazon becomes genuinely competitive for frontier training, the preferential allocation advantage weakens and with it the central justification for the premium pricing and the capital structure.
- If Nvidia’s stock falls significantly — whether from trade restrictions, competitive pressure, or a general AI capex correction — the equity investment relationships become more complicated, and the GPU-backed debt covenants could come under pressure simultaneously.
These are correlated risks. They are not independent. A scenario in which Nvidia faces competitive pressure on hardware is also a scenario in which GPU collateral values fall and the neocloud debt structure comes under stress. The circular financing means the risks are coupled in ways that a linear supply chain would not be.
The Numbers That Should Make Engineers Think Twice
Let me put some specific numbers on the exposure that currently exists in the system.
CoreWeave’s estimated FY2026 revenue is $12.6 billion. Against this, it has approximately $12 billion in debt as of its last disclosed filing, with additional debt being raised to fund ongoing construction. The company is not yet cash-flow positive on a GAAP basis. Its path to cash-flow positivity depends on a combination of factors: continued rapid revenue growth from existing contracts being activated as infrastructure comes online, stable or improving GPU utilization rates, and the ability to roll and expand debt at current interest rates.
Nebius is earlier stage: FY2026 revenue of approximately $3.4 billion, with aggressive construction targets for its European data center footprint. Nebius has the advantage of Yandex’s original engineering team and a European market position that is somewhat differentiated from CoreWeave’s primarily North American focus, but the structural capital challenges are similar.
The combined contracted power footprint — 7 GW between the two companies — represents a future construction obligation in the tens of billions of dollars. The construction pipeline is substantially larger than either company’s current revenue. This is not, by itself, a red flag — hyperscaler data centers routinely have construction pipelines that dwarf current revenue. But hyperscalers fund those pipelines from free cash flow and investment-grade bond issuance. CoreWeave and Nebius are funding theirs through a combination of GPU-backed debt, hyperscaler prepayments, and equity from sources including Nvidia.
What Engineers Choosing Cloud GPU Providers Should Actually Know
I have been asked by several engineering teams whether the CoreWeave circular financing story should change their infrastructure decisions. Here is my honest assessment.
In the near term — the next 12 to 18 months — the concentration of hyperscaler commitments creates substantial revenue visibility for CoreWeave and Nebius. Microsoft and Meta are not going to unwind $122 billion in commitments because of a concern about circular financing. The contracts exist, the hardware is deploying, and the revenue is real. If you need GPU capacity for a large workload and CoreWeave’s pricing and network topology are right for your use case, the financing structure is not a reason to avoid them today.
The medium-term picture — 2027 to 2029 — is where the structural questions matter. By 2027, hyperscalers will have invested several hundred billion dollars in their own data center construction, much of which will be coming online. The “fast deployment” advantage of neoclouds will narrow. The question is whether CoreWeave and Nebius have used the current window to build software differentiation (utilization tooling, networking software, model serving optimization) that justifies continued premium pricing once the raw hardware access advantage narrows.
For engineers making multi-year commitments to AI infrastructure platforms — whether as customers, as architects designing systems that depend on a specific GPU provider’s APIs, or as companies evaluating whether to run on neocloud vs. hyperscaler vs. self-hosted — the relevant question is not “is CoreWeave solvent today?” It is “what does my application dependency look like if this particular infrastructure layer goes through financial stress in 2027 or 2028?”
Having a migration path matters. Running on multiple providers matters. Avoiding deep API-level integration with a single neocloud’s proprietary tooling — CKS, SUNK, Tensorizer — matters if you are worried about lock-in risk that would make it expensive to migrate if pricing or availability changes. These are not exotic concerns. They are standard resilience thinking applied to an infrastructure layer that currently has unusual financial concentration.
The Broader Pattern: AI Infrastructure as Synthetic Demand Creation
The circular financing story is interesting on its own terms. But it also illustrates a broader pattern in the current AI infrastructure buildout that deserves explicit attention.
The demand for GPU compute that is driving the neocloud buildout is, in part, demand created by the availability of GPU compute. This sounds circular because it is. When OpenAI can access 100,000 H100s at low enough marginal cost on CoreWeave, it runs experiments and trains models that it would not have run at a higher price. Those experiments generate research results that validate the investment in further GPU access, which supports further hardware commitments, which makes CoreWeave a more creditworthy borrower, which makes it easier for Nvidia to argue that the demand for its hardware is durable and growing.
This is not fraud. This is not a Ponzi scheme. It is a legitimate infrastructure investment cycle with real end products — large language models, image generation systems, AI coding assistants — that have demonstrable user value. But it is also a cycle in which the investment itself is creating demand signals that feed back into further investment, and in which the financial relationships between the participants — Nvidia, CoreWeave, Nebius, OpenAI, Microsoft, Meta — are interlocked in ways that make it genuinely difficult to measure where “fundamental demand” ends and “financing-driven demand” begins.
I spent a week earlier this year going through CoreWeave’s S-1 and Nebius’s annual report in detail, specifically looking at how they characterize the demand signals that justify their capital structure. Both documents are candid about the hyperscaler concentration risk. CoreWeave discloses that Microsoft is a dominant customer. Nebius discloses that Meta commitments are central to its buildout thesis. Neither company pretends that the demand is diversified across thousands of small customers in the way that, say, AWS’s demand is. The concentration is explicit.
What is less explicit — and what the circular Nvidia investment relationships make harder to evaluate — is the degree to which the hyperscaler commitment sizes themselves are influenced by the availability of Nvidia-backed financing for the neoclouds. If CoreWeave were not able to access GPU-backed debt and Nvidia equity, it would deploy less capacity, which would mean Microsoft would sign smaller capacity commitments, which would mean less revenue to service less debt, which would mean a smaller CoreWeave. The current scale of the system exists partly because the financing structure makes it possible, not entirely because the end-user demand requires it at that scale today.
That gap between “what the financing makes possible” and “what end-user demand requires right now” is the central risk to watch. As long as AI application demand is growing faster than the infrastructure buildout, the gap closes and the structure is self-sustaining. If application demand growth slows — whether because of model capability plateaus, regulatory friction, or simply because the most obvious use cases get saturated — the infrastructure overhang becomes visible and the financial relationships that have been masked by rapid growth become exposed.
Measuring the Actual Risk: What the Numbers Say About Debt Serviceability
To make the structural risk concrete, let me work through what CoreWeave’s debt serviceability actually looks like under a mild stress scenario. I will use public figures where available and reasonable assumptions where not.
CoreWeave entered 2026 with approximately $12 billion in debt. At a blended interest rate of around 8–9% (GPU-backed debt with private credit at current rates), annual interest expense runs approximately $960 million to $1.08 billion. Against FY2026 estimated revenue of $12.6 billion and assuming gross margins of 55–60% (consistent with its disclosed margins as of its S-1 period), gross profit is approximately $6.9–7.6 billion. EBITDA, after significant depreciation on the GPU fleet and operating expenses, is harder to estimate precisely, but the company is not generating free cash flow sufficient to service its debt from operations alone in the near term. The capital raise cadence — repeated equity and debt issuances — is the mechanism by which the construction pipeline is funded.
This is not inherently dangerous in a growth infrastructure business. Amazon Web Services ran with significant net losses for years before becoming the dominant profit center of the company. The question is whether the growth trajectory is fast enough, and durable enough, to reach a self-funding point before the capital markets or collateral conditions change.
The specific stress scenario that concerns me most is a simultaneous softening of hyperscaler AI capex and a tightening of the GPU secondary market. If Microsoft or Meta were to defer or reduce their neocloud commitments by even 20–30% — not cancel, just slow the ramp — CoreWeave’s revenue growth would fall significantly below current analyst projections. At the same time, if Nvidia releases Vera Rubin+ and Blackwell Ultra on a compressed timeline, the H100 and H200 collateral that backs a substantial fraction of the GPU-backed debt would see accelerated depreciation.
Neither of these scenarios is a prediction. Both of them are plausible within a two-to-three year window. And the circular financing means they would arrive together rather than independently: a slowdown in hyperscaler AI spending is also a scenario in which enthusiasm for new Nvidia hardware procurement moderates, which affects Nvidia’s revenue, which affects Nvidia’s ability to make strategic investments, which affects the creditworthiness signaling that the Nvidia equity relationships provide.
The AMD Question That the Circular Financing Obscures
There is another risk that the current structure systematically underweights: AMD’s MI300X and MI350 lineup. I have spent time with MI300X clusters and the honest engineering assessment is that for inference workloads — particularly large context window inference and mixture-of-experts architectures — the MI300X’s 192 GB HBM3 memory pool is genuinely competitive with H100 for a meaningful fraction of production inference scenarios.
The CoreWeave and Nebius competitive advantage depends substantially on preferential Nvidia hardware access. This is not an advantage if AMD closes the hardware gap to within 15–20% on the most common inference workloads, because at that point the “access to newest Nvidia generation” argument weakens relative to the pricing and availability story that AMD is aggressively pushing with hyperscalers. Google’s TPU v5 and Amazon’s Trainium 2 add further diversity to the custom silicon landscape.
The circular financing relationship between Nvidia and the neoclouds creates a structural bias in which both Nvidia and the neoclouds have strong incentives to present the Nvidia hardware advantage as larger and more durable than the underlying technical benchmarks necessarily support. I am not saying the benchmarks are misleading — Nvidia’s GPU performance leadership on training workloads is real and the H200/GB200 generation represents a genuine step function in performance per watt. But the financial relationships between the parties create an environment in which independent third-party benchmarking matters more than it otherwise would. Engineers evaluating GPU procurement should treat vendor-associated benchmarks, including those published by or in partnership with CoreWeave and Nebius, with the same skepticism they would apply to any supplier’s performance claims about their own products.
The Engineering Takeaway
I am not writing this as a financial warning about whether to buy NVDA or CRWV or NBIS. I am writing it because the infrastructure layer that a large fraction of the AI engineering ecosystem currently depends on has a financial structure that creates correlated risks that most engineers have not looked at carefully.
When you run a training job on CoreWeave, or use an API that runs on Nebius infrastructure, or build an application that ultimately depends on compute that is financed through GPU-backed debt backed by hyperscaler contracts backstopped in part by Nvidia’s strategic investments — you are embedded in a financial structure, not just a technical one. That structure is currently healthy. The commitments are real, the revenue is growing, and the hardware is excellent.
But “currently healthy” and “structurally robust under stress” are different things. The circular financing between Nvidia, the neoclouds, and the hyperscalers is not a secret. It is disclosed in public filings. What is underappreciated is the degree to which these financial relationships mean that a stress event in any one part of the system — a significant drop in Nvidia’s stock, a softening of hyperscaler AI capex, a major shift in GPU market share toward AMD or custom silicon — does not affect that one part in isolation. The interlocking financial relationships mean that stress propagates.
Engineers who understand this are better positioned to make decisions about which infrastructure dependencies to treat as durable and which to treat as requiring a hedge. The technical merits of CoreWeave’s infrastructure are real. The financial structure that underlies that infrastructure’s continued existence is worth understanding with the same rigor you would apply to evaluating any other dependency in your system.
The GPU boom is real. The circular financing that is partly powering it is also real. Both things can be true at the same time.
Discussion