Microsoft and Meta rent only to cover gaps in their own builds. Three customers pay 72% of sales.
Is it a good business? · 6 · What does its record show?
Sales soared; profit turned to a loss
Sales, $B
0.2
1.9
5.1
Dec 2023Dec 2024Dec 2025
Sales$0.2B → $5.1B2023 to 2025
Profit per $116.9% → -1.9%operating profit per $1, 2024 to this year
Interest$846Ma quarter, growing with the debt
Each new site brings rent, power and wear on the chips before it brings sales. Management says profit turns once sites fill.
Is it a good business? · 7 · How much room is there to grow?
The specialist AI cloud tripled in a year
Over $25Bthe specialist AI cloud market in 2025
$5.1BCoreWeave's 2025 sales, about a fifth of the market
30% a yearexpected growth in AI chips' electricity use
Nobody publishes how many chips are rented out in total.
Is it a good business? · 8 · Who is in charge?
Founders in control, paid at the board's discretion
About 72%of the votes held by insiders
100%of target bonus, paid at the board's discretion with no metric
784insider sales in the past year, and no purchases
Chief executive Michael Intrator and the other founders still run it.
The story right now · 9 · Why do AI labs rent chips?
AI labs need huge computers faster than they can build them
72 GPUsin one NVIDIA rack; thousands of racks run one job
Over $100Bof rent signed with CoreWeave
Every new AI model takes more computing than the last, often weeks of nonstop work. Labs rent clusters rather than wait years to build their own.
The story right now · 10 · Why does signed rent not mean profit?
The chips are bought now; the rent comes over years
A lab signs a contractPays monthly for years.
CoreWeave borrows$46.9B of debt at 30 September.
It buys chips and buildsPaid for at the start.
Rent repays the loansIf it can borrow again cheaply.
Over $10B of debt is due within a year. Whether CoreWeave can roll it over cheaply decides how much of the rent is left for shareholders.
The story right now · 11 · What changed lately?
Sales beat the forecast, but it lost $709M
$3,661Mquarterly sales, from $1,365M
$379Mprofit before interest, above forecast
-$709Mafter interest
Near 15%what its bonds paid in October
Strong quarter
Bond buyers priced its debt like a borrower in trouble. It plans contract-backed loans at 7% to 7.5% instead, but none is priced yet.
The story right now · 12 · What is everyone arguing about?
Can it refinance its debt cheaply?
Many investors think
The leading AI cloud, with a huge signed order book
We think
Real contracts; it turns on borrowing cheaply
Against us: its average borrowing cost has fallen every year, to about 8%. If that continues, the funding worry fades.
The price and our view · 13 · What is a share worth?
No single value here: it is priced as a bet
0.8×if it works
13%kept if it fails
Multiples of the company's value when the bet was last struck, 9 Nov 2026.
We put no single value on a share here, because the business does not yet earn a lasting profit. Instead we ask what the shares pay if the plan works, and what is left if it fails.
The price and our view · 14 · How could we be wrong?
Debt it cannot roll cheaply, and a few big customers
Debt it cannot rollBiggest
$6.2B due in 2027, more to borrow.
A big customer stumblesWatch
Three pay 72% of sales.
Prices fall with supplyShowing now
July's rise came in a shortage.
Contract pricesUnknown
No one publishes the price per chip-hour.
Profit assumes each chip earns rent for six years. If chips go out of date faster, the losses will be bigger.
The price and our view · 15 · How will we know?
February shows whether the plan holds
2026 sales of $13.0B to $13.3B, and a $20B yearly paceGood sign
Quarterly interest of no more than $1.15BGood sign
A miss on this year's raised targetsWe turn negative
FEB 202725
Fourth-quarter 2026 results. Companies report every three months; each report is a test of the story.
Our view moves on these business events, and turns positive if the company's value, debt included, falls to 7 times a year's sales.
The price and our view · 16 · So, what do we think?
Labs sign for its chips; lenders doubt the debt
Priced for successthe kind of bet this is
NeutralMedium conviction
CoreWeave rents NVIDIA's chips to AI labs on contracts they must pay, and over $100B of rent is signed. But it is built with debt, interest swallows the profit, and the price already pays for success. Neutral.
↑ ↓ to move · C for contents
What it does
More about the company
Who it is
Training a large AI model takes tens of thousands of chips working as one machine for weeks. Few companies can build that themselves, so they rent it. CoreWeave buys NVIDIA's newest chips, wires them into huge clusters inside data centres it leases, and runs software that keeps them healthy and busy.
Its customers are AI labs, big tech companies and trading firms. Without it they would rent from Amazon, Microsoft or Google, from a rival specialist such as Nebius, or wait to build their own.
How it makes money
Almost all of its money, 98% in the June quarter, comes from customers who reserve a block of chips for one to six years and must pay whether they use it or not, often with part paid up front. The rest is rented by the hour. Each dollar arrives month by month over the contract, while the chips and buildings are paid for at the start.
Growth and profit
What growth costs
Growth is making it less profitable for now. Office and sales costs are small, but every new site brings rent, power and depreciation before it brings sales, and interest grows with the debt. The operating margin went from positive in 2024 to slightly negative this year; management says it turns once new sites fill.
Its record
The numbers
Is it growing?Fast
Sales grew 168% in the year to Dec 2025, slower than the 736% in the year before.
Is it profitable?Losing
Operating margin is −0.9%.
Does growth pay off?Costly
Running costs take 73¢ of each sales dollar, down from 76¢ in the year to Dec 2023.
How long can it keep going?Short
It used $7.25B of cash in the latest year, so the $3.16B it holds lasts about 0.4 years at that pace.
Is the balance sheet safe?Strained
It owes $18.2B more than the cash it holds, about 6.0 years of the cash its operations bring in.
Why it wins
The evidence
Right now the scarce thing in AI is not the chip but a powered building ready to run it. CoreWeave has more of that running and contracted than any other specialist, and NVIDIA, an investor, gets it the newest systems early.
The evidence is that it is raising prices, not cutting them: about 25% in July, and outside price indices for rented chips are rising too. The test is a glut: if new supply arrives faster than demand, this edge shrinks.
Who could beat it
The rivals in full
The biggest threat is its own customers. Microsoft, which paid two-thirds of last year's sales, and Meta build their own data centres and rent only to cover shortfalls. Rivals sell the same thing: Nebius now holds the same top rating, and Nscale has huge Microsoft and Anthropic contracts. Google sells its own chips to labs instead of NVIDIA's.
The whole job: the same NVIDIA clusters, now rated in the same top tier, sold to the same labs and to Microsoft and Meta.
Nscaleprivate
The same job for Microsoft and Anthropic, with statements of work worth more than CoreWeave's Meta and OpenAI orders combined.
Google TPU capacity sold to outside labsGOOGL
The NVIDIA GPU itself: a lab trains and serves models on Google's chips instead of rented NVIDIA clusters.
Microsoft's and Meta's own data-centre buildsMSFTMETA
Renting capacity at all: the largest customers build the halls themselves once power and chips reach them.
Amazon Web Services and Oracle Cloud InfrastructureAMZNORCL
The same rented GPU clusters sold inside a general cloud, often bundled with storage and data a customer already keeps there.
Inference platforms of the general clouds and model labsAMZNMSFTGOOGL
Managed inference: selling finished model answers per token rather than GPU time.
How big it could get
The market and who has signed
The market
One research firm measured the specialist AI cloud market at more than $25B in 2025, growing about threefold in a year, and the world's energy agency expects the electricity used by AI chips to grow 30% a year. CoreWeave was about a fifth of that market. Nobody publishes how many chips are rented out in total, so the size of the prize is a range, not a count.
Who has signed
Meta has committed up to about $35B across two orders, OpenAI up to about $6.5B and the trading firm Jane Street about $6B. With others, $103.7B was signed at June 30, and management says more than $25B more was added since. None of the customers has confirmed the contracts in its own filings.
Who runs it
Leaders, pay and the cash
Who runs it
The three founders, led by chief executive Michael Intrator, still run it and control it: their special ten-vote shares give insiders about 72% of the votes. Bonuses are paid at the board's discretion and name no sales or profit target, and their shares vest with time. Insiders bought no shares on the market in the past year and sold many, mostly on vesting days or under pre-set plans.
What they did with the money
Every dollar goes into chips and buildings. It is raised from loans backed by customer contracts, high-interest bonds, convertible bonds and share sales, including $2B to NVIDIA and $1B to Jane Street this year. In September it set up a program to sell up to 35M new shares. It pays no dividend and buys back no stock.
The last report
The latest report
CoreWeave's sales reached $3.66B, beating its guide, but it lost $709M
What it did to the thesis Strong quarter
It answered the demand question well: the guide was beaten, the year raised and prices kept rising. The funding question stays open: debt rose fast and the cheaper loan the call promised is not yet priced. The view stays Neutral.
The report
In the September 2026 quarter sales rose to $3,661M from $1,365M a year earlier, above the $3.45B to $3.6B guide. Profit before interest, on the company's own measure, was $379M against a guided $200M to $260M, and the year's targets were raised.
After $846M of interest it lost $709M, and debt reached $46.9B. Interest next quarter is guided above the profit before it.
On the call, management put its average borrowing cost at 7.9%, said prices kept rising, and expects its next contract-backed loan at 7% to 7.5%, not yet agreed.
What's happening now
The background
The September quarter then beat its guide and the year was raised again. The open question is the debt: $46.9B at September 30, over $10B due within a year. On the call management said it will fund 2027 with contract-backed loans at 7% to 7.5% rather than new bonds; none is priced yet.
Where we disagree
Both sides in full
The market
Most analysts see the leading AI cloud with a huge signed backlog; bond buyers price its debt like a borrower in trouble.
We believe
The contracts are real and the prices good; the shares turn on refinancing next year's debt without paying 15% for it.
The argument
Bulls treat the backlog as if it were cash, but rent arrives over years while the chips are paid for now. Bears treat the debt as if customers could walk away, but most contracts must be paid regardless.
Against us: the company's average borrowing cost has fallen every year, to about 8% now, and its contract-backed loans have won strong ratings. If that continues, the funding worry fades.
What it's worth
How we value it
No fair value: Q3 2026 was the first quarter with operating profit after stock pay, $207M, but interest of $846M left a pre-tax loss and nine months' operating profit was $14M, so there is no lasting profit to value yet; the bet prices it instead.
The bet
The bet is the new service that sells AI answers rather than chip time. If it grows to the size management's start suggests, it adds only a small slice to a company valued in the tens of billions, so it does not pay much on its own.
If it never grows, shareholders keep the chip-rental business, valued at this year's sales like the cheapest big cloud, which barely covers what the company owes.
The bet
Seen as a bet
The bet is the new service that sells AI answers rather than chip time. If it grows to the size management's start suggests, it adds only a small slice to a company valued in the tens of billions, so it does not pay much on its own.
If it never grows, shareholders keep the chip-rental business, valued at this year's sales like the cheapest big cloud, which barely covers what the company owes.
What could go wrong
Every risk
The risks
Debt it cannot roll cheaply
With $6.2B due in 2027 and tens of billions still to borrow, refinancing near 15% would hand most of the contracts' profit to lenders.
A big customer stumbles
Three customers pay 72% of sales; OpenAI and Jane Street are private, and one failing to pay would leave chips idle and debt unpaid.
Prices fall when supply arrives
July's 25% rise came during a shortage; a wave of new capacity in 2027 could push contract prices back down.
Chips wear out faster
Profit assumes each chip earns rent for six years; if new generations make them obsolete sooner, write-offs rise.
One supplier, more shares
Every chip comes from NVIDIA, and a 35M-share sale program and convertible bonds can dilute shareholders.
What we still do not know
Which customers are behind the 36% and 26% of sales, and how much of the backlog each holds.
What price per chip-hour the large contracts actually pay; no customer or index publishes it.
What the rating agencies would need to see before calling the company's own debt investment grade.
Whether the new inference service becomes a real business; its figures so far come only from the call.
What changes our view
Every route to a new view
Every route to a new view
→ Attractive
The company's value, debt included, falls to 7 times its past year's sales, from 7.3 on November 9, with this year's targets still intact. New borrowing over the next two reports costs 8% or less, or a rating agency raises the company's credit rating.
→ Unattractive
The December-quarter report misses this year's raised sales or profit targets, or puts the yearly billing pace below $20B. A customer paying more than a tenth of sales cuts, delays or fails to pay for capacity it committed to.
The next test
It supports the case if
Revenue and adjusted operating income inside or above the guide, a run-rate of at least $20B, interest at or below $1.15B and the largest customer below 30% of revenue.
It pressures the case if
Revenue or adjusted operating income below the guide, a run-rate under $20B, interest above $1.15B, or a customer above 10% of revenue cutting or deferring capacity.
We would also be wrong if
The yearly billing pace at the end of 2026 comes in below $20B.
New contracts are signed below July's prices.
The verdict
Scores and how we weighed it
The scores
Growth
3 of 4
Business quality
3 of 4
Financial position
2 of 4
Execution record
3 of 4
Per share capture
2 of 4
Potential
3 of 4
Risk profile
2 of 4
Valuation
3 of 4
Evidence quality
3 of 4
How we weighed it
ValueMixed
No fair value is drawn: the forecasts show no lasting profit after stock pay, so the bet prices it.
BusinessMixed
Demand is real and contracted, but the business is built with debt that eats the profit.
GrowthPositive
Sales nearly tripled on a year ago, the year's guide was raised again and the signed backlog keeps growing.
RisksNegative
Large debts come due soon, three customers pay most of sales, rental prices face new supply and chips wear out.
RivalsMixed
Including debt it sits mid-pack among cloud rivals.
The groupMixed
AI hosting names move together on AI spending and funding news.
Latest newsMixed
The September quarter beat every part of the guide and raised the year; on the call prices kept rising and cheaper contract-backed loans were promised, but none is priced yet and interest still exceeds profit before it.