CoreWeave

IThe business

CoreWeave sells computing time. It buys the chips that train and run artificial intelligence, mostly NVIDIA’s, installs them in data centers that it mostly rents, wires them together with its own software, and rents the result by the hour or, far more often, by multi-year contract. It calls itself the Essential Cloud for AI. A fair description is an equipment lessor with a software layer on top.²³

The company was founded in September 2017 and launched its cloud platform in 2020. Before 2022 it had little revenue, mostly from mining crypto currency, which it stopped. It listed on Nasdaq on 28 March 2025 at $40 a share. Revenue went from $229M in 2023 to $1.9B in 2024 and $5.1B in 2025. The twelve months to June 2026 brought in $7.6B C (calculated). That is a short record: three audited years, and every one of them ended in a loss at the bottom line.²³¹

Who pays

The business is a few big customers. Microsoft was 67% of 2025 revenue, 62% in 2024 and 35% in 2023.² In the second quarter of 2026 the three largest customers were 36%, 26% and 10% of revenue, 72% together, and the quarterly report does not name them.³ The annual report names OpenAI and Meta as significant customers from 2026: OpenAI committed up to about $11.9B through October 2030 in a March 2025 contract and up to about $6.5B more through May 2031 in a September 2025 order. Meta committed up to about $14.2B through 2031 in September 2025, and in March 2026 signed another order for up to about $21.0B, including an option it exercised.²³ About 93% of the second quarter’s revenue increase came from customers it already had.³

Customers buy capacity on take-or-pay contracts with a weighted average length of about five years. The company says these are underwritten before signing, with prepayments, credit checks and sometimes letters of credit. It carries no credit insurance on any of them.² Customers have also paid in advance: deferred revenue was $9.7B at 30 June 2026 R (reported).³

The backlog

The company calls its contracted revenue the backlog. The accounting version, remaining performance obligations, was $103.7B at 30 June 2026, up from $60.7B at the end of 2025 and $15.1B a year earlier. Of the $103.7B, 41% is due in the first 24 months, 39% in months 25 to 48, and 20% in months 49 to 78 R (reported). More than $25B of new commitments were added in the first weeks of the third quarter and are not in that figure.²³⁵

Two cautions on reading it. First, the rate of turning it into revenue: the first 24 months hold about $42.5B, an average of $21B a year, against a fourth-quarter 2026 run-rate the company puts at $18.5B to $19.5B a year.³⁷ The backlog, on its own, supports a business of about the size it will already be at the end of this year. Growth beyond that is new contracts. Second, quality. More than half of the backlog sits in contracts where delivery has started; management says it expects more than two thirds by the end of 2026. The rest is a promise to build.⁷

How big it has become, in power

The unit the company counts in is active power, the megawatts of electricity that are energised and available to run chips. It had about 70 MW at the end of 2023, about 360 MW at the end of 2024, over 850 MW at the end of 2025 and 1.5 GW at 30 June 2026, nearly 500 MW of it added in one quarter. Management guides to more than 1.85 GW by year-end and says it has excellent visibility to at least 8 GW by 2030. It had contracted about 3.1 GW at the end of 2025, 3.7 GW at 30 June 2026 and about 4.2 GW by 11 August.²⁵⁷⁹

The box, in plain words

The way the company pays for all this is the thing to understand first. CoreWeave signs a contract with a customer. It then sets up a separate company, owned by CoreWeave, that holds that contract, the chips to serve it and the right to the data center. Lenders lend to that company against the contract’s cash, in a delayed draw term loan that is drawn as the chips arrive and repaid month by month from what the customer pays. CoreWeave’s chief executive, Michael Intrator, calls such a structure a box. The idea is that the customer’s payments go first to the data center, power, interest and principal, and only the remainder reaches CoreWeave.³⁷¹⁵

The company says the customer’s payments over a typical five-year contract fully repay the debt in that box, with a good deal left over, and that the chips are then free to be rented again.⁷ I test that in the valuation section, because it is the hinge of the whole case. For now: at 30 June 2026 the company had $13.6B outstanding under such loans, $5.1B of financing from equipment makers and a software vendor, $10.0B of unsecured senior notes, $6.6B of convertible notes, and a revolver with $2.0B of room.³

How it got here

• March 2025: listed at $40. The stock closed at a low of $35.42 on 21 April and a high of $183.58 on 20 June 2025.¹
• May 2025: bought Weights & Biases, a developer tools business, and later OpenPipe, Marimo and Monolith AI, paying mostly in stock.²
• January 2026: sold about 23 million shares to NVIDIA for $2.0B at $87.20. NVIDIA held 10.66% of the Class A shares at 15 April 2026.³⁸
• March 2026: a new order from Meta of up to about $21.0B.³
• April 2026: $1.0B of stock sold at $109.00 in a private placement, which I take to be the $1B Jane Street investment the company’s release describes, though the 10-Q does not name the buyer; and $4.0B of convertible notes due 2032 at 1.75%.³⁵
• May and June 2026: the first publicly syndicated delayed draw term loan on chips, $3.1B at SOFR plus 4.50%, and a first euro bond, €2.0B at 8.5%.³⁵
• September 2026: $4.2B of convertible notes due 2033, with a conversion price of about $97.85, and an equity programme for up to 35 million more shares.¹⁰⁹

The fund Situational Awareness LP, run by Leopold Aschenbrenner, held 7,479,558 shares of CoreWeave Class A stock at 30 June 2026, valued in its filing at $744.5M, which was 3.68% of the $20.2B it reported, according to its 13F filing made on 14 August 2026. That is a dated fact from a regulatory filing. A 13F does not show short positions or holdings outside the United States, and it says nothing about what the fund has done since or will do, or why it held the shares.¹¹

IIThe moat

The question for this company is what stops a rival or a customer from doing the same thing. It has two things that might protect it and one that threatens to take the protection away.

What might protect it

• Time. It takes about two years to find power, build or rent a hall, order and install racks. CoreWeave has 4.2 GW of power under contract and more than 1.5 GW of further options, which the company counts separately. A buyer who wants capacity in 2027 has fewer places to go than one who is planning for 2029 E (estimate).⁷
• Software. The company sells a Kubernetes service, a storage layer that caches data on the chips, a fleet manager called Mission Control, and, since 2025, Weights & Biases, which tracks model training runs. It says the non-GPU lines passed $400M of annual recurring revenue in the second quarter and that managed inference rose from $1M to more than $100M of booked annual recurring revenue in a few months.²⁷
• Operating record. In 2025 SemiAnalysis rated it the only Platinum provider in its ClusterMAX ranking, according to the 10-K. That is a third party rating, quoted by the company.²

What threatens it

Everything in the box can be bought by someone else. The chips are NVIDIA’s, and the 10-K says a limited number of suppliers provide the key components. The racks come from the same server makers. The data centers are leased from developers that also lease to others. A customer with its own money can do the same thing: Microsoft, Meta and OpenAI are customers and also builders of their own capacity. The 10-K names Amazon, Google, Microsoft and Oracle as competitors with more resources and longer histories.²

What stops customers leaving is the contract. What keeps the contract is the shortage. The company says near-term capacity is effectively sold out, that demand comes from several buyers for each chip, and that prices for current and older chips are at or above levels of a year ago; in July it raised list prices by about 25% across its chip types.⁷ That is a statement about a scarce market. Protection that depends on a shortage is worth less when the shortage ends. This one has not been tested against that.

IIIThe competitive advantage

Where does CoreWeave beat the alternatives, and where does it lose? It beats them on speed, on focus and on willingness to take the risk of a single AI-driven buyer. It loses on the price of money.

Speed

In the second quarter it added nearly 500 MW of active power, and in June alone more than 300 MW, more than any single earlier quarter. Construction in progress rose to $11.9B from $9.6B in a quarter, which management reads as capacity about to arrive in July.⁷³ Few companies can install equipment at that rate. The cost is visible as well: the company says it continues to incur ramp costs, rent and power before a cluster earns.⁷

The cost of money

The company has pushed its weighted average cost of debt down by almost 300 basis points in a year, which it says saves about $1.1B a year at the end-of-June debt load.⁷ The notes themselves say what lenders ask. At 30 June the company paid effective rates of 7% to 15% on its term loans, 9% to 10% on unsecured notes, 11% on equipment financing and 15% on its oldest term loan; the convertible notes carry about 2% because the holder also gets a call on the stock.³ Weighted by amount, the $35.6B of debt averages about 8.4% including the converts and about 9.9% without them C (calculated). Press reports in mid-October described credit fears and put yields on some of its unsecured bonds near 15%; I did not find a primary quote for that figure.¹⁷ The Treasury yield on ten years is 4.98%.¹³ The company says it aims to move its credit profile toward investment grade, which tells you where it is not yet.⁹ A spread of that size over risk-free money matters in a business whose product is a financed asset.

Utilisation, price and the engineering

The company says its clusters score above reference targets on model flop utilisation in NVIDIA’s own benchmarks and that its inference stack was fastest on open models on a public leaderboard.²⁷ These are company statements about tests it ran or chose. They matter less than the fact the customers renewed: 93% of the second quarter’s revenue growth came from existing customers.³

Price is the strongest sign. The prospectus supplement says contracts signed since June are short, three to six months, at about $40M of annualised revenue per megawatt of power the clusters need.⁹ That figure is about four times the $10M per megawatt the company’s year-end revenue guide implies for the whole fleet C (calculated).⁷ I read it as a price for a scarce, new, short contract, not as a number that lasts for five years, and I do not use it in my cases. It tells us how tight the market was in July and August. It cannot tell us how tight it will be when the contracts end.

IVManagement

How they are really paid

The pay is plain. In 2025 Intrator and the co-founder Brian Venturo each earned a $750,000 salary and a $750,000 bonus, set at 100% of target for every executive. Brannin McBee, the third co-founder, earned $600,000 and $600,000, and the chief financial officer, Nitin Agrawal, $500,000 and $500,000. Intrator’s reported total was $3.45M, including $1.89M for filing fees that regulators require for his shares, of which $1.04M was a tax gross-up the company paid.⁸

The real pay is stock. In March 2025 each received restricted stock units with a four-year time-based vest and a condition that was met at the IPO: $28.5M of target value for Intrator, $16.0M for Venturo, $10.8M for McBee and $10.5M for Agrawal. None of it depends on return on capital, free cash flow or the cost of debt. A new chief revenue officer, Jonathan Jones, joined from Amazon Web Services in October 2025 with $22M of target stock.⁸ Stock pay in the cash flow statement was $630M in 2025 and $318M in the first half of 2026 R (reported).²³

What the pay plan rewards

It rewards staying and growing, and the share price. A plan with no capital charge pays an executive the same for a dollar of revenue that earns 5% as for one that earns 25%. In a company whose capital costs 9% to 10%, that matters. It is also the common arrangement among founders of capital-heavy companies. I do not read it as a flaw in these particular executives; I note that nothing in it pushes them to say no to a contract that fails to earn its keep.

How much of their own money is in, and what they have taken out

They are large owners with unusual voting power. At 15 April 2026 Intrator held 38.70% of the votes, Venturo 20.30% and McBee 14.59%, through Class B shares that carry ten votes each; officers and directors together held 72.32% of the votes. Magnetar, the hedge fund that lent to CoreWeave early, held 17.07% of the Class A shares, NVIDIA 10.66% and Jane Street 6.53%.⁸

They have also sold, a lot of it. Adding up the sales in each of their Form 4 filings from the IPO to 30 October 2026, Venturo sold about 11.8M shares for roughly $1.2B, Intrator about 10.1M for roughly $1.0B and McBee about 9.8M for roughly $0.95B, about $3.1B between them C (calculated). Several lines repeat on the same day, which can be separate accounts or duplicates, so treat the totals as good to a few percent. Intrator and McBee sold another $0.25B after the second-quarter results on 11 August. Magnetar’s affiliated funds also sold, a larger amount; several of its entities file for the same trades and I did not try to total them.¹² Selling is ordinary after an IPO. It means the founders’ wealth is no longer tied only to the price, and it shows what they think of $100 prices; it says little about $55.

What they did with capital over a full cycle

There is no full cycle yet. The company has existed through one build phase. Its weighted average share count went from 192M in 2023 to 218M in 2024, 435M in 2025 and 551M in the second quarter of 2026, a rise of 187% in three years C (calculated), from IPO shares, acquisitions paid in stock, the NVIDIA and other private placements, and exercised options.²³ Class A shares outstanding were 458.9M on 31 July 2026, with another 94M Class B; the prospectus supplement lists options for 23.6M shares, units for 29.3M, warrants for 11.9M, and room for 35M more shares to be sold through banks.⁹³

Of the dollars it spent, nearly all went to equipment. Cash capital spending was $2.9B in 2023, $8.7B in 2024, $10.3B in 2025 and $14.1B in the first half of 2026. Equipment and buildings on the books, before depreciation, went from $12.9B at the end of 2024 to $52.6B at 30 June 2026 R (reported). It spent small amounts on acquisitions and began to invest in joint ventures: $1.7B committed to two data center developments, with $479M carried at 30 June.²³

Controls and the lawsuit

The 10-Q states that disclosure controls were not effective at 30 June 2026, because of material weaknesses it first reported at the IPO: general controls over finance systems, separation of duties, and enough qualified finance staff. It says it has not found a misstatement and is fixing them.³ A putative securities class action filed on 12 January 2026 in the District of New Jersey alleges false and misleading statements, with derivative suits stayed behind it. The company has accrued nothing.³

VFinancials

CoreWeave has three audited years, 2023 to 2025, and the first half of 2026 from unaudited reports. I cannot build a ten-year history, and I cannot say how the business behaves in a downturn, because it has never had one. The table below shows what there is.

Read the table in two places. The growth is real: revenue is up 33 times in three years. The bottom line says that growth has not yet paid for itself. Depreciation took $0.86B in 2024 and $3.99B over the latest twelve months. Interest went from $0.36B to $1.87B. In 2024 operating income was $324M, and the company was close to break-even before interest; in the next two years depreciation and rent grew faster than revenue.²³

Cash, and why free cash flow tells two stories

Operating cash flow looks healthy and rising: $6.9B over the latest twelve months. Two things lift it. Customers prepay, and the increase in deferred revenue added $4.8B to operating cash in that period. And much of the equipment is not paid in cash at the time of purchase. Of the $34.0B of equipment and buildings added in twelve months, only $20.6B left as cash capital spending; the other $13.5B was not paid in cash in the period. Most of it is owed to suppliers, financed by equipment makers or sitting in the accrued purchases line, which was $5.5B at 30 June, with a little capitalised interest and stock pay C (calculated).³² The usual free cash flow figure, operating cash less cash capital spending, is −$13.7B for the twelve months, and it understates the outflow by $13.5B.

What an owner can really take out

An owner wants what is left after the business has paid for its customers’ computing, its debt, and replacing what wears out. I build it in steps. Start with operating cash, which already includes cash interest. Take away the customers’ prepayments, which are borrowing from customers. Take away stock pay, which operating cash flow adds back though it costs owners. Take away depreciation as a rough charge for replacing worn equipment. What remains is what would be left if the company stopped growing.

That is the first honest number on this page. Without growth and without customer prepayments, the business consumed $2.5B over the last twelve months. The business is not yet earning what it spends to stay the same size. This does not mean it will not. The latest quarter had an adjusted operating margin of 5%, up from 1% in the first, and the company guides to the low teens in the fourth quarter. But it shows how far the first margins are from an owner’s cash.⁶⁵⁷

Debt funded about two thirds of the cash. The rest was operating cash, of which $4.8B was customers’ prepayments, and a little stock. The company says it raised about $18B across debt, convertible notes and stock in the second quarter alone, and has over $32B of debt and equity secured to date.⁷

The debt

Principal due, from the debt note: $4.4B in the rest of 2026, $6.2B in 2027, $4.4B in 2028, $2.4B in 2029, $3.2B in 2030 and $14.9B after that. Current debt on the balance sheet was $7.5B. Cash and equivalents were $5.5B, with $1.4B restricted and $10.0B available to draw under existing facilities, for $15.6B of liquidity R (reported).³ Since 30 June it has added a $2.6B term loan (with $1.2B drawn), a $1.2B draw on the revolver and $4.2B of convertible notes, and has filed to sell up to 35 million shares.³¹⁰⁹ Interest was $640M in the second quarter and the company guides to $860M to $940M for the third.⁷

The shares are a thin layer on a large structure. At $54.87 the equity is worth about $33.8B, and debt plus customer prepayments less cash is $39.9B. Value the whole at that and you have an enterprise of about $73.7B, which is 9.7 times the last twelve months of revenue and 12 times the second quarter’s adjusted EBITDA at an annual rate C (calculated).³⁵⁹

VIValuation

Start with what would ruin it

Before any number, I ask what would most likely hurt the owners of this business, and work backward from there. I see three ways, and they are linked.

• The chips stop earning before the loans and the stock are paid for. Every box is sized so a contract repays its lender in about five years. What is left for the shareholder is what the chips earn after that, and nobody has run these chips for nine years at this scale. If prices fall when the shortage ends, or the chips become hard to rent after six years, the stock is worth little.⁷²
• A large customer fails or renegotiates. Three customers were 72% of revenue in the second quarter, and the only one the filings name for 2025 was Microsoft at 67%. OpenAI, a private company, has committed in two contracts to pay up to about $11.9B and $6.5B, and the 10-K itself says a customer that fails could leave CoreWeave holding the chips, the leases and the loans.²³
• The money stops. The plan needs the lenders, again and again. In my base case the company adds about $52B of net debt and equity before it turns cash positive in 2030. Principal due runs at $4.4B in the rest of 2026, $6.2B in 2027 and $4.4B in 2028 before the new borrowing. Mid-October showed what a weak week does: the shares fell from $69.69 on 15 October to $54.24 six trading days later, down 22%, and the press reported credit fears and unsecured bond yields near 15%.¹³¹⁷

The model below tests each one. The first is the chip life and the price per megawatt. The second shows up in the bear case. The third is the funding walk.

What a megawatt really earns

Take one megawatt of active power. From the filings, the equipment and buildings added since the start of 2024 total $68B for the 1.78 GW added to date, with $11.9B still in construction, which puts the all-in cost between $32M and $38M a megawatt; I use $35M, of which $30M is the chips, racks and network and $5M is the fit-out of the hall E (estimate).²³⁷ Fleet revenue is about $10M a megawatt a year now and the company’s year-end guide implies $10.3M; I use $11.5M for a mature cluster after the July price rises E (estimate). With an adjusted EBITDA margin of 64%, a level the company is heading toward, that is $7.4M a year.⁷⁵

Now finance it the way the company does. A lender advances 80% of the $30M, $24M, at 9.5% over five years. The annual debt payment is $6.25M. The box covers its debt payment 1.18 times in a full year, a thin margin that fits the company’s statement that a five-year contract repays the loan.⁷ The shareholder puts in $11M and receives the $1.1M of surplus for five years, and then everything. After the fifth year the chips are free to rent again, and how long they earn decides the answer.

Read it this way. At five years the equity is almost lost. At six, the accounting life, it earns about 2%. At eight years it earns about 15% and at ten about 20% a year. The investment is a bet on the length of the chips’ life, much more than on demand this year. The unlevered return on the whole $35M is 12.2% at eight years, 14% at nine and 6% at six, so at eight years the project earns just about my 12% discount rate.

I need evidence on the life. The company says it signed a contract for the A100, a chip launched in 2020, that runs to 2029 at an attractive price, and that the older fleets that are coming off contract are a very limited part of what it has.⁷ That is one contract, which was priced during a shortage. The accounting life is six years, and it was five until 2023.²

What the price implies

I find this useful. The price is not asking for a miracle. It asks the base case to be a little better in one respect: the chips earn nine years instead of eight, or prices run 9% higher, or the discount rate is two points lower. Equity in a business like this is the thin slice that is left over, so a small change in an input moves it a long way. That is why I give a wide range and do not trust a point.

The discount rate

I use 12.0%, the Treasury yield of 4.98% plus a premium for a business I cannot see far into.¹³ The build: the company’s equity costs more than 16% by any measure of its price swings, which include a 22% drop in six trading days this month. Its unsecured debt costs about 10% (the 9.75% notes due 2031) and its box loans cost 7% to 15%.³ Roughly half of the financing is the cheaper secured layer, which pulls the blend down to about 12% E (estimate). Higher is easy to defend and lower is not. I show 10% to 15% below.

The three cases

The bear is the story told by short sellers: demand cools, only 3.7 GW of the contracted power is lit by 2030, prices per megawatt slide to $8.6M, and the chips earn six years. In that case the fleet’s cash flows do not cover the debt and the equity is worth zero. The model gives it a negative equity value, about −$176 a share if the lenders were somehow paid in full; I show $0 because a share cannot be worth less than nothing, and the debt takes the loss.

The base has power arriving a little late, 5.6 GW at the end of 2030 against the company’s target of at least 8 GW, revenue per megawatt near the July 2026 level, margins rising to 64% as the new clusters mature, and chips that earn eight years. Revenue is $63B in 2030, about 8 times the TTM. It is worth $9 a share E (estimate).

The bull has the 8 GW target met a year or so late, the July price increases holding, 65% margins and chips earning 9.5 years. Revenue is $83B in 2030 and $111B in 2035. It is worth $157 a share after treating the three convertible issues as converted, which adds about 100M shares E (estimate).³¹⁰

From cash flows to a value per share

Look at the second line. In the base, the cash flows to 2035 are negative, −$6B in present value, because the company spends more building than it earns until 2030. All of the value of the operations, $51B, comes from after 2035, when the fleet is mature and cash comes out. After the claims are paid, $5B is left for the owners. The steady-state year in my model has revenue of $85B, EBITDA of $55B and replacement capital of $34B, so about a quarter of EBITDA is free cash.

The funding walk

In the base, the company’s net claims rise from $39.9B to about $119B by the end of 2029, a cumulative shortfall of $52B in owner cash plus interest on the way. They fall as a multiple of EBITDA from 5.3 times at the end of 2027 to 3.5 times in 2029 and 3.0 times in 2030. The company says it has $10.0B undrawn, and it raised about $18B in the second quarter. The walk is feasible if the lenders keep lending against contracts, which is what the boxes are for. What it cannot survive is a closed market in 2027 or 2028, when the shortfall peaks.³⁷

Equity dilutes this. Suppose 30% of the shortfall is raised in stock at $54.87. In the bull, that cuts the value per share from $157 to $123, because the shares are sold below what they are worth in that case. In the base the same sale raises the value per share from $9 to $23, because I think the shares are worth less than $54.87 and the buyers pay more than they are worth. Both effects are real, and the company has already sold stock at $87.20, $109.00 and in the Jane Street deal.³ The bull also does not count the capped calls bought with the convertible notes, which would soften the dilution up to a cap of $230 on the 2032 notes.³

How the odds shrink

Four coin-flip bets multiply to about 7%. They are linked, because strong demand lifts power, price and the life of the chips together, so I weight the bull at 20%, the base at 45% and the bear at 35%. That gives $35 E (estimate). A reader who thinks the bull is 35% likely and the bear 20% gets about $59, around the price.

A stress with no probability

The bear keeps all customers paying. The stress takes away one. If the largest customer, 36% of second-quarter revenue, did not renew or did not pay, the box that serves it would still be repaid first from that customer’s contract if it held, but any revenue from the freed chips would have to be sold again in a market that I assume is weaker at that moment. The unsecured notes, $10.0B, and the converts, $6.6B, sit behind the secured boxes and would take the first loss. The equity is gone well before. I do not put a number on it, because the loss depends on contract terms I cannot see, and I cannot see whether a failure in one box leaves the others untouched. Of the $13.6B of term loans, $10.8B is listed as recourse debt and DDTL 5.0 is unconditionally guaranteed by the parent. Only DDTL 4.0, $2.8B, is non-recourse except for limited guarantees.³

My margin-of-safety price

My margin-of-safety price is $4, half of my base value E (estimate). I cannot set it from the bear, because the bear is worth zero at any price. Half the base is a judgment. I want a wide cushion because the base itself moves from about $0 to about $59 with two points of discount rate. At $4 the whole business, debt included, earns 12.3% a year on my base cash flows and 18.8% on the bull’s, against the 12.0% I require. That rate moves little because the shares are a thin layer on about $40B of claims; the bear still loses everything. The price sits 1272% above it. The lowest close of the past year was $54.24, a week before the price date.¹

The next best use of the same dollar

A ten-year Treasury pays 4.98% and cannot lose.¹³ CoreWeave’s own 9.75% senior notes due 2031 rank ahead of the stock and pay 9.75% a year on their face; the press reported yields much higher in mid-October, which I did not verify.³¹⁷ These rates are not like for like. The Treasury and the notes pay a lender a stated rate. The 10.1% is the rate my base cash flows give for the whole business, debt included, at today’s price; a shareholder gets a thinner, more variable slice and nothing at all in my bear. On this site, Oracle is the nearest comparison: it also borrows to rent out AI capacity, and my base there returns 10.3% a year on the same kind of measure.

Quality and price, separately

Quality first. The business is a lessor of fast-ageing equipment funded mostly with borrowed money, with three customers making up 72% of revenue and a record of three years. Growth is outstanding and the contract backlog is large. The returns on capital are not yet proven to be above its cost of capital: the second quarter’s adjusted operating margin was 5%, interest alone was 25% of revenue, and my one-megawatt test makes the equity return depend on a chip life nobody has measured. I would call it a high-growth business of unproven quality. Price is a separate matter. At $54.87 the shares are 518% above my base value and 65% below my bull; the market is paying for some of the bull and little of the bear. That the share price has fallen 60% from its high does not by itself make it cheap.

Why the market may be right

The shares are an option on the working life and price of AI chips, and options are worth more than their expected payoff when the outcomes are wide. My weights put the expected value at $35, but with a bull at $157 and a 20% weight, a reader with a bull at 30% gets near the price. The market may also be right that demand lasts: the company says near-term capacity is sold out and prices for older chips are firm. My model is also conservative on one thing, the terminal value, which is 3 times EBITDA in 2035. If the fleet were sold in 2030 for 4.3 times that year’s EBITDA, not the 3.5 times my cash flows imply, the price is worth it. Chips and halls change hands at higher multiples than that in good times.

What I cannot judge

• How long a GPU earns. The accounting life is six years and one contract runs nine. I do not know which is typical, and the answer moves my base from $0 to $75.
• What prices do when the shortage ends. The company says prices have risen 25% since July and that pricing for older chips holds. Those are claims about a scarce market.⁷
• Whether the boxes are as isolated as described. The debt note lists $10.8B of the $13.6B of term loans as recourse, and I cannot tell what a default would trigger.³
• Whether OpenAI, a private company, can pay what it signed. I cannot see its finances.²
• The third-quarter results, due after the price date. Guidance is $3.45B to $3.6B of revenue and $200M to $260M of adjusted operating income.⁷

Where this could be in 5 and 10 years

No owner cash reaches shareholders in any case before 2036. Every dollar the fleet earns repays debt, and in the bull the net claims reach about zero only around 2036. So the value in a later year is today’s value grown at my 12.0% rate, with the debt taking its share on the way. The base reaches $16 in October 2031 and $28 in October 2036 E (estimate), which is −6.7% a year from $54.87.

For the base to hold at five years, power must reach about 6 GW by the end of 2031 and revenue $69B, while the net claims, $119B at the peak in 2029, start to fall. At ten years it must also get through the first big replacement of chips bought in 2027 to 2029, at about the same cost per megawatt and the same price. The bear and the bull differ in three numbers: how many years the chips earn (6 against 9.5), revenue per megawatt ($8.6M against $13.3M) and how much power gets built (4.2 GW against 9.0 GW). Each is a judgment about a market that is five years old.

The most likely way my base is wrong is the chip life. If the chips earn nine years, not eight, the base is $45 today, close to the price. If they earn seven, it is zero. I have no data to choose between those.

These are estimates of value, not forecasts of the share price, and the market price can sit above or below value for years. I also assume new capital raised at fair value, which the company has not always done.

VIIThe case for the price

What most people miss

Three facts cut its way. It just signed an A100 contract, a 2020 chip, running into 2029.⁷ It raised prices about 25% in July and says short contracts now sign near $40M of annualised revenue per megawatt.⁷⁹ And software outside chip rental passed $400M of annual recurring revenue.⁷

The bull case

The bull case is management’s, on the Q2 call of 11 August 2026. Intrator said near-term capacity is effectively sold out, with visibility to at least 8 GW by 2030. Agrawal said each five-year deal repays its debt within the term, with cash left over.⁷

Intrator explained the structure behind that, the box, to Jason Calacanis on the All-In podcast, in an interview recorded at NVIDIA’s GTC in March 2026.¹⁶ I did not watch it. I read a machine transcript of the clip replayed in the October episode below (4:13 to 5:44). In my words: the customer’s contract, the chips and the data center lease go into one box; the customer pays the box; it pays the data center, power and lenders, and the rest goes to CoreWeave; each box is separate; and it pays for itself in about two and a half years of a five-year deal.¹⁵ The 10-Q backs only part: DDTL 4.0, $2.8B, is non-recourse, but the parent guarantees the other five term loans, $10.8B of $13.6B, so those boxes are not sealed.³

A reply from the other side

The episode, “CoreWeave Crash, AI Trade Cracks, Anthropic IPO Red Flags, Bill Gates Doomer”, went up on YouTube on 16 October 2026 Pacific time.¹⁴ YouTube’s chapters put the CoreWeave segment at 2:18, which the transcript’s clock matches. On it, Chamath Palihapitiya answers from about 6:00, with his $5B breakeven estimate in the block at 11:30, before the host turns to David Sacks at about 13:14. The transcript has no speaker labels; the host asks him. From the block at 6:07: “a coin flip whether they default or not”.¹⁵

His case, in my words: credit default swaps jumped and imply roughly 55% default over five years; higher rates lift borrowing costs; about $35B is still to be raised; and fixed prices give no gain from higher spot prices, with three customers at about 75% of revenue.¹⁵

Against the filings: three customers were 72% of second-quarter revenue.³ Assets of $77.1B cover $35.6B of debt at book value.³ Fixed prices apply to the signed backlog, not new business.⁷⁹ His $35B is the order of my $52B shortfall less the $32B secured.⁷ I cannot verify his default odds; swaps refer to the parent’s unsecured debt, not the boxes, and overstate real odds when investors want extra pay for risk.

Against reality

If CoreWeave reaches 5.6 GW by 2030 at today’s prices and a 64% EBITDA margin, and its chips earn eight years, the shares are worth about $9. At 8 GW, with the July increases kept and nine and a half years, about $157. Revenue is $7.6B over twelve months against $63B in 2030 in my base: about $143B of equipment and $52B of net borrowing.³⁷

What a 10x would take

Ten times the price is $548.70, or about $390B of equity on roughly 717M shares including the converts C (calculated). My bull has $40B of owner cash in 2033; ten times that is about $400B, so a 10x needs all of the bull case and a multiple of ten. The stock rose 4.6 times to $183.58 in June 2025 and sits 70% below it.¹ I put a 10x well below my 20% for the bull.

What they have to hit

Where I come down

The reply moved me on timing and concentration more than on the shares, because the plan needs lenders for two more years. The unit economics did not move me; the shareholder keeps what is left, a bet on how long the chips earn. Proof that old chips rent near full price for nine years would raise my values.

VIIIRisks, and what would change my view

The valuation section covers the three main ways this goes badly. These are the others, and the signs I will read.

• Delays and power. 393 MW of electrical power at one leased site and 355 MW at others were undelivered at 30 June, with delivery between 2026 and 2028. Rent under one of those leases is based on construction costs up to $14.7B.³
• Supply. The 10-K says a limited number of suppliers provide key components. Memory is a pinch point: the company made a long-term deal with Solidigm for storage this year and says it passes through component price rises.²⁷
• Customers. OpenAI is private and its two contracts total up to about $18.4B. The largest customer was 36% of revenue in the second quarter against Microsoft’s 67% for 2025, and the quarterly filings do not name the top three.²³
• Dilution. The share count rose 187% in three years and the company has registered for 35 million more shares, plus $4.2B of 2033 converts at $97.85.⁹¹⁰
• Controls and law. Material weaknesses in financial controls remain open, and a securities class action filed in January 2026 is pending.³
• Control. Officers and directors together held 72.32% of the votes at 15 April 2026; outside shareholders have little say if they disagree.⁸

Geopolitical risk, and what it means for the value

CoreWeave sells to American customers from data centers it leases in the United States, Europe and the United Kingdom; 94% of 2025 revenue was booked in the United States.² Its geopolitics sits upstream. Every chip is NVIDIA’s, NVIDIA relies on TSMC in Taiwan, and the filings name the growing potential for military conflict between China and Taiwan as a supply risk.²³ The rest runs through Washington and the local power grid.

The table reruns my model for each case. I changed one thing at a time. Odds are my judgment, not a calculation. Taiwan cuts both ways: new capacity stops, but chips already installed become scarcer, so I show both. Delays move the base little, because in my base the new capacity earns about its cost of capital. The model charges nothing for halls left empty.

Export controls matter less to CoreWeave than to NVIDIA, because its customers are American. The January 2026 rule lets H200-class chips go to China case by case, capped at half of US sales of each chip, with a certification that the exports do not cut US supply.¹⁹ It is a statement of policy that can change. The 25% Section 232 duty from 15 January 2026 covers a narrow band of chips and exempts imports for US data centers; Commerce owed an update on that market by 1 July.²⁰ The customers are a different risk. OpenAI’s two contracts are up to about $18.4B, against a backlog of $103.7B, and a 4% lower price per megawatt across the fleet already takes the base to $0.²

What is already in the price: the shares are 60% below their May high on credit fear, so some of that risk is there. I see no sign of Taiwan or tariffs. The bear stays at $0 in every case. The base falls to $7 or less in four of five. The bull stays between $97 and $135, above $54.87, so these risks mostly cut the bull, which is the case the price is paying for.

What would raise my base: chips that earn nine years, prices that hold above $11.5M per megawatt, and cheaper debt. What would lower it: a loss or restructuring of a large contract, a closed credit market in 2027 or 2028, or a first renewal at a discount. A lower share price would not change my base value. It would change the answer to whether the bear is paid for.