$35 Billion for GPUs. CoreWeave Bet the House.

From a pool table in New Jersey to the most leveraged company in AI

Three commodities traders, a pile of graphics cards on a pool table, and a grandfather’s garage in New Jersey. That was CoreWeave in 2017 — then called Atlantic Crypto — when Michael Intrator, Brian Venturo, and Brannin McBee were mining Ethereum with the kind of scrappy energy that venture capitalists love to romanticize in retrospect. One GPU became hundreds, hundreds became tens of thousands snapped up at fire-sale prices during the crypto winter of 2018, and by 2019 the trio had pivoted to renting compute to anyone who needed it. The cloud business grew 271 percent in its first three months. When Ethereum switched to proof-of-stake in 2022 and killed GPU mining overnight, CoreWeave did not flinch. It had already moved on. Then ChatGPT launched two months later, and every AI lab on Earth suddenly needed more GPUs than Amazon, Microsoft, and Google could deliver. CoreWeave was sitting on the supply.

On April 9, 2026, that garage-to-hyperscaler arc reached its most audacious chapter yet. CoreWeave and Meta announced a $21 billion expanded AI infrastructure agreement running through December 2032, layered on top of a $14.2 billion deal inked in September 2025. Total committed spend from a single customer: $35 billion. The contract makes Meta the largest commercial relationship in CoreWeave’s history and includes some of the first commercial deployments of NVIDIA’s Vera Rubin platform, the next-generation architecture unveiled at GTC 2026 in March. CEO Michael Intrator called it proof that “leading companies are choosing CoreWeave’s AI cloud to run their most demanding workloads.” Wall Street rewarded the announcement with a stock bump that pushed CRWV back above $90, more than double its $40 IPO price from March 2025.

The numbers are extraordinary by any standard. CoreWeave posted $5.1 billion in full-year 2025 revenue, up 168 percent year-over-year. Management guided for $12 billion to $13 billion in 2026 revenue, roughly 140 percent growth at the midpoint. The contracted revenue backlog stands at $66.8 billion as of year-end 2025 — a number that has since ballooned past $80 billion with the Meta expansion. NVIDIA invested $2 billion in CoreWeave stock at $87.20 per share and committed to a partnership that will see CoreWeave deploy more than five gigawatts of AI factories by 2030. The company that started with one GPU on a felt-covered table now operates approximately 470 megawatts of connected power and has contracted 2.2 gigawatts more. It is building the physical substrate of artificial intelligence, one rack at a time.

But there is a reason CoreWeave’s CDS spreads imply a 40 percent default probability over five years. The company carries more than $21 billion in long-term debt. Q4 2025 interest expense alone was $388 million — more than double the prior year — and the full-year net loss widened to $1.167 billion. Capital expenditure guidance for 2026 is $30 billion to $35 billion, roughly three times the revenue it expects to generate. CoreWeave is not a cloud company that happens to carry debt. It is a leveraged finance operation that happens to rent GPUs. The question is no longer whether the business can grow. It is whether the debt can be serviced if anything — demand, GPU pricing, a single mega-customer — goes sideways.

Meta’s $135 billion capex budget for 2026 is the tailwind. But tailwinds reverse. And the distance between $35 billion in contracted revenue and $35 billion in collected cash is measured in execution risk, technology transitions, and the brutal physics of GPU depreciation. CoreWeave is the most important company in AI that most people have never heard of. Whether it becomes the next great infrastructure franchise or the most spectacular leveraged blowup in tech history depends entirely on what happens in the next thirty-six months.

Follow the money through the GPU supply chain

To understand why Meta is writing $35 billion in checks to a company that was mining cryptocurrency five years ago, you need to understand the structural bottleneck that created the neocloud market. The three hyperscalers — AWS, Azure, and Google Cloud — collectively control roughly 65 percent of global cloud revenue. But when the generative AI wave hit in late 2022, their GPU procurement pipelines were optimized for general-purpose compute, not the massive clusters of NVIDIA A100s and H100s that training frontier models demands. Wait times for GPU capacity at the hyperscalers stretched to six months, then twelve. Labs with billions in funding and deadlines measured in quarters could not afford to wait. They turned to a new breed of GPU-focused cloud providers — the neoclouds — that had done one thing the hyperscalers had not: stockpiled NVIDIA hardware.

CoreWeave was the largest and fastest-moving of these. Its early relationship with NVIDIA, forged during the crypto years, gave it preferred-partner access to new GPU generations. When Blackwell shipped in late 2025, CoreWeave had racks online before most hyperscaler regions. When Vera Rubin was announced at GTC 2026, CoreWeave was named among the first commercial deployers. That access is the moat. NVIDIA does not sell GPUs to everyone equally — it allocates supply to partners who can deploy at scale, pay upfront, and demonstrate technical credibility. CoreWeave earned that position, and it translates directly into commercial advantage: customers like Meta, OpenAI, and Microsoft sign multi-year contracts not because CoreWeave’s software stack is superior to AWS, but because CoreWeave can deliver the specific hardware they need months before anyone else can.

The financial engineering behind the growth is as sophisticated as the hardware. CoreWeave closed an $8.5 billion financing facility in March 2026, the first investment-grade-rated GPU-backed financing in history. The company has raised approximately $28 billion in combined debt and equity over the past twelve months. For its OpenAI contract — originally $11.9 billion, later extended past $22 billion — CoreWeave created a special-purpose vehicle that owns the GPUs and data center assets, keeping billions in liabilities off its main balance sheet. The structure mirrors what you see in infrastructure project finance: long-dated revenue contracts collateralize the debt, and the assets (GPUs) serve as the physical backstop. It is elegant in a rising market. The problem is that GPU rental rates have fallen 64 to 75 percent from their Q4 2024 peak, with H100 spot prices declining from $8-to-$10 per hour to $2.50-to-$3.50. If the collateral value of your GPU fleet is dropping faster than your debt service obligations, the elegant structure starts to resemble a margin call waiting to happen.

The market backdrop reinforces both the opportunity and the risk. Mordor Intelligence estimated the neocloud market at $35 billion in 2026, up from $24 billion a year prior. Gartner projects neoclouds will capture 20 percent of the $267 billion AI cloud market by 2030. Q1 2026 saw investors pour $300 billion into startups globally, with AI companies claiming 80 percent of the total — up from 55 percent a year earlier. CoreWeave is not operating in a vacuum. It is riding the largest capital deployment wave in the history of technology. The question is whether the wave carries it to shore or whether the undertow of its own leverage pulls it under. For context, the $852 billion OpenAI valuation round and Anthropic’s own $30 billion raise both depend on the kind of compute infrastructure CoreWeave provides — making the neocloud market simultaneously the beneficiary and the systemic risk of the AI capital cycle.

Meta’s decision to spread its infrastructure bets across multiple vendors is instructive. Three weeks before the CoreWeave expansion, Meta signed a $27 billion deal with Nebius — the AI cloud provider that emerged from the restructuring of Yandex’s international operations — for dedicated capacity through 2031. Combined with its owned data center buildouts, including the 2,250-acre Hyperion campus in Louisiana and the Prometheus campus in Ohio, Meta is assembling a hybrid infrastructure empire that mixes owned and leased capacity across at least three major providers. CoreWeave CEO Intrator framed the diversification positively — noting that no single customer will represent more than 35 percent of total sales — but the flip side is also true: Meta is deliberately avoiding single-vendor lock-in. If CoreWeave stumbles on delivery timelines or Vera Rubin deployment, Meta has alternatives already contracted and funded. The $35 billion is a commitment, not a guarantee.

The three fault lines beneath the $35 billion floor

The most immediate risk is the GPU depreciation spiral. CoreWeave’s debt is collateralized by physical hardware that loses value every quarter. H100 rental rates have stabilized in the $2.50-to-$3.50 range after crashing from their 2024 peak, but the arrival of Blackwell — and soon Vera Rubin — means that each generation of hardware makes the previous generation less valuable. CoreWeave’s $8.5 billion GPU-backed financing facility was structured when H100s commanded premium prices. If Vera Rubin delivers the performance-per-dollar improvements NVIDIA has promised, H100 collateral values could compress further. The company must continuously cycle into new hardware, financing each generation at scale, while the residual value of existing inventory erodes. This is the treadmill problem: you cannot stop running without falling off, but running faster requires borrowing more.

The second fault line is customer concentration risk that the Meta deal ironically both solves and deepens. In 2024, Microsoft represented 62 percent of CoreWeave’s revenue. By mid-2025, that number was 72 percent. The Meta expansion diversifies the customer base on paper — Intrator’s 35-percent ceiling is a meaningful improvement — but it replaces one form of concentration with another. CoreWeave’s revenue base is now dominated by three customers: Microsoft, Meta, and OpenAI. Collectively, those three likely account for more than 80 percent of contracted backlog. If any one of them renegotiates terms, delays deployment, or builds out internal capacity faster than expected, the revenue shortfall would cascade through CoreWeave’s debt covenants. Meta alone has $115-to-$135 billion in 2026 capex earmarked for AI — much of it flowing through the closed-source Muse Spark strategy that demands proprietary infrastructure — but that spending is discretionary and subject to board approval each quarter. Zuckerberg has staked his reputation on AI infrastructure, but reputations are not contractual obligations.

The third fault line is competitive compression. CoreWeave’s moat — early access to NVIDIA hardware — is narrowing as NVIDIA’s own production scales. Nebius, the former Yandex spinout, raised $4.2 billion in September 2025 and now holds an estimated $6 billion in cash with contracts from both Microsoft (up to $19.4 billion) and Meta ($27 billion). Lambda continues to compete aggressively on price and simplicity. The hyperscalers themselves are catching up: AWS, Azure, and Google Cloud have dramatically expanded their GPU capacity over the past twelve months, and all three are developing custom AI chips (Trainium, Maia, TPU) that reduce their dependence on NVIDIA entirely. CoreWeave’s adjusted EBITDA margin of 57 percent in Q4 2025 is impressive, but margins in infrastructure businesses tend to compress as competition intensifies and hardware commoditizes. The neocloud market is growing at 46 percent annually, but the number of well-funded competitors is growing faster.

There is a scenario where all three fault lines converge. GPU prices continue to decline as Vera Rubin supply ramps. A major customer delays or restructures a contract. Competitors undercut CoreWeave’s pricing to win the next wave of deals. In that world, CoreWeave’s $21 billion in long-term debt — carrying roughly $1.2 billion in annual interest expense — becomes unsustainable. The CDS market is already pricing this possibility: a 40 percent implied default probability over five years is not a fringe estimate. It is the bond market’s best guess at what happens when a company with $5 billion in revenue borrows $21 billion against depreciating assets in a market where pricing power is eroding. CoreWeave’s bulls argue that the revenue backlog is the answer — $66.8 billion in contracted revenue provides visibility through the end of the decade. But contracted revenue and collected revenue are different things, and the history of infrastructure finance is littered with companies that had beautiful backlog numbers right up until they did not.

The thirty-six months that decide everything

CoreWeave’s path to either vindication or catastrophe runs through three inflection points between now and the end of 2028. The first is the Vera Rubin transition. If CoreWeave can deploy the new platform on schedule — second half of 2026 for initial clusters, full-scale availability in 2027 — it maintains its hardware-access moat and justifies the premium pricing embedded in its Meta and OpenAI contracts. A delay of even two quarters would give competitors time to close the gap and could trigger contract renegotiations. The second inflection is profitability. Management has signaled GAAP profitability by 2027, which would require interest expense to stabilize while revenue roughly triples. That is possible at the guided growth rates but leaves zero margin for execution error. The third inflection is refinancing. CoreWeave has $986 million in debt due in 2025 and $4.2 billion in 2026 that must be refinanced or repaid. If credit markets tighten — or if CoreWeave’s operating metrics disappoint — refinancing at favorable rates becomes significantly harder.

The bull case writes itself: AI infrastructure demand is insatiable, CoreWeave has the best customer list in the industry, NVIDIA’s $2 billion investment validates the platform, and the $66.8 billion backlog provides multi-year visibility that no other neocloud can match. At 32 analysts covering the stock, the consensus price target of $120.79 implies 28 percent upside from current levels. CoreWeave is building the physical layer of artificial intelligence — the racks, the power, the cooling, the fiber — and that layer does not go away even if individual model architectures change. The analogy to early cloud infrastructure is compelling: AWS was unprofitable for years before becoming Amazon’s highest-margin business. CoreWeave could follow the same trajectory if it survives the leverage.

The bear case is equally straightforward. CoreWeave is spending $30 billion to $35 billion in capex this year against $12 billion to $13 billion in projected revenue. It is financing that gap with debt collateralized by GPUs that depreciate faster than servers ever did. Its three largest customers are simultaneously building their own infrastructure and developing custom silicon that could reduce their dependence on NVIDIA hardware entirely. The neocloud market is growing, but so is competition from well-capitalized rivals like Nebius, which has $6 billion in cash and contracts worth more than $46 billion. And the entire thesis depends on AI demand continuing to grow at rates that have no historical precedent. If the AI investment cycle slows — not crashes, just slows — the mismatch between CoreWeave’s fixed debt obligations and its variable revenue stream could unravel quickly.

Here is the proprietary math that neither the bulls nor the bears are doing. CoreWeave’s $66.8 billion backlog converts to revenue over approximately five years. At the guided 57 percent EBITDA margin, that backlog generates roughly $38 billion in cumulative EBITDA. Against approximately $21 billion in long-term debt carrying $1.2 billion in annual interest ($6 billion over five years), the coverage ratio looks manageable — if every contract pays on time, at the contracted rate, with no renegotiation. But strip out the interest payments, subtract the $30-to-$35 billion per year in capex required to maintain and refresh the GPU fleet, and the free cash flow picture turns negative through at least 2028. CoreWeave is not building toward profitability. It is building toward a future refinancing event where it must convince credit markets that the next generation of contracts is as reliable as the current one. That is a bet on the permanence of AI demand — which may well be correct — financed by the most aggressive leveraged structure in the history of technology infrastructure.

For operators and investors watching from the outside, the checklist is short:

• Watch the Vera Rubin deployment timeline. Any slip past Q4 2026 for initial commercial availability is a yellow flag for contract execution risk.
• Track Meta’s quarterly capex disclosures. The $115-to-$135 billion 2026 guide is the floor assumption in CoreWeave’s model. A reduction changes the math.
• Monitor H100 and Blackwell spot rental rates. If H100s fall below $2.00 per hour and Blackwell rates follow the same compression curve, CoreWeave’s collateral values are in trouble.
• Follow the refinancing. The $4.2 billion in 2026 maturities is the near-term stress test. The terms of that refinancing will reveal how much confidence credit markets have in the backlog.
• Compare CoreWeave’s customer concentration reports each quarter. The 35-percent ceiling Intrator promised is the metric that tells you whether diversification is real or aspirational.

CoreWeave turned a crypto winter into a cloud empire and a garage into a $90-per-share public company. The $35 billion Meta deal is the capstone of a transformation story that has no parallel in modern tech. But capstones are heavy, and the foundation beneath this one is made of borrowed money and depreciating silicon. The next thirty-six months will determine whether CoreWeave is the AWS of the AI era — or the most spectacular confirmation that leverage and growth are not the same thing.

In other news

Visa enables AI agents to make card payments autonomously — Nevermined integrated Visa Intelligent Commerce and Coinbase’s x402 protocol to let AI agents purchase digital goods and services within cardholder-defined guardrails, including budget caps, merchant restrictions, and time-based spending windows. It is the first practical mechanism for machines to transact at the point of sale without human checkout flows (Dataconomy).

Meta AI app rockets to No. 5 on the App Store — Following the Muse Spark model launch on April 8, Meta’s standalone AI app surged from No. 57 to No. 5 on the U.S. App Store in less than 24 hours. Daily web visitors in the U.S. rose more than 450 percent day-over-day, hitting an all-time high.

AI scribes are raising healthcare costs — and nobody agrees on a fix — A roundtable of insurers, providers, and investors unanimously acknowledged that AI ambient scribes are driving higher medical coding intensity and inflating claims. Health economists warned of an “AI coding arms race” between provider upcoding tools and insurer denial algorithms that could hurt vulnerable patients most (STAT News).

Shield AI raises $2 billion at $12.7 billion valuation — The defense AI startup closed a $1.5 billion Series G co-led by Advent International and JPMorgan, plus $500 million in preferred equity from Blackstone, after the U.S. Air Force selected its Hivemind autonomous pilot platform for the Collaborative Combat Aircraft program. Valuation jumped 140 percent in one year (TechCrunch).

Q1 2026 venture funding shatters all records at $300 billion — Crunchbase data shows investors poured $300 billion into 6,000 startups globally in Q1, up more than 150 percent year-over-year. AI companies claimed 80 percent of the total, with OpenAI’s $122 billion round, Anthropic’s $30 billion raise, and xAI’s $20 billion close collectively accounting for 65 percent of all global venture capital (Crunchbase).