CoreWeave announced $28 billion in deal in 6 days. Are they good for it?
The celebration was immediate. The questions came later.
In the first two weeks of April 2026, CoreWeave, Inc. did something remarkable. The company announced a $21 billion expanded agreement with Meta Platforms on April 9. The next day, it revealed a multi-year deal with Anthropic reportedly worth $6.8 billion. Five days later, Jane Street committed $6 billion in cloud services and invested $1 billion in equity at $109 per share. In six days, CoreWeave added roughly $28 billion to its revenue backlog, pushing the total past $88 billion.
The stock surged from $69 on March 30 to $117 by April 15. A 50% rally in sixteen days. Analysts raised price targets. Headlines declared the “neocloud” model vindicated. CoreWeave, the company that had crashed from post-IPO highs to $25 barely a year earlier, was being discussed alongside the hyperscalers.
On the same day it announced the Meta deal, CoreWeave also priced $4.25 billion in new debt: $1.25 billion in senior notes and $3.0 billion in convertible notes. That offering was upsized to $5.25 billion within forty-eight hours. The Meta deal was packaged as Exhibit 99.3, “Supplemental Information,” provided to potential bond investors. The deal announcement was marketed to lenders, not filed as a binding legal event.
During those same sixteen days, company insiders sold approximately $1.03 billion in stock. Magnetar Capital, the hedge fund that had turned a $50 million convertible note into a position once worth $12.5 billion, dumped 4.9 million shares for $577 million. All three co-founders were selling on pre-arranged plans they had adopted within a single week in November 2025.
Nobody bought.
This is the story of what the public filings show, what they do not show, and the questions that nobody seems to be asking about the most aggressively financed company in the AI infrastructure boom.
---
A Brief History of CoreWeave
CoreWeave was not born as an AI company. It was born as a cryptocurrency mine.
In 2017, three friends from the New Jersey commodities trading world launched a company called Atlantic Crypto. Michael Intrator, Brian Venturo, and Brannin McBee were energy and options traders, not engineers. They understood one thing clearly: the relationship between electricity prices, hardware costs, and the value of what the hardware produces. They bought NVIDIA GPUs by the thousands, packed them into rented facilities, and mined Ethereum.
For a year, this worked. Then the 2018 crypto crash erased 80% of Ethereum’s value and the mining economics collapsed overnight. Atlantic Crypto was sitting on warehouses full of NVIDIA GPUs with nothing profitable to mine.
This is where the origin story diverges from a dozen other failed mining operations. Rather than liquidate the hardware, the founders asked a different question: who else needs GPUs and cannot get them? The answer, in 2018 and 2019, was machine learning researchers. NVIDIA’s chips were designed for parallel computation. Training neural networks is parallel computation. The major cloud providers (AWS, Azure, Google Cloud) offered GPU instances, but availability was limited, pricing was opaque, and researchers often waited weeks for allocation.
Atlantic Crypto began renting its mining GPUs to academic labs and small AI startups. The company brought on Peter Salanki as Chief Technology Officer to build a proper cloud orchestration layer. By 2019 the pivot was complete. The Ethereum miners were now serving inference and training workloads. In October 2021, the company formally rebranded as CoreWeave.
The timing of what came next was extraordinary. OpenAI released ChatGPT on November 30, 2022. Within months, every technology company on earth wanted GPU capacity. Microsoft, unable to build fast enough internally, signed a multi-billion dollar agreement with CoreWeave in 2023, becoming the company’s anchor customer and eventually its largest by a wide margin (62% of revenue, rising to 67% in fiscal year 2025).
CoreWeave went public on March 28, 2025 at $40 per share, below its initial $51 target, raising $1.5 billion. At that point the company was eighteen months removed from being a rebranded crypto mining operation with a few hundred million in annual revenue.
The founders’ original insight, the tradeable relationship between hardware cost, energy price, and output value, never changed. What changed was the output. First it was Ethereum. Now it is tokens. The commodity is different. The bet is structurally identical: acquire specialized hardware using leverage, sell its output for more than the carrying cost, and refinance into the next generation before the current one depreciates below its debt obligation.
Whether this is a technology company or a leveraged commodity trade dressed in a software multiple is the question that determines whether CoreWeave is worth $117 per share or something much closer to zero.
---
How the Machine Works
To understand CoreWeave, you need to understand the machine it runs on. The machine is not a data center. It is a financing loop.
CoreWeave is a “neocloud,” a term for companies that exist to buy NVIDIA GPUs and rent them to others. The company went public on March 28, 2025 at $40 per share, below its $51 target. It generated $5.13 billion in revenue in fiscal year 2025, an increase of 168% over the prior year. It lost $1.17 billion. It spent $14.9 billion on capital expenditures, roughly three dollars for every dollar of revenue. Its free cash flow was approximately negative $13 billion.
Those numbers are not the result of mismanagement. They are the business model working as designed. CoreWeave operates a capital-intensive flywheel that works like this:
NVIDIA invests equity in CoreWeave. That equity strengthens CoreWeave’s balance sheet, enabling it to raise debt collateralized by NVIDIA GPUs. CoreWeave uses the debt proceeds to buy more NVIDIA GPUs. NVIDIA books those purchases as data center revenue. The revenue growth supports NVIDIA’s stock price, generating capital for more equity investments. The cycle repeats.
The multiplier is striking. For every $1 of equity NVIDIA invested in CoreWeave, approximately $15 in total capital expenditure was unlocked. NVIDIA’s $2.25 billion in equity (a $250 million IPO anchor investment in March 2025 and a $2 billion private placement in January 2026 at $87.20 per share) helped catalyze an $8.5 billion debt facility in March 2026, which fed into $30 to $35 billion in planned 2026 capital expenditure, overwhelmingly directed back to NVIDIA for GPU purchases.
NVIDIA also signed a $6.3 billion capacity backstop agreement through April 2032, committing to buy any CoreWeave compute capacity that goes unsold. This is a revenue guarantee. If CoreWeave cannot find customers for its GPUs, NVIDIA will pay for them anyway. The backstop de-risks CoreWeave’s debt structure while ensuring NVIDIA’s largest neocloud customer cannot fail without NVIDIA explicitly allowing it.
As of January 2026, NVIDIA holds 47.2 million shares of CoreWeave Class A common stock, representing 11.5% of the class and making it the second-largest shareholder after CEO Michael Intrator. That position is worth approximately $5.5 billion at the April 19 closing price of $116.99. NVIDIA has filed no Form 4 sales through April 20. The shares are likely subject to lock-up restrictions.
This structure is disclosed across multiple SEC filings. But it is spread across enough filings, entities, and footnotes that the full picture rarely assembles in one place. The chips are simultaneously the product sold, the collateral pledged, and the reason the customer exists.
NVIDIA’s broader strategic investment program now exceeds $110 billion in commitments across the AI ecosystem: up to $100 billion in OpenAI (staged in ten tranches of $10 billion), up to $10 billion in Anthropic, $2 billion in Nebius, $2 billion in xAI, and smaller stakes across Crusoe Energy, Lambda Labs, Nscale, and formerly Applied Digital. On top of those, $13 billion flows to supply chain partners including Intel, Synopsys, Marvell, Lumentum, and Coherent. The total: $110 billion in commitments against approximately $165 billion in trailing twelve-month revenue, a ratio of 67%. No analyst on NVIDIA’s earnings call has asked what percentage of reported revenue originates from entities NVIDIA has funded. The question sits in the gap between what is disclosed and what is asked.
CoreWeave is not profitable on a GAAP basis. Fiscal year 2025 operating income was negative $46 million, down from positive $324 million the prior year, as interest expense growth outpaced revenue growth. The company lost $1.17 billion on the bottom line. Adjusted EBITDA, the metric CoreWeave prefers, was approximately $3 billion, with margins around 57%. But EBITDA is not cash flow. It excludes the interest payments that consume a quarter of revenue, the principal repayments running $250 to $300 million monthly, and the stock-based compensation that dilutes shareholders while appearing nowhere in cash expense.
---
The Deals
The April 2026 deals pushed CoreWeave’s stated backlog toward $88 billion. That number headlined every news cycle. The details beneath it deserve closer scrutiny.
Meta: $21 billion. The expanded agreement was announced April 9, 2026, structured as a new order form under an existing Master Services Agreement dated December 10, 2023. The deal extends through December 2032 and references initial deployments of NVIDIA’s Vera Rubin platform, which has not yet shipped in volume.
The critical finding is how this deal was filed. CoreWeave’s 8-K reported the Meta expansion under Item 7.01 (Regulation FD Disclosure) and Item 8.01 (Other Events). It was not filed under Item 1.01 (Entry into a Material Definitive Agreement). The deal appeared as supplemental information packaged alongside a $4.25 billion debt offering. The actual order form was not filed as an exhibit.
Compare this to the prior Meta deal. In September 2025, CoreWeave filed the original $14.2 billion Meta order form under Item 1.01, with the redacted MSA included as an exhibit. That filing disclosed specific termination provisions: either party may terminate for cause. No convenience termination was disclosed.
The April 2026 expansion received notably lighter regulatory treatment. We do not know why. We do not know the termination provisions of the new order form because it was not filed. This matters because the $21 billion headline is the number that supports the $88 billion backlog, which in turn supports the $8.5 billion debt facility that used Meta’s $19.2 billion contract backlog as collateral.
Anthropic: reportedly $6.8 billion. This deal generated no SEC filing at all. No 8-K was filed under any item. CoreWeave’s press release described a “multi-year deal with Anthropic to power Claude AI models,” deliberately vague on financial terms. The $6.8 billion figure comes from a single media source, FinancialContent/MarketMinute. CoreWeave declined to disclose the deal’s value to CNBC. The phased roll-out “with the potential to expand over time” suggests this may be a smaller initial commitment with options, not a $6.8 billion binding obligation from day one.
Jane Street: $6 billion cloud agreement plus $1 billion equity. The 8-K filed April 15 covers only the equity piece: 9,174,311 shares of Class A common stock at $109.00 per share, a 7% discount to the prior closing price. The $6 billion cloud agreement, the vastly larger component, was not filed under Item 1.01. No contract was filed as an exhibit.
A quantitative trading firm committing $6 billion to AI cloud compute is an extraordinary claim. Jane Street is legitimately one of the world’s largest consumers of compute for quantitative modeling. But the deal structure, pairing a $6 billion cloud commitment with a $1 billion equity investment at a discount, creates a dual relationship where Jane Street is simultaneously customer and shareholder with interests aligned to CoreWeave’s stock price performance. The equity investment arrived squarely within the unlimited equity cure window for CoreWeave’s covenant compliance, potentially serving double duty as both strategic capital and covenant relief.
The filing pattern across all three deals looks like this: of $28 billion in announced deal value, zero dollars were filed as Material Definitive Agreements, zero contracts were filed as exhibits, and the only document actually submitted to the SEC was a press release.
CoreWeave describes its contracts as “take-or-pay.” But its own filings qualify all backlog as “subject to the satisfaction of delivery and availability of service requirements.” CoreWeave must physically build data centers, procure NVIDIA GPUs that have not yet shipped, and deliver working infrastructure before contracted revenue is recognized.
This qualifier is not hypothetical. In February 2025, Core Scientific began flagging construction delays at CoreWeave’s Denton, Texas data center. Heavy rains and winds caused 60-day concrete pour delays that summer. CoreWeave did not publicly disclose these issues until November 10, 2025, nine months after its construction partner first raised concerns. The disclosure came during Q3 earnings and included a revenue guidance cut and a capex reduction from $20-23 billion to $12-14 billion. CEO Michael Intrator initially characterized the issue as affecting a “singular data center” on CNBC before clarifying it involved a “singular data center provider” with multiple affected sites. The stock fell 34% between November 10 and December 16, 2025, destroying $14 billion in market capitalization.
Securities class action lawsuits were filed by more than ten law firms, including Pomerantz LLP, Robbins Geller Rudman & Dowd, and Hagens Berman, alleging CoreWeave concealed known delays and overstated its ability to meet customer demand. The class period spans from the March 2025 IPO through December 15, 2025. Named defendants include CEO Intrator, CFO Nitin Agrawal, and Chief Development Officer McBee.
CoreWeave separately disclosed material weaknesses in internal controls over financial reporting in its IPO S-1 filing, citing insufficient IT general controls, inadequate segregation of duties, and staffing gaps in accounting and finance. Remediation is expected to continue through at least 2026.
The connection between construction delays and the covenant amendment is direct. The Wall Street Journal revealed the full scope of the Denton delays on December 15, 2025. Sixteen days later, CoreWeave executed the covenant amendment that slashed the liquidity floor to $100 million and granted unlimited equity cures.
Now CoreWeave has committed to $30 to $35 billion in 2026 capital expenditure, roughly $2.60 for every dollar of new revenue guided. All three April deals reference Vera Rubin deployments. If NVIDIA’s next-generation platform ships on schedule in the second half of 2026, CoreWeave captures its contracted revenue ramp. If it slips, CoreWeave faces what the Denton experience already demonstrated: debt service continues while revenue recognition is deferred.
---
The Debt Clock
CoreWeave carries approximately $23 to $26 billion in drawn debt as of mid-April 2026, with total capacity of $32 to $34 billion across its facilities. This is against $5.13 billion in fiscal year 2025 revenue. The debt stack is layered and complex.
The secured facilities tell the story of acceleration. The first delayed-draw term loan, $2.3 billion led by Blackstone and Magnetar in August 2023, carried a floating rate of approximately 15%. The second, $7.6 billion from multiple institutional lenders in May 2024, carried an average variable rate of roughly 11%. A third facility of $2.6 billion closed in July 2025. Then in March 2026, the $8.5 billion DDTL 4.0 closed at SOFR plus 2.25%, backed by all assets of a special purpose entity and collateralized by Meta’s $19.2 billion contract backlog. This was the first investment-grade rated GPU-backed financing in history. Moody’s rated it A3. There is zero precedent for how these structures perform under stress.
On top of the secured facilities sit $5.5 billion in unsecured senior notes carrying coupons between 9.0% and 9.75%, and approximately $6.1 billion in convertible senior notes at 1.75%. A $2.5 billion revolving credit facility rounds out the stack.
Interest expense consumed roughly 25 to 27% of fiscal year 2025 revenue, estimated at $1.3 to $1.4 billion. That figure had quadrupled year-over-year; Q4 2025 interest expense alone was approximately $388 million, implying an annualized run rate of $1.55 billion. For 2026, with the additional debt layered on, projected annual interest reaches $2.0 to $2.5 billion.
Interest expense tripled. Margins compressed from 60% adjusted EBITDA in Q4 2024 to 57% in Q4 2025. Nobody celebrated that math.
The 2026 cash demands are straightforward arithmetic. Capital expenditure guidance: $30 to $35 billion. Interest expense: $2.0 to $2.5 billion. DDTL principal repayments that began in January 2026: an estimated $3 to $5 billion, with monthly payments running approximately $250 to $300 million. Operating expenses: $6 to $8 billion. Total 2026 cash needs: approximately $42 to $50 billion. Revenue guidance: $12 to $13 billion.
The funding gap is $30 to $37 billion. That gap must be filled through new debt, equity, and asset-level financing. The April offerings, the DDTL 4.0, the Jane Street equity, and whatever comes next are not signs of momentum. They are the continuous transfusions that keep the patient alive. The machine must keep running.
The January Amendment. On December 31, 2025, sixteen days after the Wall Street Journal revealed the full scope of the Denton construction delays, CoreWeave executed an amendment to its DDTL 3.0 credit agreement. The amendment was filed as an 8-K on January 2, 2026.
Three things changed. First, the minimum liquidity requirement was slashed to $100 million for the March through April 2026 monthly payment dates.
One hundred million dollars is the liquidity floor for a company with more than $21 billion in debt.
Second, debt service coverage ratio testing was postponed to October 31, 2027. Third, and most importantly, the amendment granted unlimited equity cure rights for covenant failures prior to October 28, 2026.
In plain language: CoreWeave’s lenders acknowledged the company could not meet its covenant tests. Rather than declaring a breach, they gave CoreWeave a ten-month window where any covenant failure could be “cured” by injecting equity. If cash flow from operations proved insufficient to cover debt service (it was, and is), the company could plug the gap by issuing stock or receiving equity investments, and lenders would not trigger a default.
After October 28, 2026, the unlimited equity cure provision expires. Equity cures become limited to three consecutive months in any four-month rolling period. A persistent coverage shortfall in November and December 2026 would mean that by the fourth month, the covenant must be met organically or a technical default occurs on the DDTL 3.0. Whether that triggers cross-defaults to other facilities is not publicly known. It is one of the most important unanswered questions in this story.
---
The Collateral Question
Everything in CoreWeave’s debt structure ultimately rests on a single assumption: that NVIDIA GPUs hold their value long enough to support the loans they collateralize. The evidence suggests this assumption is under severe pressure.
CoreWeave depreciates its GPU fleet over six years on a straight-line basis, an assumption it adopted in January 2023 after extending from four years. The timing is notable. The extension occurred just as CoreWeave was scaling its GPU-backed debt facilities. The $2.3 billion Magnetar/Blackstone facility closed in August 2023. A longer depreciation period reduces annual expense, inflates EBITDA, and improves debt service coverage ratios, all metrics that matter when borrowing against the same assets being depreciated.
CoreWeave’s six-year assumption ties with Microsoft and Google as the most aggressive in the industry. Amazon tells a different story. After extending server life from three years to four years in 2020, then to five in 2022, then to six in 2024 (adding $900 million to Q1 2024 net income), Amazon reversed course on January 1, 2025, shortening servers back to five years. The reason cited: “accelerated AI technology development.” The reversal cost Amazon $700 million in reduced 2025 operating income.
Amazon, with arguably the deepest infrastructure expertise of any hyperscaler, concluded that six years was too aggressive and ate a $700 million hit to correct the assumption. CoreWeave, with the most levered balance sheet in the sector, maintained it. In March 2026, the company stated that “six years remains appropriate” and that “usage could extend beyond that.”
Nebius, a direct competitor, uses four-year depreciation for equivalent GPU assets.
The secondary market tells its own story. H100 rental rates have collapsed approximately 60 to 75% from peak, falling from $8 to $10 per GPU-hour in late 2023 to $2 to $3 per hour by April 2026, with spot rates occasionally below $2.50. AWS cut H100 pricing by 44 to 45% in June 2025, triggering a broader market reset. Used H100s trade at $12,000 to $22,000, roughly 40 to 53% of CoreWeave’s original purchase price of $25,000 to $30,000. Under six-year straight-line depreciation, the book value at year three is approximately $15,000. Market value is already at or below book value, with three years of depreciation remaining on the books.
Jim Chanos, the short seller who famously identified Enron, has been the most vocal critic of the depreciation math. On the “Monetary Matters” podcast in December 2025, Chanos laid out the arithmetic: “If the chips last for three years, you have to depreciate a third of what you spend.” Using CoreWeave’s own numbers, he argued, annual economic depreciation at a seven-year useful life (a figure CEO Intrator has used publicly) would be $2.729 billion, exceeding Q2 annualized EBITDA of $2.624 billion. At CoreWeave’s own aggressive assumptions, depreciation alone nearly consumes all EBITDA.
Michael Burry, who rose to fame shorting subprime mortgage-backed securities, has estimated $176 billion in cumulative earnings understatement across major AI infrastructure players between 2026 and 2028 from aggressive depreciation policies. His assessment of actual useful life: two to three years.
Kerrisdale Capital’s September 2025 short report, “Artificial Returns,” called CoreWeave “a debt-fueled GPU rental business with no moat, dressed up as innovation” and targeted $10 per share, a 90% decline.
Now consider what comes next. NVIDIA’s Vera Rubin platform, offering five times Blackwell’s inference performance at ten times lower cost per token, ships in the second half of 2026. First customer samples have already been delivered. All three of CoreWeave’s April deals explicitly reference Vera Rubin deployments.
CoreWeave is simultaneously deploying next-generation silicon that will make its existing H100 and H200 fleet economically obsolete for premium workloads. SemiAnalysis estimates that an H100 would need to rent at $0.98 per hour to match Vera Rubin’s price-per-output. Current spot rates are $1.25 to $2.50. Another 20 to 60% decline is implied.
Jensen Huang himself warned about this dynamic at GTC 2025:
“When Blackwell starts shipping in volume, you couldn’t give Hoppers away.”
The math on collateral erosion is blunt. If Vera Rubin drives H100 market values down to $6,000 to $11,000 per unit while book value remains at approximately $15,000, the gap of $4,000 to $9,000 per GPU across hundreds of thousands of units amounts to billions in potential undercollateralization. Dave Friedman’s analysis noted that lenders “assumed 50% value retention at 3 years,” and H100s have already exceeded that decline timeline by eighteen months. If lenders demand collateral top-ups, CoreWeave would need to inject additional capital, pledge additional depreciating GPU assets, or negotiate yet another covenant waiver.
CoreWeave’s projected debt service coverage ratio is 1.26x through 2031. The minimum covenant threshold is 1.15x. That leaves eleven basis points of cushion. If depreciation expense were to double under a three-year useful life assumption, the additional $2.52 billion in annual expense would collapse that cushion entirely and push EBITDA-based covenants into breach territory.
The split in credit ratings tells the story visually. CoreWeave’s secured DDTL 4.0 is rated A3 by Moody’s and A (low) by DBRS. Investment grade. The corporate family rating is Ba3. The unsecured senior notes are rated B1. The gap between secured and unsecured reflects that lenders to the secured facilities trust the GPU collateral and the contracted revenue streams. The corporate entity itself carries substantial risk. If collateral values erode to the point where the secured facilities are no longer truly over-collateralized, that gap collapses. The investment-grade tranche would become indistinguishable from the rest of the capital structure.
The Level-Headed Investing Substack captured the paradox:
“The 2.5-year EBITDA payback looks impressive until you realize the hardware needs replacing in 3-4 years. This works beautifully in a capex boom but becomes a refinancing nightmare when credit tightens or demand softens.”
---
The Insider Signal
Since CoreWeave’s IPO on March 28, 2025, insiders have sold more than $5 billion in stock. In the same period, the number of open-market purchases by officers or directors is zero.
Not one executive. Not at any price. Not when the stock was at $120, or $80, or $25. The only two “purchase” transactions on record are pre-IPO private placements: director Glenn Hutchins ($20 million on November 14, 2024 at $47.28 per share) and director Karen Boone ($500,000 on March 3, 2025 at $47.56). Both occurred before the stock was publicly traded.
Between December 2025 and April 17, 2026, SEC Form 4 filings document 100 insider sale transactions totaling approximately $1.248 billion across 11.3 million shares. In April 2026 alone, as the stock surged above $118 on the Meta and Anthropic deal announcements, insiders sold approximately $1.03 billion.
The largest seller was Magnetar Capital. The firm’s CoreWeave story is one of the most extraordinary return narratives in hedge fund history. Magnetar’s original investment was $50 million in convertible notes during CoreWeave’s Ethereum mining era. Those notes converted to equity. By September 30, 2025, Magnetar held 91.4 million shares, approximately 23% of the company, valued at roughly $12.5 billion. That single position represented 72% of Magnetar’s entire $20.5 billion portfolio.
Then the liquidation began. Magnetar sold approximately $372 million in late September 2025. Between April 15 and 17, 2026, the firm dumped another 4.9 million shares for $577 million at prices between $118.30 and $118.85. It also wrote exchange-traded call options on 2 million shares at a $160 strike expiring December 18, 2026, capping upside while generating premium. The remaining position: approximately 2.8 million shares. Magnetar has liquidated more than 97% of its CoreWeave holdings.
The co-founders present a different pattern but the same directional signal. All three adopted Rule 10b5-1 trading plans within a single week in November 2025: Brian Venturo (Chief Strategy Officer) on November 13, Brannin McBee (Chief Development Officer) on November 17, and Michael Intrator (CEO) on November 20. These plans governed the systematic selling that followed through April 2026.
Chief Strategy Officer Brian Venturo sold 1,125,000-share blocks on April 1, April 6, and April 13 for combined proceeds of approximately $327 million through West Clay Capital LLC and the Venturo Family GST Exempt Trust. CEO Michael Intrator sold 307,693 shares on April 14 at $116.94 for $36 million. Chief Development Officer McBee sold through multiple grantor retained annuity trusts named Canis Major 2025 and Canis Minor 2025, executing 46 sale transactions in approximately five months.
The 10b5-1 plans were adopted while CoreWeave was negotiating deals that would later be announced in early 2026. Rule 10b5-1 requires adoption when the insider does not possess material nonpublic information. The plans executed at dramatically higher prices than the $65 to $75 range where the stock traded in November 2025. Under SEC rules, the 90-day cooling period meant the earliest these plans could execute was approximately February 11 to 18, 2026.
NVIDIA has not sold. Its 47.2 million shares remain intact per SEC filings through April 20. But NVIDIA’s shares are likely subject to lock-up restrictions from the January 2026 private placement.
One additional context point frames all of this: Microsoft accounted for approximately 67% of CoreWeave’s fiscal year 2025 revenue, roughly $3.44 billion. That concentration actually increased from 62% in fiscal year 2024, meaning Microsoft’s spend grew faster than CoreWeave’s overall revenue despite diversification efforts. The April 2026 deals diversify the backlog on a forward-looking basis, but current cash flows remain dominated by a single customer that is simultaneously building out its own capacity, diversifying across multiple neoclouds (Nebius for $19.4 billion, Nscale for $23 billion, IREN for $9.7 billion), and that declined a $12 billion option to expand its CoreWeave commitment in March 2025. Microsoft froze 1.5 gigawatts of self-developed data center capacity in Q1 2026. Its data center leasing dropped from $11.1 billion in one quarter to $6.7 billion the next.
The insiders who adopted 10b5-1 plans in November 2025 would have known about Microsoft’s trajectory. The question of what they knew about the April 2026 deals cannot be answered from public filings. The question of what the selling pattern implies is left to the reader.
---
The Echo
Three companies from a prior era built enormous revenue streams by financing customers who could not otherwise afford their products. All three saw those revenue streams evaporate when the customers failed.
Lucent Technologies. Peak revenue of $38.3 billion in fiscal year 1999. Vendor financing outstanding: $8.1 billion, representing 24% of fiscal year 2000 revenue. The SEC later found Lucent “fraudulently and improperly recognized approximately $1.148 billion of revenue,” of which $637 million “should not have been recognized at all.” Bad debt provisions: $3.5 billion. Stock decline: peak of $82.31 to eventual merger with Alcatel at $3.01, a 96% decline. Revenue collapsed from $38.3 billion to $8 billion by fiscal year 2006.
Nortel Networks. Peak revenue of $30.3 billion in 2000. Vendor financing extended: $3.1 billion across 45 deals, mostly interest-free and unsecured. Between 2000 and 2003, 47 Competitive Local Exchange Carriers went bankrupt. Not one or two. Forty-seven. The CLECs had all been funded by the same vendors, buying the same equipment, building the same networks, competing for the same customers. When demand failed to materialize as projected, they failed together because their business models were structurally identical. Nortel’s collateral, telecom equipment installed in failed companies, proved worthless. Peak market capitalization: $366 billion. Stock at bankruptcy: $0.39. A 99.7% decline. The largest corporate bankruptcy in Canadian history. The neocloud sector today shares the structural feature that made CLECs vulnerable: identical business models, identical suppliers, overlapping creditors, and correlated failure conditions.
Cisco Systems. Revenue from financed customers: approximately 10% of $20 billion in annual revenue. Inventory write-off in 2001: $2.25 billion. Bad loan write-offs: approximately $900 million. Stock decline: from approximately $80 to below $10, an 88% decline. Cisco survived. It was more diversified, had a stronger balance sheet, and recovered over the subsequent decade.
Now consider the comparison table.
NVIDIA’s strategic investment commitments exceed $110 billion, equivalent to 67% of its trailing twelve-month revenue. Lucent’s vendor financing was $8.1 billion, or 24% of revenue. NVIDIA’s relative exposure is 2.8 times larger. NVIDIA’s top two customers represent 39% of revenue; Lucent’s top two (AT&T at 10%, Verizon at 13%) represented 23%. NVIDIA’s customer concentration is nearly double. The equipment Lucent financed had a useful life of 10 to 15 years. NVIDIA’s GPUs have an economic life of 2 to 3 years, according to short sellers and at least one hyperscaler’s revealed preference.
The key difference: NVIDIA is enormously profitable. It generated $97 billion in free cash flow in fiscal year 2026. Lucent could not absorb $3.5 billion in loan losses. NVIDIA could theoretically absorb a $30 billion write-down without an existential crisis. NVIDIA’s exposure is equity, not direct loans; a mark-to-market loss, not a bad debt provision. NVIDIA is not a creditor in any neocloud bankruptcy.
The key similarities: the economic function is identical. Capital flows to customers to buy the vendor’s product. The accounting opacity is worse than Lucent’s. NVIDIA categorizes its neocloud investments as “strategic.” There is no “vendor financing” line item. The aggregate $110 billion investment program is described by NVIDIA leadership as “ecosystem development.” On NVIDIA’s Q4 fiscal year 2026 earnings call, Morgan Stanley analyst Joe Moore asked about the strategic investment program. Jensen Huang responded: “Fundamentally, at the core of everything NVIDIA is its ecosystem, which is what everybody loves about our business. We want to make sure we continue to invest into our ecosystem.”
No analyst followed up on circular financing. No analyst asked about the Lucent parallel. No analyst asked what percentage of NVIDIA’s $193.7 billion in data center revenue originates from companies NVIDIA has funded. Fifty-seven of sixty covering analysts rate NVIDIA a Buy.
NVIDIA’s 10-Q reveals that four direct customers each exceeded 10% of total revenue in Q3 fiscal year 2026. The company uses anonymized designations. Top two customers: 39% of revenue. Top four: 46%. For context, Lucent’s top two customers represented 23% of revenue. NVIDIA’s concentration is nearly double.
Our estimate, based on reconstructing neocloud capital expenditures: 15 to 30% of NVIDIA’s data center revenue may originate from companies in which NVIDIA holds equity stakes. NVIDIA does not disclose this figure. CoreWeave alone spent $14.9 billion on capital expenditure in fiscal year 2025, the majority directed to NVIDIA hardware, with 2026 guidance of $30 to $35 billion. If CoreWeave is spending $15 to $30 billion annually, and Nebius, Lambda, Crusoe, Nscale, and Applied Digital collectively spend a comparable amount, neocloud GPU purchases could represent $30 to $60 billion of NVIDIA’s data center revenue.
The Cisco parallel is instructive as much for its differences as its similarities. Cisco survived the dot-com collapse despite $2.4 billion in customer loan losses and inventory write-offs. It was diversified, profitable, and had the balance sheet to absorb the hit. NVIDIA today has $97 billion in free cash flow and genuine end-customer demand from hyperscalers building their own infrastructure. The neocloud layer could unwind entirely and NVIDIA could survive, diminished but intact. This is not a prediction of NVIDIA’s failure. It is a question about how much of the reported growth rate is organic and how much is circular.
---
The Systemic Question
CoreWeave is the largest single concentration of GPU-collateralized debt. It is not the only one.
The neocloud sector carries an estimated $40 to $55 billion or more in GPU-backed debt across a handful of companies. Crusoe Energy has approximately $10.75 billion, including $9.6 billion in project finance for a single facility in Abilene. FluidStack has secured approval to borrow more than $10 billion using NVIDIA GPUs as collateral. Nebius carries approximately $4.89 billion with 40% of its 2026 capital expenditure still needing to be raised. Applied Digital has $2.6 billion in debt. Cipher Digital has approximately $2 billion in senior secured notes. Lambda Labs has approximately $775 million. The $20 billion figure cited by industry analysts as of early 2026 significantly undercounts when project finance, convertible debt, and approved-but-undrawn facilities are included.
These companies share characteristics that amplify correlated risk. They are all secured by the same rapidly depreciating asset class. They are often financed by overlapping lenders. Blackstone led both CoreWeave facilities of $2.3 billion and $7.5 billion and is active in the broader GPU-backed lending market. Macquarie financed Lambda’s $500 million GPU loan and FluidStack’s $10 billion credit line. JPMorgan participates in Lambda, Crusoe, and the CoreWeave DDTL 4.0 bank syndicate. They depend on the same small group of hyperscaler customers: Microsoft, Meta, and Google.
NVIDIA sits at the center as supplier, investor, customer, and implicit guarantor of collateral value. The circular dynamic that operates at CoreWeave operates across the entire sector. Lambda’s relationship illustrates the pattern in its most explicit form: NVIDIA sells GPUs to Lambda, Lambda collateralizes those GPUs for a $500 million Macquarie loan, and NVIDIA leases 18,000 GPUs back from Lambda for $1.5 billion over four years, making NVIDIA simultaneously Lambda’s supplier, investor, and largest customer.
NVIDIA’s behavior toward Applied Digital is the most telling data point in the neocloud portfolio. NVIDIA invested $160 million in funding and $63.66 million in equity. The stake appreciated to over $239 million by early December 2025. Then NVIDIA sold its entire position in early 2026, all 7.7 million shares for approximately $182 million. Applied Digital is CoreWeave’s primary tenant. NVIDIA exiting its bet on CoreWeave’s landlord while maintaining its CoreWeave position has not been publicly explained.
If CoreWeave were to default and its fleet of 250,000 or more GPUs flooded the secondary market, the impact would not be confined to CoreWeave. A forced liquidation at that scale, the largest AI hardware liquidation in history, would collapse used GPU prices across the market, repricing collateral values for every other neocloud debt facility simultaneously. Shared lenders would tighten terms across all GPU-backed lending. The contagion mechanism is not theoretical; it is structural.
Stability AI provides a precedent, but at roughly fifty times smaller scale. Stability owed nearly $100 million to cloud providers and creditors by early 2024. Suppliers, including AWS, Google Cloud, and CoreWeave itself, ultimately forgave more than $100 million in debt and $300 million in future spending obligations. At that scale, suppliers could absorb the losses. At CoreWeave’s scale, lenders cannot simply forgive $21 billion.
The CFA Institute’s Enterprising Investor blog captured the systemic concern in December 2025: “When the same dollar gets counted as demand by the chip maker, revenue by the neocloud, and collateral by the lender, you don’t have a market signal. You have circular flow masquerading as growth.”
The equity and credit markets are already pricing this bifurcation for CoreWeave specifically. In April 2026, the equity market prices a $61.4 billion transformative AI infrastructure winner. In December 2025, the credit default swap market priced 773 basis points, implying a 47.5% cumulative five-year default probability at standard 40% recovery assumptions. Both positions have rational justification. Historical precedent suggests that in 2007-2008, credit markets signaled distress in financial firms six to twelve months before equity markets priced the same risk. In 2000-2001, credit spreads on telecom infrastructure companies widened well before equities bottomed.
Short interest in CoreWeave stands at 22.59% of float as of April 2026, representing 60.73 million shares. The borrow cost has fallen from a peak of 343% to 10.69% annually, making the position more accessible. Options flow is net bearish: put notional of $6.58 million nearly doubles call notional of $3.42 million on large block trades, despite an overall open interest put/call ratio of 0.81 that suggests retail holds calls while institutions quietly build bearish positions.
---
What We Don’t Know
Intellectual honesty requires acknowledging the gaps. Several critical questions remain unanswerable from public sources.
We do not know the termination provisions of the April 2026 Meta order form. The prior order form, filed in September 2025, allowed termination for cause only. The new $21 billion order form was not filed as an exhibit. If there are convenience termination rights, the headline number is less meaningful than reported.
We do not know whether the Anthropic deal is a binding $6.8 billion commitment, a smaller initial obligation with expansion options, or something else entirely. The figure comes from a single media source. CoreWeave declined to confirm it. No SEC filing was made.
We do not know what percentage of NVIDIA’s $193.7 billion in data center revenue comes from companies in which NVIDIA holds equity. Our estimate of 15 to 30% is based on reconstructing neocloud capital expenditures. NVIDIA does not disclose this figure. It is the central question of the circular financing thesis.
We do not know the cross-default provisions across CoreWeave’s debt stack. If the DDTL 3.0 enters technical default after October 28, does it trigger defaults on senior notes, convertible notes, and other term loan facilities? The answer determines whether a covenant breach is containable or cascading.
We do not know whether the April deals are genuinely take-or-pay obligations or cancellable capacity reservations. CoreWeave’s own language qualifies all backlog as “subject to the satisfaction of delivery and availability of service requirements.” The actual contract terms were not filed.
We do not know the Jane Street deal’s true nature. Is it a genuine customer relationship, a financing arrangement dressed as a customer deal, or both? The bear case notes the structure: $6 billion cloud commitment paired with $1 billion equity at a 7% discount, with no lockup disclosed and no demand registration rights. The bull case notes that Jane Street is legitimately one of the world’s largest compute consumers for quantitative modeling and stated that early access to Vera Rubin silicon represents genuine competitive advantage. The truth probably contains elements of both. Public filings cannot resolve it.
We do not know whether CDS spreads have compressed following the April equity rally. The five-year spread peaked at 773 basis points in December 2025, implying a 47.5% cumulative default probability at standard recovery assumptions. April 2026 CDS data is not publicly available through our sources. If spreads have compressed significantly, it would suggest credit markets are moving toward the equity narrative. If they remain wide, the divergence between equity euphoria and credit skepticism persists.
We do not know whether Amazon’s depreciation reversal, shortening from six years back to five, will trigger similar reassessments at Microsoft, Google, or CoreWeave itself. If it does, the impact on EBITDA-based covenant calculations would be immediate and severe.
We do not know the actual recovery rate on GPU collateral in a forced liquidation. The 40% five-year default probability derived from CDS spreads uses standard recovery assumptions of 40%. But no precedent exists for large-scale GPU seizure. Used H100 prices of $12,000 to $22,000 assume orderly sales, not a fire-sale of 250,000 units hitting the market simultaneously. If recovery rates are 20 to 30 cents on the dollar, lender losses under a CoreWeave default would be substantially worse than standard models predict, and the entire GPU-backed lending asset class would require repricing.
We do not know whether the industry is structurally shifting from take-or-pay contracts to pay-as-you-go arrangements. McKinsey has warned of a “material shift from take-or-pay to pay-as-you-go” in neocloud contracts. CoreWeave’s entire financing model depends on take-or-pay structures, because guaranteed revenue streams are the only thing that makes GPU-backed asset financing possible. A structural shift to pay-as-you-go would undermine the collateral model at its foundation.
These are the questions that determine whether CoreWeave’s $88 billion backlog represents the revenue foundation of a transformative AI infrastructure company or the largest collection of conditional commitments ever assembled atop a debt stack that cannot service itself from operations.
---
October 28
The stock surged 50% in sixteen days. The headlines declared CoreWeave vindicated. Magnetar sold $577 million. The co-founders sold on plans they had adopted in November. Nobody at the company bought a single share.
The debt clock did not stop. Interest expense runs at approximately $1.55 billion annualized and rising. Principal repayments of $250 to $300 million per month continue. The 2026 funding gap of $30 to $37 billion requires continuous access to new capital markets. The machine must keep running.
The $28 billion in April deals is real, or at least announced. But the contracts were not filed. The termination provisions are not known. The revenue depends on infrastructure that has not been built, powered by chips that have not shipped, at a company that concealed construction delays for nine months the last time execution faltered.
Equity markets and credit markets are telling different stories about the same company. The equity market sees an $88 billion backlog and a 168% revenue growth rate and prices a $61.4 billion AI infrastructure winner. The credit market saw 773 basis points in December 2025 and priced a meaningful probability of financial distress. In prior cycles, credit was right earlier. Whether that pattern holds here depends on execution.
On October 28, 2026, the unlimited equity cure window closes. After that date, CoreWeave must meet its covenant obligations organically or face restricted remedies. Either the April deals convert to revenue on schedule, the debt gets refinanced on favorable terms, the covenants get amended a second time, or the question resolves in the other direction. Six months from now, the structure will have answered the question that the April headlines could not.
We present no prediction. We present the public record and its absences. The celebration and the concern are both supported by the same set of filings. That is the most unsettling thing about this story.
The stock surged 50%. The debt clock didn’t stop.
---
This article was produced by Vesper: Public Intelligence. All claims are sourced from SEC filings, corporate press releases, earnings call transcripts, published analyst reports, and credible financial journalism cited throughout. It does not constitute investment advice. No material nonpublic information was used. Readers are encouraged to verify all claims independently through the cited sources.
For questions, corrections, or additional source material: vesperosint.substack.com
---
Sources and Key Filings:
CoreWeave 8-K, April 9, 2026 (Meta expansion + debt offerings), SEC EDGAR Acc-No: 0001193125-26-148912. CoreWeave 8-K, September 25, 2025 (Original Meta MSA + $14.2B order form), SEC EDGAR Acc-No: 0001769628-25-000050. CoreWeave 8-K, April 15, 2026 (Jane Street equity), SEC EDGAR Acc-No: 0001769628-26-000167. CoreWeave 8-K, January 2, 2026 (DDTL 3.0 covenant amendment), SEC EDGAR. CoreWeave 10-K, FY2025. NVIDIA Schedule 13G/A, January 26, 2026, SEC EDGAR. NVIDIA Q4 FY2026 earnings call transcript, February 25, 2026. Kerrisdale Capital, “Artificial Returns: CoreWeave Inc (CRWV),” September 2025. Masaitis v. CoreWeave, Inc., 2:26-cv-00355, U.S. District Court, District of New Jersey. Jim Chanos, “Monetary Matters” podcast with Jack Farley, December 2025. CoreWeave Form 4 filings (multiple), SEC EDGAR, CIK 1769628. MarketBeat short interest data. S3 Partners borrow cost data. Fintel short interest data.