The GPU Rental Arbitrage: CoreWeave Hit $50 Billion by Reselling Nvidia's Chips. The Margins Are Not What You Think.

The GPU-as-a-Service market is worth an estimated $5.7 billion to $8.2 billion in 2025, depending on which research firm you ask. Projections for 2030 range from $26 billion to $50 billion. Over 130 active GPUaaS companies operate globally. The sector has a name — "neoclouds" — and a poster child: CoreWeave, a company that went from Ethereum mining to a $50 billion public market valuation in under three years.

The business model sounds simple. Buy Nvidia GPUs. Rack them in data centers. Rent them out by the hour. Collect the spread. It is, in essence, a rental arbitrage — buying hardware that is supply-constrained and leasing it to companies that need compute faster than they can build it themselves. The early margins were enormous. The current margins tell a different story entirely.

CoreWeave's Q4 2025 adjusted EBITDA margin was 57%. Its net margin was negative 18%. Between those two numbers lies the entire economic reality of the neocloud sector — a business that looks wildly profitable before you account for the cost of the hardware, the interest on the debt used to buy it, and the price compression that is eroding the rental rates every quarter.

The CoreWeave Phenomenon: $5.1 Billion in Revenue, $18.8 Billion in Debt

CoreWeave IPO'd on March 28, 2025, pricing at $40 per share — below its expected $47-$55 range. Nvidia invested $250 million to bolster the offering. The stock closed flat on day one, dropped 10% the following Monday, then surged 42% on Tuesday. Since the IPO, shares have climbed roughly 200%, trading around $95.45 as of February 2026 with a market cap approaching $49.8 billion.

The revenue numbers are staggering. CoreWeave posted $1.92 billion in 2024 revenue — a 700%-plus year-over-year increase. In 2025, that grew to $5.1 billion, up 168%. The company became the fastest cloud provider in history to reach $5 billion in annual revenue. Management guided $12 billion to $13 billion for 2026, with an annualized run-rate of $17 billion to $19 billion exiting the year. The contracted backlog — pre-committed revenue from customers — reached $66.8 billion by end of Q4 2025, quadrupling during the year.

The customer list reads like a who's who of AI infrastructure demand. Microsoft accounted for 62% of 2024 revenue. OpenAI signed an initial $11.9 billion contract in March 2025, expanded it by $4 billion in May, and is pouring roughly $12 billion total into CoreWeave infrastructure over five years. Meta signed a $14 billion deal. By late 2025, CoreWeave had diversified enough that no single customer exceeded 35% of revenue.

But the debt side of the balance sheet is where the story gets complicated. CoreWeave carried approximately $18.8 billion in total debt as of September 2025. The stack includes a $2.3 billion GPU-backed credit facility from 2023, a $7.5 billion private credit facility from 2024, a $2.6 billion term loan, and $2 billion in convertible notes. Capital expenditure hit $8.2 billion in Q4 2025 alone — more than the company's total annual revenue. The debt-to-revenue ratio stands at 3.7x. Some bond market analyses price in a roughly 40% default risk.

GPU-backed debt is a new and untested asset class. Unlike real estate or manufacturing equipment, GPUs depreciate on technology cycles, not wear-and-tear schedules. A 2023-vintage H100 is worth meaningfully less in 2026 when Blackwell B200s are shipping. If demand slows or pricing compresses faster than expected, the collateral backing billions in loans could be worth a fraction of its original value.

The Margin Illusion: 57% EBITDA, 6% Operating, Negative 18% Net

The most important thing to understand about neocloud economics is the gap between headline margins and actual profitability.

CoreWeave's Q4 2025 adjusted EBITDA of $898 million at a 57% margin looks like a software business. But EBITDA strips out the two largest costs in the GPU rental business: depreciation of GPU hardware and interest on the debt used to finance it. Once you add those back, adjusted operating income was $88 million — a 6% margin, down from $121 million in Q4 2024 despite revenue more than doubling. After interest payments, the company lost $284 million in the quarter.

Industry-wide, the picture is even less flattering. Gross profit margins for GPU rental businesses run 14-16% after labor, power, and depreciation — lower than many non-tech retail operations. The McKinsey neocloud report describes bare-metal-as-a-service economics as "fragile." Neoclouds typically use large, low-margin offtake agreements with hyperscalers to finance fleet acquisition, then attempt to extend the economic life of the hardware by renting it at lower rates to enterprise customers.

CoreWeave management targets 25-30% operating margins long-term. Reaching that number requires scaling revenue faster than depreciation and interest accumulate, and it requires GPU utilization rates to stay high as competition intensifies. Neither is guaranteed.

The Price Collapse: H100 Rates Down 64% From Peak

The pricing environment has shifted dramatically against neocloud providers. H100 rental rates have collapsed from approximately $8 per GPU per hour at peak to $2.85-$3.50 — a 64% decline. AWS H100 spot instances dropped 88% between January 2024 and September 2025. Over 300 new providers entered the H100 cloud market in 2025.

The current pricing landscape is bifurcated. Budget neocloud providers like Vast.ai charge roughly $1.87 per H100 hour, and RunPod's community cloud runs about $1.99. Lambda Labs offers on-demand at $2.99. CoreWeave's H100 PCIe tier sits at $4.76. Hyperscaler pricing ranges from Google Cloud at $3.00 (after recent cuts) and AWS at $3.90 (after a 44% price reduction in June 2025) up to Microsoft Azure at $6.98 and Oracle at $10.00.

The forces driving further compression are structural, not cyclical. A100 and H100 units from expiring reservations are entering the secondary market. Nvidia's Blackwell B200 GPUs are launching broadly in 2026, which will push older-generation pricing down further. Analysts expect an additional 10-20% decline in GPU cloud rates through the year.

For companies that financed GPU fleets with debt based on 2023-2024 pricing assumptions, this math is unforgiving. The revenue per GPU-hour is declining while the debt service remains fixed.

The Supporting Cast: Lambda, Crusoe, and Together AI

CoreWeave is the largest and most visible neocloud, but it is not the only one navigating these economics.

Lambda Labs raised $1.5 billion in a November 2025 Series E led by TWG Global, bringing total funding to roughly $2.3 billion at a valuation north of $4 billion. Revenue hit an estimated $425 million in 2024, with an annualized run rate of $500 million by mid-2025. Lambda's most notable deal is a $1.5 billion agreement with Nvidia to lease back 18,000 GPUs over four years — making Nvidia simultaneously Lambda's largest supplier and its largest customer. Lambda also signed a multi-billion-dollar deal with Microsoft to deploy tens of thousands of Nvidia GPUs, including next-generation GB300 NVL72 systems.

Crusoe Energy brings a differentiated angle — energy. The company is vertically integrated, building its own power generation (natural gas turbines) alongside its data centers. A $1.375 billion Series E in October 2025 valued Crusoe at over $10 billion, a 3.6x jump from its $2.8 billion valuation just seven months earlier. Revenue grew from roughly $276 million in 2024 to a projected $500 million to $1 billion in 2025, with $2 billion projected for 2026. Crusoe's highest-profile project is the 1.2 GW campus in Abilene, Texas, the flagship site for OpenAI's Stargate initiative. The eighth and final building topped off in late 2025, with completion expected mid-2026.

Together AI occupies a slightly different niche, focused on open-source model inference and training. The company raised $305 million in a February 2025 Series B at a $3.3 billion valuation. Revenue reached an estimated $130 million in 2024 and approximately $300 million annualized by September 2025. Together AI runs two revenue lines: per-token API usage (30-40% of revenue) and GPU server rentals (60-70%). The API business offers a potential path beyond pure hardware arbitrage — but for now, GPU rentals remain the majority of the business.

Nvidia's Shadow: Supplier, Investor, Backstop, and Now Traffic Controller

No discussion of neocloud economics is complete without understanding Nvidia's extraordinary role in the ecosystem. Nvidia controls 92% of the discrete GPU market and an estimated 97%+ of the data center GPU accelerator market. Its data center revenue reached approximately $170 billion for fiscal year 2026. It has sold over $180 billion worth of Blackwell processors since launch.

But Nvidia isn't just the supplier. It is simultaneously an investor, a customer, and a financial backstop for the companies it sells to. Nvidia invested $250 million in CoreWeave's IPO, then poured in another $2 billion in January 2026 at $87.20 per share. It committed to purchasing up to $6.3 billion in unsold CoreWeave cloud capacity through April 2032. It is an investor in both Lambda Labs and Together AI. It leases back GPUs from Lambda under a $1.5 billion agreement. Critics have described this arrangement as "round-trip finance" — Nvidia funds companies that then buy Nvidia hardware, inflating Nvidia's own top line.

A pivotal shift occurred in September 2025 when Nvidia announced it would stop competing directly with AWS and Azure through its DGX Cloud offering. The team was reorganized and folded into core engineering. The reason was straightforward: competing with your own largest customers creates channel conflict. In its place, Nvidia launched DGX Cloud Lepton, a marketplace that routes workloads to partner providers including CoreWeave, Crusoe, Lambda, and others.

This was a major de-risking event for neoclouds. Nvidia chose to be the platform, not a competitor. But it also means Nvidia now sits at the center of the entire GPU cloud ecosystem — as supplier, financier, investor, customer, and traffic director. If Nvidia's interests ever diverge from those of the neoclouds it supports, the consequences would be immediate and severe.

The Risks That Could Unravel Everything

The neocloud sector faces a convergence of structural risks that could reshape the landscape within the next 12 to 18 months.

Commoditization is the existential threat. McKinsey's analysis is blunt: bare-metal-as-a-service economics are fragile. Renting GPUs by the hour is a commodity business. Neoclouds must move up the stack into AI-native services — model serving, inference optimization, workflow orchestration — or risk being squeezed between hyperscalers with deeper pockets above and budget providers with lower prices below.

Customer concentration remains dangerous. CoreWeave's 62% dependence on Microsoft in 2024 was an acknowledged risk. The company has diversified, but many smaller neoclouds remain dependent on one or two hyperscaler or AI lab contracts. The loss of a single deal can be existential.

The debt wall is approaching. CoreWeave's $18.8 billion in debt against $5.1 billion in revenue creates a 3.7x leverage ratio on a business with a negative net margin. GPU-backed lending is untested at this scale. First-generation 2021-2022 GPU deployments are hitting depreciation limits in 2026. If utilization drops or pricing compresses faster than expected, the collateral backing billions in loans loses value rapidly.

Hyperscalers are building their own GPU capacity. AWS, Azure, and Google Cloud are all constructing massive GPU clusters internally. Google's custom TPUs reduce dependence on Nvidia entirely. Amazon's Trainium chips could erode whatever cost advantage neoclouds currently offer. The hyperscalers were caught flat-footed in 2023-2024 when GPU demand spiked. They will not be flat-footed again.

A consolidation wave is expected in 2026. With over 130 GPUaaS providers, the market is fragmented far beyond what demand can sustain. McKinsey projects $3.1 trillion will flow into chips and computing hardware by 2030 — but only well-capitalized players will be around to capture it. Weaker, undifferentiated providers will fade or be acquired. The question is how many of the 130 survive.

What the GPU Rental Business Actually Is

Strip away the hype and the GPU rental arbitrage is, at its core, an infrastructure financing business. Neoclouds are financial intermediaries. They borrow money, buy depreciating hardware, and rent it out on contracts that they hope will generate enough cash flow to service the debt, replace the hardware, and eventually produce a profit.

This business model works when three conditions hold simultaneously: GPU supply is constrained, demand is accelerating, and pricing is stable or rising. In 2023 and early 2024, all three conditions were true. In 2026, supply constraints are easing, demand growth is uncertain beyond the hyperscaler and AI lab cohort, and pricing is falling.

CoreWeave's $50 billion valuation is a bet that it can thread the needle — that its $66.8 billion backlog converts to actual revenue, that it can move up the stack from bare-metal rentals into higher-margin services, that GPU demand continues to grow faster than supply, and that its debt service remains manageable as rates and hardware cycles evolve. The company has the scale, the contracts, and the Nvidia relationship to make this work. But the margin of error is razor-thin.

The broader neocloud sector's fate depends on an even simpler question: is renting out someone else's chips a sustainable business, or is it a transitional arbitrage that exists only because the hyperscalers were temporarily short on GPUs? The next 18 months will provide the answer. McKinsey, the bond market, and 130 GPU cloud startups are all watching the same numbers. The margins, it turns out, are not what anyone thought they were.