CoreWeave

The downsized IPO that quadrupled, halved, and — seventeen months in — still doubled: AI demand bailed out every disclosed risk except the income statement

CoreWeave priced at $40 on March 28, 2025, cut from a $47-55 range, and the market's initial verdict matched the skeptics: a 0% first-day pop and a trough of $35.42 (-11.4%) by April 21. Then the disclosed concentration risk inverted into the bull case. The ~$11.9 billion OpenAI contract signed in March 2025 (with OpenAI taking a $350 million stake in the IPO) began a customer land-grab, and by June 20, 2025 the stock hit $183.58 — up 359% from the IPO price in under three months, the most violent re-rating of any 2025 listing.

The backlog kept outrunning the skeptics. OpenAI expanded its deals by up to $6.5 billion to roughly $22.4 billion total; Meta signed a $21 billion infrastructure agreement; Anthropic and Jane Street (a $6 billion cloud commitment plus $1 billion equity investment) followed; and by Q1 2026 CoreWeave reported a revenue backlog approaching $100 billion against $15.1 billion of remaining performance obligations at filing. Fiscal 2025 revenue came in at $5.13 billion, up 168%. Microsoft dependence was genuinely diluted — but replaced by concentration in OpenAI and Meta, so the top-heavy structure survived with new names. The company even joined the Nasdaq-100 about fifteen months after listing.

What the demand boom couldn't fix was the model's cost of capital. Fiscal 2025 operating income was roughly breakeven (-$17 million) — but interest and financing costs drove a $1.2 billion net loss, worse than 2024's, and the trailing-twelve-month loss reached $1.93 billion by mid-2026. Total debt scaled from the $8.0 billion disclosed in the S-1 to roughly $35.1 billion by June 30, 2026, with a further $8.5 billion financing sought in February 2026 collateralized by the Meta contracts and a $3.5 billion convertible notes offering in April 2026. The S-1's 'innovative financing structures' pitch proved scalable — and proved exactly as expensive as the risk factors warned.

The stock told both stories at once. The attempted $9 billion all-stock acquisition of Core Scientific — CoreWeave's move to own its power and data center shells — was rejected by Core Scientific shareholders on October 30, 2025, and between October 31 and November 28, 2025 the stock fell from $133.71 to $73.12 as AI-capex sentiment cracked across the sector. 2026 brought whipsaws between roughly $72 and $112 (52-week range $60.55-$153.20), a $2 billion Nvidia investment at $87.20 per share in January 2026 deepening the circular supplier-customer-investor relationship, and an August 28, 2026 close of $84.23 — up 110.6% from the IPO but down 54% from the June 2025 peak.

Seventeen months is early for a company whose contracts run to the 2030s. The demand thesis has been vindicated beyond the prospectus's own best case; the financial-structure risks have compounded rather than resolved; and whether a business that borrows at this scale to buy depreciating GPUs ever converts a $100 billion backlog into profit remains the open question the S-1 posed and hindsight has not yet answered.

CoreWeave is a specialized 'AI Hyperscaler' — a GPU cloud purpose-built for AI training and inference workloads, selling access to Nvidia-based infrastructure through multi-year take-or-pay committed contracts. The pitch: hyperscalers can't build AI-optimized capacity fast enough, so a focused operator that pioneered GPU-collateralized financing can out-deploy them and capture the AI compute build-out, converting a $15.1 billion backlog into contracted, recurring revenue.

Revenue went from $16 million to $1.9 billion in two years, and the demand side of the story writes itself: every frontier AI lab needs more compute than exists. If scaling laws hold, CoreWeave's take-or-pay contracts (96% of revenue, two-to-five-year terms, prepayments of 15-25% of contract value) give it hyperscaler-grade revenue visibility without hyperscaler legacy costs. Its pioneering GPU-backed debt structures let it deploy capital faster than anyone, and as it graduates to cheaper financing — securitizations, rated parent-level debt — unit economics improve while the moat of operational expertise (first to new Nvidia generations, purpose-built data centers) compounds. The company becomes the fourth cloud, built for the AI era.

With hindsight: The growth half of the best case was exceeded: revenue nearly tripled to $5.13 billion in 2025, the take-or-pay model attracted OpenAI (~$22.4B), Meta ($21B), Anthropic, and Jane Street, and backlog approached $100 billion. But the economic half — falling cost of capital producing profitability — failed so far: debt more than quadrupled to ~$35.1 billion, net losses widened to $1.2 billion in 2025 and $1.93 billion TTM by mid-2026, and the stock at $84.23 (+110.6%) sits less than half its June 2025 peak. Hypergrowth realized; the profitable fourth cloud, not yet.

• Hypergrowth is real and contracted, not speculative — revenue grew 118x in two years.

  “Our revenue was $16 million, $229 million, and $1.9 billion for the years ended December 31, 2022, 2023, and 2024, respectively” source ↗

• AI demand for compute keeps compounding as models scale and end-user demand grows.

  “As demand continues for AI services, AI providers, including our customers, have sought increased compute capacity to enable advancements in their AI models and service the demands of end users” source ↗

• CoreWeave's financing innovation is itself a moat, and its cost of capital should fall with scale.

  “we have pioneered and scaled innovative financing structures that have enabled us to grow our business through timely and flexible access to capital” source ↗

1. Microsoft is 62% of revenue — extreme customer concentration

   partly came true

   platform dependence · structural

   Two customers drove 77% of 2024 revenue, and one of them — Microsoft — was 62% alone. Any change in Microsoft's demand, its ability or willingness to perform, or the broader relationship flows straight through the income statement. Concentration is expected to persist for the foreseeable future.

   “For the years ended December 31, 2023 and 2024, our largest customer was Microsoft, which accounted for 35% and 62% of our revenue, respectively.” source ↗

   What happened: The feared loss of Microsoft spend never hit — revenue grew 168% in 2025. But concentration itself persisted, merely rotating: OpenAI's contracts reached ~$22.4 billion and Meta signed a $21 billion agreement, so a handful of frontier labs still dominate the backlog. The risk transformed from Microsoft-dependence into frontier-lab-dependence.

2. $8 billion of debt against a two-year-old revenue base

   came true

   financing · structural

   CoreWeave carries $8.0 billion of total indebtedness plus $2.6 billion of operating lease liabilities, built through delayed-draw term loans (DDTL 1.0 and 2.0) at variable rates. Roughly a third of operating cash flow already goes to debt service, and the entire model depends on continuously rolling and expanding this leverage.

   “We have a substantial amount of debt, which requires significant interest and principal payments. As of December 31, 2024, our total indebtedness was $8.0 billion” source ↗

   What happened: Leverage deepened dramatically: total debt reached roughly $35.1 billion by June 30, 2026, with an additional $8.5 billion financing sought in February 2026 against Meta contracts and $3.5 billion of convertibles in April 2026. No default occurred, but financing costs turned near-breakeven 2025 operating results (-$17 million) into a $1.2 billion net loss — the cash-flow-diversion harm the risk factor described, at 4x the scale.

3. AI demand could plateau if scaling laws break

   didn't happen

   market · structural

   The whole business is a levered bet on continued exponential demand for AI training compute. The filing concedes that compute-efficiency advances (it was written weeks after DeepSeek) or a plateau in scaling could shrink demand — a fatal scenario for an operator with take-or-pay obligations to suppliers and landlords.

   “In the event that existing scaling laws do not continue to apply as they have in the past, demand by our customers for compute resources, including our solutions and services, may not continue to increase over time” source ↗

   What happened: Through August 2026, demand kept compounding: backlog approached $100 billion by Q1 2026 and revenue grew 168% in 2025. Market sentiment about AI capex cracked twice — the stock halved between October and November 2025 and slid again in mid-2026 — but the customers kept signing bigger contracts.

4. Total dependence on Nvidia as sole GPU supplier

   didn't happen

   platform dependence · structural

   Every GPU in the fleet is Nvidia's — contractually required by customers — and three suppliers were 76% of 2024 purchases, procured on just-in-time purchase orders with no long-term supply guarantees. Blackwell delays already demonstrated the exposure; Nvidia is simultaneously supplier, customer, and the industry's kingmaker.

   “as a result of our obligations in our current customer contracts, all of the GPUs used in our infrastructure today are NVIDIA GPUs” source ↗

   What happened: The dependence deepened but paid off: CoreWeave was first to deploy Nvidia's GB200 NVL72, GB300 NVL72, and Rubin platforms, and Nvidia invested $2 billion in CoreWeave at $87.20 per share in January 2026. No supply disruption materially harmed the business — though the circular supplier-customer-investor relationship grew tighter, not looser.

5. Insatiable capital expenditure needs

   partly came true

   financing · structural

   Growth requires buying GPUs and building data centers years ahead of revenue, forever. If capital markets close or terms worsen, CoreWeave cannot fund the expansion its contracts assume; if it raises equity or convertible debt, existing holders get diluted. The growth model is a capital-markets treadmill.

   “We require substantial capital expenditures to support our growth and respond to business challenges. We have made significant financial investments in our business, and we intend to continue to make such investments in the future” source ↗

   What happened: The capital hunger arrived exactly as disclosed — funding UK expansion (£1.5 billion), a $6 billion Lancaster, PA data center, and Asia-Pacific entry — but the feared inability to raise never came: markets supplied tens of billions in debt, converts, and equity (including Nvidia's $2 billion and Jane Street's $1 billion) through mid-2026, at the cost of a balance sheet carrying ~$35 billion of debt.

6. Widening net losses with no proven path to profitability

   came true

   profitability · structural

   Net losses grew from $31 million to $863 million in two years alongside a $1.5 billion accumulated deficit, with interest expense compounding as debt scales. Post-IPO stock-based compensation from performance RSUs adds another expense wave.

   “We incurred net losses of $31 million, $594 million, and $863 million for the years ended December 31, 2022, 2023, and 2024, respectively, and we may not achieve or, if achieved, sustain profitability in the future” source ↗

   What happened: Net loss widened from $863 million (2024, per the S-1) to roughly $1.2 billion in fiscal 2025 and $1.93 billion on a trailing-twelve-month basis by June 30, 2026. Operating income swung from +$324 million in 2024 to -$17 million in 2025; interest expense did the rest. Profitability has not arrived.

7. GPU useful-life estimates could prove optimistic

   unclear

   tech & security · structural

   Reported economics hinge on assumptions about how long GPUs stay productive and whether they can be redeployed past their contracted life. Faster Nvidia release cycles compress the value of prior generations; any change in useful-life estimates would hit depreciation and reported results across a multi-billion-dollar asset base.

   “This requires us to make certain estimates with respect to the useful life of the components of our infrastructure and to maximize the value of the components of our infrastructure, including our GPUs, to the fullest extent possible. We cannot guarantee that our estimates will be accurate” source ↗

   What happened: GPU depreciation became the central bear argument against the whole neocloud sector in late 2025 and 2026, but no impairment, restatement, or useful-life revision at CoreWeave has been verified. With Nvidia's release cadence accelerating (Blackwell to Rubin within the period), the question remains open and unresolved.

8. Competing against hyperscalers who are also its customers

   didn't happen

   competition · structural

   AWS, Google, Microsoft, Oracle, and IBM all offer competing GPU clouds with vastly deeper pockets — and several are simultaneously CoreWeave customers who could in-source capacity at any time. Smaller AI-focused clouds (Crusoe, Lambda) attack from below.

   “Key competitors that offer general purpose cloud computing as part of a broader, diversified product portfolio include Amazon (AWS), Google (Google Cloud Platform), IBM, Microsoft (Azure), and Oracle, a number of which are also our current customers” source ↗

   What happened: Through mid-2026 CoreWeave won share rather than losing it — Meta, OpenAI, Anthropic, and Jane Street chose it over or alongside the hyperscalers, and it entered the Nasdaq-100. The hyperscalers' own capacity build-outs continued, but no competitive displacement of CoreWeave's business is evident yet.

9. The take-or-pay contract model is unproven through a cycle

   didn't happen

   growth · serious

   96% of revenue comes from committed take-or-pay contracts — a model only two years old at this scale. If the industry shifts toward pay-as-you-go, revenue visibility, prepayment-funded economics, and the debt structures collateralized by those contracts all weaken together.

   “For the years ended December 31, 2022, 2023, and 2024, committed contracts accounted for 20%, 88%, and 96% of our revenue, respectively. There is no guarantee that in the future customers will continue to be willing to enter into, and that the industry will continue to support, a take-or-pay model” source ↗

   What happened: The model scaled instead of breaking: committed take-or-pay contracts grew from $15.1 billion of remaining performance obligations at filing to a backlog approaching $100 billion by Q1 2026, with the industry's largest players (OpenAI, Meta) signing multi-year commitments. It still hasn't been tested by a genuine demand downturn.

10. Power is the binding constraint on growth

    unclear

    operations · serious

    CoreWeave has already run short of power for customer projects and paid service credits for it. Data center expansion depends on utilities, grids, and regulators it doesn't control, in a market where every competitor is chasing the same megawatts.

    “We have in the past experienced, and we may in the future experience, insufficient power to service a customer’s project. Any limitation on the delivered energy supply would limit our ability to operate our platform” source ↗

    What happened: Power remained the industry's choke point and motivated CoreWeave's $9 billion bid for Core Scientific's powered data center footprint — which shareholders rejected in October 2025, leaving CoreWeave leasing rather than owning its power. No specific CoreWeave power failure has been verified, but the constraint clearly shaped strategy.

11. Restrictive covenants and GPU-collateralized facilities invite foreclosure in a downturn

    didn't happen

    financing · serious

    The credit facilities restrict dividends, acquisitions, asset sales, and additional debt, require financial covenant compliance, and are secured — a default lets lenders seize the collateral behind the DDTL facilities, the very infrastructure the business runs on.

    “the lenders under our DDTL 1.0 Facility and DDTL 2.0 Facility instituting foreclosure proceedings against their collateral, any of which could materially adversely affect our business, operating results, financial condition, and future prospects” source ↗

    What happened: No default or foreclosure occurred; lenders kept extending. The collateralized-financing model actually expanded — the February 2026 $8.5 billion financing was secured by Meta infrastructure contracts — meaning the structure the risk factor warned about became the company's primary growth engine.

12. Unremediated material weaknesses in internal controls

    unclear

    governance · serious

    The company identified material weaknesses spanning IT general controls, segregation of duties, and insufficient qualified accounting personnel. Remediation is expected to run into 2026 — meaning a company with $8 billion of debt and 62% customer concentration is reporting numbers through admittedly deficient controls.

    “while preparing the financial statements that are included in this prospectus, we identified material weaknesses in our internal control over financial reporting” source ↗

    What happened: The S-1 projected remediation extending into 2026. No restatement, control failure, or confirmation of successful remediation has been verified in the sources checked; the company continued filing timely results through mid-2026.

13. Counterparty credit risk from cash-burning AI customers

    unclear

    other · serious

    Future growth depends on selling multi-year committed contracts to AI startups that are themselves unprofitable and dependent on venture financing. CoreWeave carries no credit insurance; a customer failure converts contracted backlog into stranded capacity.

    “We do not currently maintain credit insurance to insure against customer credit risk. If our customers fail to fulfill their contractual obligations, it may have an adverse effect on our business, operating results, financial condition, and future prospects” source ↗

    What happened: No customer default has been verified — but the exposure grew enormously: the largest counterparty, OpenAI, committed ~$22.4 billion while itself burning cash at historic rates. Whether frontier labs can pay for their committed compute is now the systemic question hanging over the entire AI-infrastructure trade.

14. Series C 'put right' could force cash redemptions of stock

    unclear

    financing · serious

    Series C holders can put their converted Class A shares back to the company for cash at the original issue price plus accrued dividends unless the stock sustains 175% of the Series C price — an unusual structure that makes public shareholders junior to a preferred investor's downside protection.

    “We may not have enough available cash at the time we are required to repurchase the Put Shares, and we may be required to adopt one or more alternatives, such as selling assets, restructuring debt or obtaining additional equity or debt capital on terms that may be onerous or highly dilutive” source ↗

    What happened: No forced redemption has been reported. Given the stock's run to $183.58 (4.6x the $40 IPO price) by June 2025, the price-based termination condition for the put right was plausibly triggered, but the Series C original issue price was blank in the filing and no confirmation of termination has been verified.

15. Dual-class structure entrenches the three co-founders

    unclear

    governance · serious

    Class B shares carry ten votes and are held entirely by the co-founders, who also hold equity exchange rights letting them convert future option shares into super-voting stock — concentrated control layered on top of staggered boards and other anti-takeover provisions.

    “Our Class B common stock has ten votes per share, and our Class A common stock has one vote per share.” source ↗

    What happened: The structure persists and the co-founders retain voting control; no governance dispute or shareholder harm attributable to it has been verified in the period. Notably, entrenchment cut the other way at Core Scientific, whose shareholders were free to — and did — reject CoreWeave's $9 billion offer on October 30, 2025.

16. A crypto-mining pivot with a two-year operating history at scale

    didn't happen

    other · serious

    Before 2022 CoreWeave was substantially a crypto miner (founded as Atlantic Crypto). The AI cloud business — the entire investment case — has effectively two years of history, making every forecast about margins, GPU economics, and demand durability an extrapolation from a tiny base.

    “Moreover, prior to 2022, we had limited revenue, most of which was derived from our crypto mining offerings, which we have discontinued” source ↗

    What happened: The short operating history didn't prevent execution: the former Atlantic Crypto scaled revenue to $5.13 billion in 2025 and $7.59 billion TTM by mid-2026 while deploying successive Nvidia generations first. Ironically, CoreWeave then tried to buy another crypto-miner-turned-AI-host (Core Scientific) for its data centers.

How the S-1 reads

A leveraged single-factor bet dressed as infrastructure: CoreWeave stacked every classic S-1 warning — customer concentration, supplier concentration, massive secured debt, material weaknesses, dual-class control, a pivot from a discredited prior business — onto one variable, AI compute demand. It is the purest example in the archive of concentration risk running in a circle: the sole supplier (Nvidia) is also a customer and investor, the largest customer (Microsoft) is also a competitor, and the debt is collateralized by the depreciating hardware the supplier sells. When every risk correlates to the same macro bet, the filing's 60+ pages of risk factors reduce to one sentence: this works if and only if AI capex keeps accelerating.

• “Investors should not rely on the revenue growth of any prior quarterly or annual period as an indication of our future performance” source ↗

  The prospectus itself warns investors off extrapolating the growth that is the entire pitch.

• “We recognized an aggregate of approximately 77% of our revenue from our top two customers for the year ended December 31, 2024” source ↗

  Concentration on both sides of the business: two customers on one end, three suppliers on the other.

• “For the year ended December 31, 2024, approximately 32% of our net cash provided by operating activities, before giving effect to the payment of interest, net of capitalized amounts, was dedicated to debt service, both principal and interest” source ↗

  A third of operating cash flow was already spoken for by lenders before the company went public.

• “for the year ended December 31, 2024, three suppliers accounted for 46%, 16%, and 14% of total purchases” source ↗

  Supplier concentration mirrors the customer concentration.

What this one teaches

• Customer concentration in a supply-constrained market can be the bull case, not the bear case: when compute is scarce, a 62%-customer S-1 became a bidding war among OpenAI, Meta, and Anthropic for the same capacity.
• A downsized IPO range tells you about the moment's sentiment, not the asset: CoreWeave cut its range, priced at $40, and traded at $183.58 within twelve weeks. Pricing dynamics and business trajectory are different signals.
• 'Innovative financing structures' scale in both directions — the GPU- and contract-collateralized debt that enabled 168% growth also quadrupled to ~$35 billion and converted near-breakeven operations into billion-dollar net losses. Leverage is the thesis, not a footnote.
• When a company's risk factors all correlate to one macro variable, the stock trades like a leveraged option on that variable: CoreWeave rose 359%, fell 45% in a month on AI-capex sentiment, and whipsawed through 2026 with a 52-week range of $60.55-$153.20.
• Vertical-integration rescues aren't guaranteed: the market cheered CoreWeave's plan to buy its power footprint, but Core Scientific's shareholders — not CoreWeave's entrenched founders — held the votes, and all-stock deals break when the acquirer's currency is this volatile.
• Judge hypergrowth IPOs on two clocks: seventeen months in, CoreWeave's demand thesis is emphatically validated and its profitability thesis emphatically not — and the stock's +110.6% vs. -54%-from-peak captures exactly that unresolved split.

The paper trail

1. 2025-03-03 S-1 filing index ↗ document ↗
2. 2025-03-12 S-1/A filing index ↗ document ↗
3. 2025-03-20 S-1/A filing index ↗ document ↗
4. 2025-03-31 424B4 filing index ↗ document ↗

Filed as CoreWeave, Inc.. All documents are public domain, served by SEC EDGAR.