On October 23, CoreWeave closed at $54.24. That is down 70% from June 2025’s high of $183.58, down 38% from the $87.20 Nvidia paid for $2 billion of stock in January, and half the $109 Jane Street paid in April.
The company grew revenue to $2.58 billion in the June quarter, up 112% year over year, and its revenue backlog reached $104.2 billion, up 246%, before counting more than $25 billion of commitments added in July and early August. And one fact that organizes everything else here: at Friday’s close the equity is worth $29.9 billion while the debt stands near $39.8 billion. The market values CoreWeave at three quarters of what it owes. Stocks don’t typically trade like that. Credits do though. The rest of this attempts to answer the question of whether the market is right to price it that way, and my initial inclination is that I don’t think it is.
Explaining The Business
Strip the branding and a neocloud is actually a pretty simple machine. Sign a long-term take-or-pay contract with someone who needs GPUs. Borrow against that contract to buy the chips, the power, and the building. Rent the computers back to the customer. Pocket the spread between the rental rate and the cost of capital plus depreciation. See, simple!
CoreWeave booked 98% of its June-quarter revenue from committed, take-or-pay contracts. This is very obviously not a cloud business in the way AWS is. It is an equipment leasing business whose equipment loses bidding power to a newer model every twelve months.
Jensen Huang (who I deem the “King of AI”) said last week at a conference in Berlin that a one-gigawatt AI factory costs “about $50 billion” now, and Bank of America models about $57 million per megawatt. On revenue, there are two prices: short-duration capacity, which is three to six months and is clearing around $40 million per megawatt annualized (CoreWeave’s September materials; Nebius sees the same $40 to $50 million on its new short-term deals, and Huang used “$40 billion per year” for a gigawatt in Berlin); and long-term contracts, which price lower: Nebius’s four recent billion-dollar, multi-year deals yield $20 to $25 million per megawatt. So the machine costs about $50 to build and prints somewhere between $20 and $40 a year depending on how long the rate is locked, while the chips stay rentable. The entire sector argument lives inside that methodology. See, simple!
The Size of the Buildout
We start with the demand, which is where the hyperscalers are spending record levels of capex. Microsoft spent $35.8 billion on capital expenditures in their last reported quarter. Alphabet spent $44.9 billion. Oracle’s remaining performance obligations — its promised-but-undelivered cloud capacity — went from $138 billion to $664 billion in the last five quarters. Nvidia booked $96.2 billion of revenue in its July quarter, and the King of AI claimed he has a vision of a trillion dollars of orders back at an event in March. BofA sizes the entire spending at roughly $5 trillion of cumulative data center capex between 2026 and 2030, of which about $1.2 trillion must be raised externally, mostly as debt issued by the companies financing it.
The neoclouds exist precisely because hyperscaler capex still isn’t enough at that eye-popping dollar figure and because the frontier LLM labs (OpenAI, Anthropic, and many others) want capacity no competitor currently has. But a $1.2 trillion debt bill means the sector’s binding constraint is not chips or customers. Demand is off the charts. The capital to build them has not kept pace.
The Players
CoreWeave is not one of several similar companies. It is the public neocloud sector, with a supporting cast at a tenth of its scale and size.
Nebius ($NBIS) is the closest true peer, running the same playbook a year behind: $582 million of June-quarter revenue growing 454%, four contracts averaging over $1 billion signed in a single quarter. Over $9 billion of customer prepayments expected this year, its first GPU-backed loan in July at SOFR plus 2.50%, and a $3.5 billion secured bridge from Citibank signed on October 17.
IREN ($IREN) is the miner-convert case study. It wrote off $639 million of bitcoin rigs this fiscal year (and took a $703 million net loss doing it), but now carries $4 billion of contracted AI cloud ARR anchored by Microsoft, with planned capex of $25 to $30 billion, per its earnings-call coverage, against $707 million of trailing revenue.
Applied Digital ($APLD) and Core Scientific ($CORZ) are landlords, not clouds: they build powered shells and lease them, overwhelmingly to CoreWeave. Core Scientific describes a single-customer concentration in its own filings. Applied Digital issued CoreWeave warrants on 21.5 million of its shares alongside the leases, which means the tenant holds equity upside in the landlord.
Cut from the comp set, with reasons: Cipher ($CIFR; $25 million of quarterly revenue, still mostly mining), TeraWulf ($WULF; same category), and Galaxy ($GLXY; its revenue line is mostly trading flow, and the Helios data center business is real but not separable on a comparable basis).
The Nvidia Relationship
The relationship all companies in the AI space have with Nvidia is probably the most important one, but maybe none more important than CoreWeave’s. Nvidia is the supplier: the $33.8 billion of technology equipment at the core of CoreWeave’s $52.6 billion gross asset base is overwhelmingly Nvidia silicon. Nvidia is a customer: a master services agreement signed April 10, 2023. Nvidia is a shareholder: $2 billion of Class A stock was purchased in January at $87.20. And Nvidia is, by its own press release, positioning as something close to a co-financier: the companies intend to “leverage NVIDIA’s financial strength” to accelerate CoreWeave’s buildout toward more than 5 gigawatts by 2030.
| the relationship | what the filings say |
|---|---|
| Nvidia sells to CoreWeave | $33.8B of technology equipment, overwhelmingly Nvidia silicon |
| Nvidia buys from CoreWeave | MSA dated April 10, 2023; $6.3B order form, September 2025 |
| Nvidia owns CoreWeave | $2.0B of Class A purchased at $87.20, January 2026 |
| Nvidia backstops CoreWeave’s utilization | $6.3B order form: must buy residual unsold capacity through April 2032, subject to delivery conditions |
| CoreWeave holds upside in its landlord | warrants on 21.5M Applied Digital shares, roughly 7% of the company |
| Customers fund CoreWeave | $9.7B of prepayments on the balance sheet today |
So the supplier invests in the customer, which pledges the supplier’s chips as collateral to borrow money to buy more of the supplier’s chips, to rent to AI labs whose own funding rounds are premised on the scarcity of those chips. None of this is hidden or nefarious, but it means every number in this sector, revenue, backlog, collateral, equity stakes, moves with one obvious variable: the price of compute.
The Contracts
The backlog went $9.9 billion, $15.1 billion, $66.8 billion, $104.2 billion across 2023 to mid-2026 (does that seem sustainable?). All the big names are seen as customers. Microsoft was 67% of 2025 revenue; the top three customers were still 72% of revenue in the most recent quarter as the business expanded. OpenAI has committed roughly $11.9 billion through October 2030 plus another $6.5 billion under a second agreement. Meta committed $14.2 billion in September 2025, then signed a new order on March 31 with an initial commitment of approximately $21 billion, which included exercising the expansion option from that earlier order and runs through April 2032.
Three important details: 1) The timing: 41% of the $103.7 billion of remaining performance obligations converts to revenue inside 24 months, and roughly $21 billion a year of already-signed revenue through mid-2028; 2) The prepayments: customers pay 15 to 25% of the contract value up front, which is why deferred revenue sits at $9.7 billion and why this company has reported generating $2 billion of quarterly operating cash flow on only $189 million of revenue. The customers are partially financing their own capacity. And 3) the price direction: As stated above, short-dated contracts are roughly $40 million per megawatt, with reported “significant demand for prior generations” of chips. For now, pricing is rising, not falling. If you are short this company or AI in general, this will need to reverse.
The Debt
| instrument | balance 6/30/26 | effective rate | matures |
|---|---|---|---|
| DDTL 1.0 facility | $1.3B | 15% | Mar 2028 |
| DDTL 2.0 facility | $3.19B | 11% | Aug 2030 |
| DDTL 2.1 facility | $3.0B | 9% | Mar 2031 |
| DDTL 3.0 facility | $2.22B | 9% | Aug 2030 |
| DDTL 5.0 facility | $1.1B | 9% | Nov 2031 |
| 9.25% senior notes | $2.0B | 10% | Jun 2030 |
| 9.00% senior notes | $1.75B | 10% | Feb 2031 |
| 9.75% senior notes | $2.75B | 10% | Oct 2031 |
| 9.625% senior notes | $1.25B | 10% | Jul 2032 |
| 8.50% EUR senior notes | $2.28B | 9% | Jul 2032 |
| 1.75% convertible 2031 | $2.59B | 2% | Dec 2031 |
| 1.75% convertible 2032 | $4.0B | 2% | Oct 2032 |
| OEM and software financing | $4.22B | 11% | 2026 to 2030 |
| Magnetar loan | $189M | 12% | Jan 2029 |
| DDTL 4.0 facility (non-recourse) | $2.84B | 7% | Mar 2032 |
| OEM financing (non-recourse) | $0.88B | 9% | 2026 to 2028 |
Contractual interest was $592 million in the past quarter; the filings’ interest expense, net line was $640 million, an annualized $2.6 billion that consumes about 42% of run-rate adjusted EBITDA. There is a $10.6 billion maturity wall across 2027 and 2028. On December 31, 2025, covenants on one facility were amended after deliveries slipped (filed January 2). The same amendment started a clock: unlimited equity cures for the loosened covenants run only through October 28, 2026, five days after this piece’s price date; after that they are capped. And $79 million of quarterly interest never touches the income statement at all, because it is capitalized into the assets.
There is a bull reading of the same table. CoreWeave’s average cost of debt has fallen every period: 14.9% in 2023, then 12.2%, then 9.0%, then 8.3%, and September’s $4.2 billion convertible was printed at 2.875%, which pulls the average down further. More of the riskiest borrowing is non-recourse, meaning that if those loans go bad, the lenders can take the specific chips and contracts behind them and nothing else. And the money is borrowed against signed take-or-pay contracts with Meta and Microsoft, not against hoped-for demand. Both readings are true at once. Which one wins depends on whether credit markets stay open to this company for the next eighteen months, and that is exactly where BofA says the strain is showing.
The Margins
June quarter: adjusted EBITDA margin of 59%. GAAP operating margin: -2%. Net loss: $626 million. The distance between those numbers is depreciation and interest and it’s the whole business.
The GPUs depreciate over six years, in a near straight line. Nvidia ships a new architecture roughly every year; CoreWeave validated Vera Rubin systems this quarter, the fourth generation inside a single depreciation window. Gross technology equipment went from $9.1 billion at the end of 2024 to $33.8 billion at June’s reporting, meaning the fleet is brand new and nearly all of its depreciation is still in front of it. Depreciation and amortization ran 55% of revenue in the first half. Even the company’s preferred adjusted operating margin fell from 16% to 5% year over year, and the adjusted EBITDA margin slipped — from 62% to 59%. Margins are compressing while revenue doubles.
Bulls state that depreciation overstates economic decay: older chips keep earning (as stated above, the company says demand for prior generations is still strong), and the contracts were priced to recover the capex within their terms. The honest version of the bear case is that nobody knows the residual value of a five-year-old GPU in 2030, but the six-year straight line is an estimate about exactly that. For the depreciation schedule to be conservative, mid-life rental pricing has to hold. Will it?
The Case Against
Bank of America has spent six weeks giving the bear case institutional cover, and I won’t attempt to rewrite their thesis on it, so if you are interested in their full coverage it is here and here, and below I will provide some key highlights.
First, the financing math doesn’t close. $5 trillion of capex needs $1.2 trillion of external money, and the marginal lender is already fatigued (see: Firmus, pulled; see: CoreWeave’s own recourse delayed-draw spreads, SOFR plus 4.50% in May to SOFR plus 5.50% by August), and U.S. business spending on AI has flatlined. Second, the depreciation is real: a six-year schedule on an asset whose pricing power is contested annually means today’s 59% EBITDA margin partly measures accounting policy, not economics. Third, the circularity cuts the safety margin: when the supplier funds the customer that collateralizes the supplier’s product, stress anywhere transmits everywhere. Fourth, the history: BofA studied earlier infrastructure cycles and landed on the conclusion that the technology can succeed and the financiers can still lose money. Railroads transformed America. So did the bondholders who funded them, into creditors of bankrupt estates. Ba dum tssss!
I think they might be describing the wrong security though. Every mechanism they cite (spread widening, funding fatigue, depreciation truth-telling) hits the equity through one channel: CoreWeave’s ability to keep rolling and raising debt. So the question is not “is the buildout risky.” It is “who is standing behind the borrower.”
The Backstop
On September 9, 2025, Nvidia signed an order form with an initial value of $6.3 billion under which, in the filing’s own words, if CoreWeave’s data center capacity is not fully utilized, Nvidia “is obligated to purchase the residual unsold capacity” through April 13, 2032, subject to delivery and service-availability requirements. That makes the supplier the buyer of last resort for unsold capacity, for another five and a half years. It is not, however, unconditional: the following is some boilerplate legalese language, but either party can terminate the underlying agreement if the other lands in a bankruptcy which is not dismissed within 90 days. The contract protects CoreWeave from empty racks; it does not protect them from insolvency. On January 23, 2026, Nvidia bought $2 billion of stock at $87.20, 61% above today’s price, and the joint release says the companies intend to “leverage NVIDIA’s financial strength” for CoreWeave’s procurement of land, power, and shell, a stated intention rather than a signed obligation. Add the supports already inside the contracts: customers prepay 15 to 25% of total contract value (there is $9.7 billion of their money on the balance sheet right now), and the biggest counterparties, the unnamed top customer at 36% of June-quarter revenue and Meta with roughly $35 billion committed, are the companies least able to afford losing their reserved capacity in the middle of training a model.
Why would Nvidia keep doing this? Not because the dollars matter to them: $6.3 billion of capacity purchases and $2 billion of stock round to nothing for a company worth $5 trillion. And the point is not that Nvidia stands ready to take over CoreWeave in a DEFCON 1 situation; what they are protecting is the broader AI trade. CoreWeave is Nvidia’s proof that it can sell chips at scale to someone other than the hyperscalers. Huang needs credible alternatives to Microsoft, Google, and Amazon, who are all building their own silicon; the neoclouds are that alternative; CoreWeave is most of the public neocloud sector by revenue. A CoreWeave bankruptcy would reprice every GPU-backed loan on earth, and Nvidia’s order book with it. When Huang says the binding constraint is capital rather than demand, as he did in Berlin, he is telling you where his incentives point: it is cheaper for Nvidia to keep CoreWeave funded and full than to let that doubt spread, and it has paid for exactly that twice, $6.3 billion in September 2025 and $2 billion more in January.
The limit of the argument needs stating too: Nvidia’s backstop covers the capacity, not the stock price. Nvidia buying unsold capacity keeps the lenders whole and the machine running; it does nothing for shareholders who overpaid. If trouble comes, the rescue will look like new money coming in at distressed prices, and existing shareholders will own less of the company afterward. The backstop keeps the data centers from sitting empty; it does not, by its own terms, keep the company from running out of financing. What it buys is time and the lenders’ confidence, which for a company that has to keep borrowing is most of what matters. Which is exactly why my target is $70 and not $100, and why my target is below what Nvidia paid for it at the beginning of 2026.
The Bonds
CoreWeave’s 9.75% senior notes due October 2031 traded near 80 cents on the dollar last week, a yield to maturity around 15.6%. In mid-August the same bonds yielded about 11.3%. That is an enormous move in two months, and it happened while the five-year Treasury sat near 4.9%. Lenders now charge CoreWeave roughly ten percentage points more than the U.S. government, a rate reserved for companies the market is actively worried about.
Here is the math for a simple stock investor (like myself). The notes cost about 80 cents on the dollar today. Hold them to maturity in 2031 and you collect the 9.75% coupon each year for five years, then get repaid at 100, about 19 points more than you paid. Add it up and that is roughly an 84% total return, so long as CoreWeave simply pays its bills. My base case for the stock returns 88%. Nearly the same number, except the bondholder gets paid before the stockholder, collects cash twice a year while waiting, and wins in three of the four futures below.
| the future | the stock (from $54.24) | the 9.75s of 2031 (from ~80) |
|---|---|---|
| Bull: signings compound | ~+252% to $191 | ~+84%, no higher |
| Base: backlog converts | ~+88% to $102 | ~+84% |
| Muddle: growth stalls but the bills get paid | flat to down | ~+84% |
| Default or restructuring | most or all of it lost | some money back, after ~$19B of secured lenders are paid first |
A few things that need to be pointed out before the bonds are treated like a pseudo-savings account: the bonds can lose money without CoreWeave ever filing for bankruptcy. The standard move for a company under stress is to offer bondholders a trade: new bonds worth less than the old ones, but backed by better collateral, and the message is take the deal or risk worse in a bankruptcy. CoreWeave’s debt already sits in separate subsidiaries built around the GPUs, which is exactly the structure that makes a trade like that easy to run. Bondholders can lose twenty points this way without a bankruptcy ever being filed.
Also, and the main problem for the people I write for: the average Joe is unable to buy them. The notes are something called 144A paper, sold only to qualified institutional buyers, so access runs through high-yield funds or institutional groups with minimums required to invest. The cleanest expression of the view “CoreWeave does not default” is locked behind a gate most cannot easily gain access through. This aspect, and knowing it exists, is genuinely irritating if you are bullish. The good news though is if I, or anyone else, wants to gamble on Bolivian squash or the weather in a random city, best believe that is accessible to us now!
It also sharpens what the stock is for: if the bonds pay 15.6%, the only reason to own the stock instead is the chance it runs well past $100, which the bonds can never do.
Valuation
It closed Friday at $54.24 with 551.5 million shares: equity of $29.9 billion, enterprise value of about $60 billion (pro forma cash is an estimate; the sources page shows the construction), against a $10.3 billion revenue run rate and roughly $6 billion of run-rate adjusted EBITDA. Three scenarios for 2027, drivers stated:
| 2027 revenue | adj EBITDA (margin) | multiple | per share | |
|---|---|---|---|---|
| Bear: deliveries slip, signings stall, spreads stay wide | $16B | $8.0B (50%) | 6x | ~$32 |
| Base: the backlog converts on schedule | $21B | $11.6B (55%) | 7.5x | ~$102 |
| Bull: signings compound, cost of debt keeps falling | $26B | $15.1B (58%) | 9x | ~$191 |
The base case requires no heroics or imagination: $21 billion is approximately what the existing backlog’s own recognition schedule produces before any new contracts are signed. The multiples are made by me, argued simply: 6x is where capital-intensive infrastructure trades when growth is doubted (and the low end of many analysts’ estimates), 9x is a premium the market has paid for contracted data center growth, and 7.5x is splitting the difference.
Those per-share values subtract today’s net debt, but the capacity that earns 2027’s revenue gets built with tomorrow’s. In other words, the table charges the company for the debt it carries now, while the data centers that produce those 2027 numbers still have to be built with money that has not been borrowed yet, and every new dollar of that debt comes out of what the stock is worth. Every additional $10 billion of net borrowing between now and the end of next year takes roughly $18 off every number in that table, and the first half of 2026 already ran a $10.5 billion funding gap. The prepayments and growing operating cash flow shrink the gap from here, but they do not close it. That drag is the reason my price target sits closer to the 6x multiple and not the 7.5x one.
What could go wrong
- The credit window shuts. Ten billion per half of new money, plus a $10.6 billion maturity wall in 2027 and 2028. BofA is documenting lender fatigue in real time. This is risk one, two, and three.
- A big customer wobbles. The top three customers are 72% of revenue, and OpenAI’s commitments across the industry, by widely reported totals, exceed its revenue many times over; its checks clear only as long as its own fundraising does.
- Delivery slippage. Take-or-pay cuts both ways: late capacity does not bill, and the December 2025 covenant amendment is the precedent for what slippage costs, with its unlimited equity-cure window closing October 28.
- Compute pricing breaks. Every collateral value and every backlog dollar reprices off GPU rental rates. Today they are rising. Nothing in this piece survives them falling 40%.
- Depreciation turns out to be true. If five-year-old chips earn little, the EBITDA margin is partly fiction and the bear scenario’s 6x is generous.
- Litigation. A securities class action (D.N.J., filed January 2026) and related derivative suits are pending.
What would change my mind
| bullish trigger | bearish trigger |
|---|---|
| Q3 (November) shows backlog above $130B with Meta ramping on schedule | any new covenant amendment or delivery-delay disclosure |
| the next big debt raise prices below 8% | new debt costing more than 10%, or an offering pulled |
| Nvidia extends or enlarges its obligation to buy unsold capacity | OpenAI missing a funding round that covers its compute bills |
The first trigger arrives in November, at Q3 earnings.
Price Target
I am a buyer at the current price and my price target is $70. This is a developing company, however, and there are better plays in the AI market where I would be more comfortable parking my money. If access to the bonds were easier and more possible, that is where I would be in on CoreWeave in a real, serious way.
Disclosure: I do not hold a position in $CRWV. I hold positions in some of the other companies mentioned in this piece. This piece is for informational and educational purposes only and is not investment advice. Prices as of the October 23, 2026 close unless noted. Bank of America research is cited via press coverage of its notes; Jensen Huang’s GTC remarks are cited via event coverage and transcripts.
View all sources · Suggested citation: Rettke, Sterling. “$CRWV & The $5 Trillion IOU.” sterlingrettke.com, October 2026.
The content on this site is for informational and educational purposes only and does not constitute investment advice, financial advice, or a recommendation to buy or sell any security. Sterling Rettke is not a registered investment adviser. The author may hold positions in securities discussed. Always do your own research and consult a qualified financial advisor before making investment decisions.