A multi-dimensional sensitivity analysis, inspired by the phase diagram of physics.
Move the levers and watch which of three states the GPU lands in. Like water, it changes state when squeezed: price is the pressure, the cost of money is the heat.
Swipe sideways to see the whole chart →
Click or drag on the diagram to move the two levers on its axes. All other levers stay where the sliders are. The three small diagrams are the same GPU seen along each pair of the three big levers. The melt path shows on price against rentable life; CoreWeave’s debt shows on the two Fed-rate views.
Rent per GPU-hour, averaged over the GPU’s rentable life. Percentages are shares of the first bar.
Swipe sideways to see the whole chart →
| Step | Per GPU-hour | Rent per GPU over its rentable life |
|---|
Each row moves one lever alone, by 10% of its value now, and holds the rest at the slider values. “Better” means a higher price, a fuller fleet or a longer life, and a lower cost or rate. Loan cover is rent banked divided by loan payments due, at the tightest year. Loan share, loan length and the owners’ required return are left out: they do not push the result in one direction.
Cumulative dollars per GPU. New financing is needed the first year rent banked ends below loan payments due.
Each row moves one lever alone and holds the rest at the slider values.
| Lever | Now | Owners miss their target when it is | New financing is needed when it is |
|---|
Range: a 1.00 point cut to a 3.00 point hike, in quarter-point steps. Floating bars move; fixed bars stay put.
Each bar is one tranche. Width is the amount owed, height is the effective interest rate, so a bar's area is its yearly interest cost. Snapshot as of September 30, 2026.
Swipe sideways to see the whole chart →
| Tranche | Owed ($M) | Effective rate | Rate type | Pricing basis | Matures |
|---|
CoreWeave sells contracts: 98% of its revenue is customers committed to pay for a fixed term. Those prices are negotiated and not published, so this chart gathers every committed or reserved H100 price that is public. They cluster between $2.46 and $3.69. The model starts in the middle of that range, at $3.00. On-demand prices, in grey, are what a walk-up customer pays by the hour and are shown only for contrast.
Does the choice within that range change the verdict? No. The filing fixes how much rent CoreWeave collects per dollar of equipment (36% a year). A higher contract price simply means a smaller share of the fleet is earning: $2.46 would imply 84% of hours, $3.00 implies 69%, and $3.69 would imply 56%. The model lands in the same place either way. “Up to 60%” is the largest discount CoreWeave advertises, so $2.46 is a floor for its published terms. The price-index figure is a median for all GPUs released since 2024, not the H100 alone.
On-demand is the dearest way to rent. Committing for months or years, or accepting that the provider can take the GPU back, can cut the price by as much as four fifths.
| Provider | On demand | Committed | Interruptible |
|---|---|---|---|
| CoreWeave“Up to 60%” off for committed use | $6.16 | from $2.46 | – |
| Amazon Web ServicesThree-year reserved; spot | $6.88 | $2.97 | $3.82 |
| Microsoft AzureSpot | $6.98 | – | $1.42 |
| Together AIReserved 31 to 90 days; preemptible | $3.99 | $3.45 | $1.99 |
| NebiusPreemptible, starting price | $4.50 | ask | from $0.79 |
| RunPodCommunity hosts, starting price | $4.39 | – | from $2.69 |
| LambdaLinked cluster of 256 GPUs, two weeks to a year | $4.59 | $5.54 | – |
Lambda’s cluster price is higher than its single-server price because the GPUs come wired together for training. One sign of direction: Crusoe’s page announces that its H100 dedicated endpoint rises from $5.50 to $6.30 on December 1, 2026. A price index puts the median for GPUs released since 2024 at $3.35 on demand and $2.92 on a twelve-month reservation.
The model only cares about this ratio: rent collected in a year, divided by what the GPU cost to buy and install. Running costs and interest come out of it afterwards.
| Case | Rent a year | Hardware | Ratio |
|---|---|---|---|
| Your levers | |||
| CoreWeave, as reportedQ3 2026 revenue, times four, against average technology equipment at cost | $14.6B | $40.4B | 36% |
| H100 at CoreWeave’s lowest advertised price$2.46 every hour of the year; cluster cost estimated for 24,576 GPUs, May 2024 | $21,600 | $37,000 | 58% |
| B200 at CoreWeave’s lowest advertised price$3.44 every hour of the year; cluster cost for 24,576 down to 576 GPUs | $30,100 | $60,000 to $77,000 | 39% to 50% |
| H100 at CoreWeave’s list price$6.16 every hour of the year; almost nobody pays this at scale | $53,900 | $37,000 | 146% |
Hardware figures are AMCompute’s cluster estimates: servers, networking and storage, not the building. A used eight-GPU H100 server was valued at $195,000 in August 2026, or $24,400 a GPU. CoreWeave’s reported ratio is lower than the H100 price-sheet row. Its equipment grew by $13.2 billion in the quarter, so some of it earned nothing for part of the period, and the total includes storage and other gear that is not rented by the GPU-hour.
Only list prices leave a public record, so this history is of on-demand prices; contract prices are assumed to follow them. Not steadily. Archived copies of the same price pages, read quarter by quarter back to 2021, show long flat stretches broken by sharp steps in both directions. Prices dropped when a seller started far above the market or a newer GPU displaced an old one. In 2026 they rose.
| GPU and seller | First price found | Then | Nov 2026 | Change | Per year | What happened |
|---|---|---|---|---|---|---|
| H100, Lambda | Oct 2023 | $2.59 | $4.59 | +77% | +20% | Up to $3.49 in 2024, down to $2.99 for 2025, up twice in 2026 |
| H100, CoreWeave | Apr 2023 | $6.20 | $6.16 | −1% | 0% | Unchanged for three and a half years |
| H100, Amazon | Jul 2023 | $12.29 | $6.88 | −44% | −16% | One cut, in June 2025, from far above the specialists |
| A100 40GB, Lambda | Jul 2022 | $1.10 | $1.99 | +81% | +15% | $1.29 from 2024, $1.99 from spring 2026 |
| A100 80GB, CoreWeave | Oct 2022 | $3.65 | $2.70 | −26% | −7% | One cut, around the start of 2025 |
| V100, Lambda | Jan 2021 | $1.50 | $0.79 | −47% | −10% | Fell to $0.55 by mid-2022 as the A100 arrived; back up in 2026 |
These are public on-demand prices, not the contract prices that make up 98% of CoreWeave’s revenue. Source: the Internet Archive’s saved copies of Lambda’s and CoreWeave’s pricing pages, one per quarter from January 2021 to July 2026, plus today’s live pages. A price change is dated to the first quarterly copy that shows it. Lambda’s pages from mid-2021 to spring 2022 showed no prices, so the V100’s fall is dated only to within that gap. Until 2024 CoreWeave priced the GPU, processors and memory separately; its figures here add 16 processor cores and 256 GB of memory per GPU at its listed rates, to match the servers it sells now. Amazon’s 2023 price is worked back from its June 2025 cut of 44%. Per-year figures are compound rates between the first price found and November 2026.
These are list prices for renting by the hour with no commitment, read on November 20, 2026. They are the only prices every seller publishes, which is why they are here. They are not what the model uses, and at CoreWeave they account for 2% of revenue. Each dot is one seller; the blue line runs from CoreWeave’s list price down to its lowest advertised contract price.
Swipe sideways to see the whole chart →
A dash means the provider does not list that GPU at a public price. CoreWeave lists the B300 and GB300 as “contact sales”; its GB200 rack system is $42.00 an hour for a four-GPU instance, or $10.50 per GPU-hour.
One phase diagram per company, on the same two axes: what its debt costs, and how much rent it collects in a year for each dollar of equipment. The black dot is the company at its average cost of debt. The blue diamond is the same company if all its debt cost what its GPU-backed loans cost. Loan length, rentable life, price fall and owners’ target are the starting values from the Server economics tab for all three.
The three are not at the same stage, and the diagrams should be read with that in mind. CoreWeave is measured against technology equipment at cost, as on the Server economics tab. IREN is measured against GPU hardware, buildings and other equipment in service at cost, because it owns its data centers; its GPU fleet nearly tripled during the quarter, so most of that equipment earned little. Nebius’s quarterly report gives only one figure for all property, including sites under construction, so its 24% is a floor: on that same broad basis CoreWeave would read 28%, not 36%. Each company’s regions differ slightly because its running costs and its share of debt differ.
Nebius and IREN are the two other listed GPU clouds whose filings can be read the same way. On rent earned per dollar of property, CoreWeave is the strongest of the three. Where it differs is the cost of its money: its GPU-backed loans cost about 9%, while the other two fund themselves mostly with convertible notes at 4% or less.
| Measure | CoreWeave | Nebius | IREN |
|---|---|---|---|
| Revenue in the quarter | $3,661M | $981M | $228M |
| Property and equipment, net, average over the quarter | $51.8B | $16.2B | $9.5B |
| Rent in a year against that property | 28% | 24% | 10% |
| Cash running costs as a share of revenue | 38% | 53% | 72% |
| Left after running costs, against property | 17% | 11% | 3% |
| Debt at quarter end | $47.5B | $14.2B | $10.0B |
| Debt against net property | 84% | 73% | 81% |
| Average cost of debt | 7.9% | about 4.0% | 3.5% |
| Rate on loans secured on GPUs | 9.4% | not in the quarterly | 10.4% to 10.7% |
| Share of debt that is convertible notes | 23% | most of it | 67% |
| Customer prepayments held (deferred revenue) | $10.0B | $9.4B | not read |
How this table differs from the model. It measures rent against all property and equipment after depreciation, including buildings and construction in progress, because Nebius’s quarterly report does not break out technology equipment. On that basis CoreWeave’s ratio is 28%; the model uses 36%, measured against technology equipment at cost. Nebius and IREN own data centers that CoreWeave mostly leases, so their property is larger and their running costs carry less rent. Nebius’s figures include its smaller non-cloud businesses. IREN’s GPU hardware nearly tripled in the quarter, from $2.0 billion to $5.8 billion at cost, and $28 million of its revenue is bitcoin mining, so its ratios say little yet. Running costs exclude depreciation and stock-based pay. Nebius’s cost of debt is its interest expense over average debt; the other two are dollar-weighted effective rates from their debt notes. Sources: CoreWeave 10-Qs for Q2 and Q3 2026; Nebius Forms 6-K of August 12 and November 12, 2026; IREN 10-Q filed November 5, 2026.
Squeezed into one average GPU, CoreWeave’s filed ratios give a loan cover of 0.97, where 1.00 is the line. Lenders are nearly whole. Owners are well short: they would get back $0.43 for each $1 put in. Whether that changes depends on four things that have not happened yet, and this page sets out what they are and how to test them.
No. The red zone means rent from a GPU does not pay off its loan in five years. A company in that position has to keep refinancing or raising new money, and that is what the filings show: debt grew during the quarter. It is a statement about what the business needs, not a forecast of a missed payment.
Close enough that one assumption decides the side. The starting case turns from red to amber at a price of $3.06 instead of $3.00, at just over 70% of fleet hours earning instead of 69%, or if prices fall less than 4.0% a year instead of 5%. Price history on the Rental prices tab puts the 5% inside the observed range, on the cautious side. It also turns amber if the loan costs under about 8.0%, and CoreWeave’s average across all its debt is 7.9%; the model uses 9.4%, the rate on its GPU-backed term loans.
Much further. Owners need about $3.47 an hour to earn their target return of 14.9% a year, against $3.00 now. No single sourced adjustment gets them there. Lenders are paid first, so everything that goes right has to cover the debt before owners see any of it.
Yes. At the Q3 pace, revenue less cash running costs is about $9.0 billion a year. Interest on debt is about $3.3 billion, and repaying $47.5 billion over five years would take about $9.5 billion a year. That is $12.8 billion needed against $9.0 billion earned. In the quarter itself CoreWeave reported operating income of $207 million and net interest expense of $846 million.
A new H100 cluster at about $37,000 a GPU, rented at the same $3.00 for 90% of hours, earns 64% of its cost a year and lands in the green zone with a cover of 2.00. The fleet earns 36%. Something drags the average down: equipment installed but not yet earning, storage and other gear that is not rented by the GPU-hour, or contract prices below the price sheet. The filing does not say which.
On the Q3 numbers the equity case rests on things that have not happened yet. Not all of them are matters of belief: 98% of revenue comes from take-or-pay contracts, so contracted capacity does pay once it is switched on. But prices at renewal, the working life of a GPU and the cost of refinancing are forecasts. The narrow finding is this: the business as it stands does not earn its owners’ cost of capital, and several favourable things must happen together for it to do so.
Four things. The whole company is squeezed into one average GPU, which hides the mix of old and new machines. Rent is measured against equipment that grew by $13.2 billion in a single quarter, so the 36% is still catching up. The yearly price fall, the loan length and the owners’ required return are assumptions. And customer prepayments, which bring cash in early, are left out.
Revenue lags equipment, and the lag is large right now. Committed contracts fix the price for their term, so a 5% yearly fall is harsher than what a contract in force actually experiences. And public prices are firm or rising: one provider has announced a 15% increase on its H100 endpoint for December 1, 2026.
The frame comes from a talented VC of a former venture of mine, who uses it to weigh an investment. What are you asking me to believe? What experiment would show whether it is true? How much money does that experiment take?
| The belief | The experiment | What it costs | Reading today | Pass | Fail |
|---|---|---|---|---|---|
| New equipment fills up | Revenue for the quarter, times four, against average technology equipment at cost, in each new 10-Q | Nothing to watch. For holders, about $3.3 billion a year of interest while waiting | 36% | Rising toward the 64% a new H100 earns | Flat near 36% while debt keeps growing |
| Prices hold at renewal | Renewal terms on the first large H100 contracts; this page’s price sheet re-read each quarter | Time. The first big renewals are probably years away | H100 list $6.16, unchanged since April 2023; lowest advertised $2.46. Lambda’s H100 is up 77% since October 2023 | List prices steady, discounts no deeper | List prices cut, or discounts beyond 60% |
| A GPU can be rented for six years | Whether 2022-era H100s still rent in 2028 | Six years. There is no shortcut | The A100, a 2020 card, still lists at $2.19 to $3.19 an hour | Old cards keep a public price near half the new one | Old cards drop off price sheets or are written down early |
| Debt can be refinanced more cheaply | The effective rate on the next facility, against 9.4% on the term loans today | Nothing to watch | Mixed. Term loans run from 15% on the oldest to 7%, but the two due in late 2031 are at 9% and 11% | New facilities at 7% or below | New facilities at 10% or above |
The third question is where the frame bites. It assumes the experiment costs less than the bet. Here the only way to test whether new capacity fills at these prices is to buy the capacity, and that ran at $13.2 billion of equipment in Q3 alone. The experiment and the investment are the same size, and both are funded with debt that must be serviced whatever the result.
Each of these would move a lever on the Server economics tab. Until they are answered, the levers are there so a reader can try their own answer and see where the GPU lands.
The model needs four numbers from a company: rent against hardware cost, cash running costs against hardware cost, the rate on its GPU loans, and the share of hardware paid for with debt. Nebius and IREN are now done, as far as their filings allow, on the Other neoclouds tab. One more has a prospectus to read, and the rest are private. Status for the companies not yet read comes from search results in November 2026.
| Company | Status | What is available | Next step |
|---|---|---|---|
| Nebius | Listed on Nasdaq | Read: quarterly reports for Q2 and Q3 2026. They omit the property and debt notes | Read the annual report for technology equipment at cost and each note’s coupon, so Nebius can go on the phase diagram |
| IREN | Listed on Nasdaq | Read: 10-Q for the quarter to September 30, 2026. It separates GPU revenue and GPU hardware | Re-read in two quarters. The fleet tripled in the quarter, so today’s ratios say little |
| Nscale | Filed to list in New York in September 2026, per press reports | A prospectus. Press reports give about 25,000 active GPUs, $141 million of first-half revenue and debt of $5.5 billion or more | Read the prospectus itself; the press figures for debt disagree. Too early in its build to model |
| Lambda | Private. No listing document found; a listing is reported for 2027 | Press reports only: two secured loans of about $1 billion each in 2026. Already on the price sheet | Watch for a listing document. Until then only its prices can be tracked |
| Crusoe, Together AI, RunPod | Private | Public prices only | Keep on the price sheet; no financials to model |
| Applied Digital, Core Scientific, TeraWulf, Cipher, Hut 8, Galaxy | Listed | Full filings, but these are landlords: they rent buildings and power to GPU clouds, not GPU-hours | Use their lease terms to source the data center rent lever, which is currently a rough split |
| Amazon, Microsoft, Oracle | Listed | GPU rental is buried inside much larger cloud segments | Prices only. The financials cannot be separated |
I find it tedious to review pro formas, market models and other analytical work that lacks the ambition to make a forecast, stick its head out and risk being wrong. I don’t do this professionally, so I have nothing to lose. Let me take the risk and predict the moves of CoreWeave’s CFO.
Cheaper hardware and cheaper running costs pull the lines down. A higher blended price lifts the GPU up. The dashed line is where the repayment line sat at Q3 2026.
Moves are applied in the order shown, each on top of the last. The order changes how much each step appears to add, but not where the plan ends.
| Who | The ask | In dollars | What has to be true |
|---|
The 70% GPU share of installed cost, and every starting value on the sliders, are assumptions made for this exercise. None comes from a filing. The model also does not count the customer prepayments CoreWeave already holds.
I, Erik Bethke, am a 54-year-old game and AI developer. I am not an equity analyst, a hedge fund manager, a venture capitalist or a private equity investor, and I hold no FINRA credentials. What I am doing is using public information, as a citizen trying to work out how to take part in the AI productivity boom, either by buying equity or by lending into the debt layer. I read the public statements and try to evaluate them from first principles. That, after all, is the whole point of a public market and its SEC filings.
This is not investment advice. I hold no position, long or short, in CoreWeave, Nebius or IREN. I am long Nvidia, and I own an S&P 500 index fund.
| Used for | Source | As of | How firm it is |
|---|---|---|---|
| Debt stack: balances, effective rates, maturities, swap notional, rate sensitivity | CoreWeave Form 10-Q, debt note and market-risk section. Open the filing | Sep 30, 2026 | Filed figures. Rate types marked * are inferred, see the debt notes below. |
| Starting case: rent and share of hours rented | CoreWeave 10-Qs for Q2 and Q3 2026 (Q2, Q3) and its public price sheet. Q3 revenue of $3,661 million is $14.6 billion a year. Technology equipment at cost averaged $40.4 billion over the quarter ($33.8 billion on June 30, $47.0 billion on September 30). Revenue is 36.2% of that. | Q3 2026 | The 36.2% is filed. The price of $3.00 is a judgement: the middle of the public contract prices for an H100 ($2.46 to $3.69, on the Rental prices tab) and close to Amazon’s three-year price of $2.97. The 69% of fleet hours earning rent is what is left over: the share that makes $3.00 on a $50,000 GPU equal 36.2%. It is not a disclosed occupancy rate. Any other price in the range gives the same verdict with a different share. |
| Starting case: cost to buy and install one GPU, $50,000 | AMCompute cluster estimates, linked on the Rental prices tab: about $37,000 a GPU for an H100 cluster and $60,000 to $77,000 for a B200 cluster. | 2026 | A round figure for a mixed fleet, not a disclosure. Only ratios matter to the model: a different cost per GPU with the same 36.2% gives the same verdict. |
| Starting case: power, rent and staff, $0.80 per GPU-hour | Q3 10-Q income statement. Operating expenses of $3,454 million less about $1,890 million of depreciation and $160 million of stock-based pay leave about $1,404 million of cash running costs in the quarter, 13.9% of average technology equipment a year. | Q3 2026 | The total is filed. Depreciation is taken as the quarter’s rise in accumulated depreciation. The split is rough: cost of revenue (84% of the total) is put into power and data center rent, the rest into staff. The filing does not separate power from rent. |
| Starting case: rentable life, 6 years | CoreWeave depreciates technology equipment over six years, per its 2025 annual report as summarised in search results. The Q3 10-Q is consistent: depreciation of about $7.6 billion a year on an average $47.9 billion of depreciable property is 6.3 years. | Q3 2026 | The life of a GPU is an open argument across the industry: some use three years, some five, some six. Six is used here to be as generous as possible to CoreWeave. It is an accounting life, which may be longer than the time a GPU can actually be rented. The annual report itself was not opened. |
| Starting case: lender’s margin, 5.5 points | Debt stack tab. The eight delayed-draw term loans, CoreWeave’s GPU-backed borrowing, total $17.6 billion at a dollar-weighted effective rate of 9.4%. That is 5.5 points above the 3.9% Fed rate. | Sep 30, 2026 | Filed balances and rates, rounded to whole percents in the filing. |
| Starting case: share paid with a loan, 75% | Q3 10-Q. Debt principal of $47.5 billion is 74% of the $64.6 billion of property and equipment at cost. | Sep 30, 2026 | A company-wide ratio. The filing does not give the advance rate on individual GPU loans. |
| Starting case: the remaining levers | No source. A five-year loan and owners requiring 5.5 points above the loan rate are assumptions. The 5% yearly price fall is set against six price histories on the Rental prices tab, which run from a 16% yearly fall to a 20% yearly rise with a middle of 3% to 4% down. | Not dated | Assumptions. The loan length is in line with the term loans’ maturities (2028 to 2032). Committed contracts fix the price for their term, so the 5% fall is on the cautious side; the “Contract prices hold” scenario removes it. |
| Fed rate lever, starting value of 3.9% | Federal Reserve statement and implementation note of October 28, 2026: target range 3.75% to 4.00%, interest on reserve balances 3.90%. Open the statement | Oct 28, 2026 | Published rate. |
| S&P 500 index fund comparison | Vanguard 500 Index Fund Admiral Shares (VFIAX), average annual total return before taxes, as shown on Schwab’s fund performance page | Sep 30, 2026 | 15.29% a year over ten years, 13.74% over five, 8.90% since the share class began in November 2000. Third-party display of fund data, re-read on November 20, 2026; not checked against Vanguard’s own page. |
| Treasury comparison | No outside figure. It compounds at the Fed rate lever. | Live | A stand-in for rolling short Treasury bills. |
Everything on this site is in two plain files: a brief written for an AI to read, and the data behind every chart. The brief carries the model itself as about sixty lines of code, every input with where it came from, and test cases with the right answers, so your AI can prove its copy works before it changes anything.
Agents can fetch the same two files directly: coreweave-pro-forma.md and coreweave-pro-forma.json, in the same folder as this page.
You are given a brief for an open financial model of CoreWeave, a GPU cloud. 1. Run the model code on every test case in the brief and confirm you get the expected answers. Tell me if any do not match. 2. Then change these assumptions: [say what you want changed]. 3. Report the state, the loan cover and what owners get back per $1, before and after. 4. List every change you made from the original, and credit Erik Bethke, erikbethke.com.