# CoreWeave: A pro forma

A multi-dimensional sensitivity analysis, inspired by the phase diagram of physics.

- Author: Erik Bethke, https://erikbethke.com
- Version 1.0, 2026-11-21
- Made with Bike4Mind, https://bike4mind.com
- Licence: Creative Commons Attribution 4.0 International (https://creativecommons.org/licenses/by/4.0/). Credit Erik Bethke, erikbethke.com, and link back. State what you changed from the original. Do not present a reworked version as the author's.
- Not investment advice. The author holds no position, long or short, in CoreWeave, Nebius or IREN; he is long Nvidia and owns an S&P 500 index fund. He is a game and AI developer, not a securities professional.

## For the AI reading this

This file is a complete, self-contained brief. Before changing anything, run the model code below on every test case and confirm the expected answers. If you rework it, list every change from the original and credit the author.

## What the model is

One fleet-average GPU, bought partly with a loan and rented by the hour. The model asks whether rent banked keeps ahead of the loan payments, and whether owners then earn their required return.

Three states:

- 0, "Needs new financing": Rent banked falls behind the loan payments in some year. New capital or a refinancing has to fill the gap.
- 1, "Lenders paid, owners short": The loan is repaid on schedule, but owners earn less than their required return.
- 2, "Profitable": The loan is repaid and owners beat their required return.

## The model code

```javascript
var HOURS = 8760;
function evaluate(p) {
  var rd = (p.fed + p.spread) / 100;
  var re = rd + p.eqprem / 100;
  var debt = p.capex * p.ltv / 100;
  var equity = p.capex - debt;
  var term = Math.round(p.term);
  var pay = 0;
  if (debt > 0) {
    pay = rd > 1e-9 ? debt * rd / (1 - Math.pow(1 + rd, -term)) : debt / term;
  }
  var u = p.util / 100;
  var keep = 1 - p.decline / 100;
  var opexHr = p.power + p.rent + p.staff;
  var nYears = Math.ceil(p.life - 1e-9);
  var horizon = Math.max(nYears, term);
  var retired = false;
  var lifeHours = 0, potential = 0, revenue = 0;
  var cumCF = 0, cumDS = 0, worst = null, worstT = null, bust = false;
  var npv = -equity, eqSum = 0;
  var series = [];
  var t;
  for (t = 1; t <= horizon; t++) {
    var priceT = p.price * Math.pow(keep, t - 1);
    var marginHr = priceT * u - opexHr;
    var frac = 0;
    if (!retired && t <= nYears) {
      if (t > 1 && marginHr < 0) {
        retired = true;
      } else {
        frac = Math.min(1, p.life - (t - 1));
      }
    }
    var hours = frac * HOURS;
    var cf = marginHr * hours;
    lifeHours += hours;
    potential += priceT * hours;
    revenue += priceT * u * hours;
    var ds = t <= term ? pay : 0;
    cumCF += cf;
    cumDS += ds;
    if (ds > 0) {
      var cover = cumCF / cumDS;
      if (worst === null || cover < worst) { worst = cover; worstT = t; }
      if (cumCF < cumDS - 0.01) { bust = true; }
    }
    var eq = cf - ds;
    eqSum += eq;
    npv += eq / Math.pow(1 + re, t);
    series.push({ t: t, cf: cumCF, ds: cumDS, eq: eq });
  }
  var collected = revenue / lifeHours;
  var cashMargin = collected - opexHr;
  var hardware = p.capex / lifeHours;
  var interest = (pay * term - debt) / lifeHours;
  var owners = cashMargin - hardware - interest;
  var state = bust ? 0 : (npv >= 0 ? 2 : 1);
  return {
    state: state, rd: rd, re: re, debt: debt, equity: equity, pay: pay, term: term,
    lifeHours: lifeHours, list: p.price,
    erosion: p.price - potential / lifeHours,
    idle: (potential - revenue) / lifeHours,
    collected: collected, power: p.power, rent: p.rent, staff: p.staff,
    cashMargin: cashMargin, hardware: hardware, interest: interest, owners: owners,
    worst: worst, worstT: worstT, eqSum: eqSum, npv: npv, series: series, horizon: horizon
  };
}
```

What it returns:

- `state`: 0, 1 or 2, as above
- `worst`: loan cover at the tightest year: cumulative rent banked divided by cumulative loan payments due. Below 1.00 is state 0
- `eqSum / equity`: what owners get back per $1 put in, undiscounted
- `npv`: owners' payouts discounted at their required return, less what they put in. Negative means state 1 at best
- `rd`: loan rate as a fraction
- `re`: owners' required return as a fraction

Units: `price`, `power`, `rent` and `staff` are dollars per GPU-hour; `capex` is dollars per GPU; `life` and `term` are years; everything else is a percentage or percentage points.

## Inputs and where each comes from

| Key | What it is | Value | Grade | Source |
|---|---|---|---|---|
| price | Price actually paid, when new | 3 | estimated | Middle of the public contract prices for an H100, $2.46 to $3.69 per GPU-hour, read November 20, 2026. CoreWeave does not publish what its contracts pay. |
| life | Years the GPU can still be rented | 6 | assumed | Six years. The industry argues over three, five or six; six is the most generous to CoreWeave and matches its accounting life. |
| fed | Fed rate | 3.9 | estimated | Short-term policy rate in November 2026. |
| decline | Yearly fall in rental price | 5 | assumed | Rental price falls 5% a year. The price history on this site has a median of about 3% to 4% a year; 5% is on the cautious side. |
| util | Share of fleet hours earning rent | 69 | derived from filing | Set so rent matches the filing: Q3 2026 revenue of $3,661M times four, divided by average technology equipment at cost of $40,423M, is 36.2% a year. At $3.00 and $50,000 that needs 69% of hours earning. |
| spread | Lender's margin above the Fed rate | 5.5 | derived from filing | CoreWeave's GPU-backed delayed-draw term loans carry a weighted rate of 9.36% (Q3 2026 10-Q), which is 5.5 points over a 3.9% Fed rate. |
| ltv | Share of the purchase paid with a loan | 75 | derived from filing | Total debt principal of $47.5B against $64.6B of property at cost is 74% (Q3 2026 10-Q). |
| term | Years to repay the loan | 5 | assumed | Five years to repay. |
| eqprem | Extra return owners require above the loan rate | 5.5 | assumed | Owners require 5.5 points above the loan rate. |
| capex | Cost to buy and install one GPU | 50000 | estimated | A blend of vendor cluster cost estimates (AMCompute): about $37,000 per H100 and $60,000 to $77,000 per Blackwell B200, all-in. Not CoreWeave's purchase prices, which it does not disclose. |
| power | Power and cooling | 0.12 | derived from filing | Split of cash running costs. In total they are 13.9% of average equipment a year (Q3 operating expenses less depreciation and stock pay), about $0.80 per GPU-hour. The split between the three cost levers is assumed. |
| rent | Data center rent | 0.55 | derived from filing | See power. Share of the $0.80 per GPU-hour of cash running costs. |
| staff | Staff, software and overhead | 0.13 | derived from filing | See power. Share of the $0.80 per GPU-hour of cash running costs. |

Grades: "derived from filing" is computed from CoreWeave's 10-Q; "estimated" comes from public prices or vendor estimates; "assumed" is the author's choice.

## Test cases

| Case | price | life | fed | decline | util | spread | ltv | term | eqprem | capex | power | rent | staff | state | loan cover | owners back per $1 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| CoreWeave, Q3 2026 | 3 | 6 | 3.9 | 5 | 69 | 5.5 | 75 | 5 | 5.5 | 50000 | 0.12 | 0.55 | 0.13 | 0 Needs new financing | 0.965 | 0.43 |
| Contract prices hold | 3 | 6 | 3.9 | 0 | 69 | 5.5 | 75 | 5 | 5.5 | 50000 | 0.12 | 0.55 | 0.13 | 1 Lenders paid, owners short | 1.142 | 1.44 |
| One new H100 | 3 | 6 | 3.9 | 5 | 90 | 5.5 | 75 | 5 | 5.5 | 37000 | 0.12 | 0.55 | 0.13 | 2 Profitable | 1.997 | 5.11 |
| Fast obsolescence | 3 | 3.5 | 3.9 | 20 | 69 | 5.5 | 75 | 5 | 5.5 | 50000 | 0.12 | 0.55 | 0.13 | 0 Needs new financing | 0.5 | -1.95 |
| Tight credit | 3 | 6 | 3.9 | 5 | 69 | 12 | 75 | 5 | 5.5 | 50000 | 0.12 | 0.55 | 0.13 | 0 Needs new financing | 0.823 | -0.25 |
| The author's prediction for the CFO | 3.365406 | 6 | 3.9 | 5 | 69 | 5 | 75 | 5 | 5.5 | 45000 | 0.12 | 0.495 | 0.13 | 2 Profitable | 1.373 | 2.25 |

## Figures taken from the filings

| Item | Value | Source |
|---|---|---|
| Revenue, Q3 2026 | $3,661 million | CoreWeave Form 10-Q, quarter ended September 30, 2026 |
| Technology equipment at cost, June 30, 2026 | $33,823 million | Form 10-Q, Q2 2026 |
| Technology equipment at cost, September 30, 2026 | $47,023 million | Form 10-Q, Q3 2026 |
| Rent a year against average equipment at cost | 36.2% | Computed: $3,661M x 4 / $40,423M |
| Cash running costs a year against average equipment | 13.9% | Computed: Q3 operating expenses of $3,454M less about $1,890M depreciation and $160M stock pay, times four |
| Total debt principal | $47.5 billion | Form 10-Q, Q3 2026, debt note |
| Average cost of all debt | 7.9% | Computed from the tranche table below |
| Weighted rate on GPU-backed term loans | 9.36% | Form 10-Q, Q3 2026 |
| Share of revenue from committed contracts | 98% | Form 10-Q, Q3 2026 |

## CoreWeave's debt, September 30, 2026

| Tranche | Owed ($M) | Effective rate (%) | Type | Basis | Due |
|---|---|---|---|---|---|
| 2031 convertible notes | 2588 | 2 | fixed | 1.75% coupon | Dec 2031 |
| 2032 convertible notes | 4000 | 2 | fixed | 1.75% coupon | Oct 2032 |
| 2033 convertible notes | 4200 | 3 | fixed | 2.875% coupon | Apr 2033 |
| DDTL 4.0, fixed-rate piece | 2383 | 7 | fixed | Treasury + 2.00%, set at each draw | Mar 2032 |
| DDTL 4.0, floating-rate piece | 2200 | 7 | float | SOFR + 2.25% | Mar 2032 |
| Revolving credit facility | 1200 | 7 | float (inferred) | Not given in this filing | Nov 2029 |
| 2032 euro senior notes | 2272 | 9 | fixed | 8.50% coupon | Jul 2032 |
| OEM and software financing (non-recourse) | 770 | 9 | fixed (inferred) | Not given in this filing | Oct 2026 to Aug 2028 |
| DDTL 2.1 | 2494 | 9 | float (inferred) | Not given in this filing | Mar 2031 |
| DDTL 3.0 | 2323 | 9 | float (inferred) | Not given in this filing | Aug 2030 |
| DDTL 5.0 | 2749 | 9 | float | SOFR + 4.50% | Nov 2031 |
| 2030 senior notes | 2000 | 10 | fixed | Fixed coupon (rate not given in this filing) | Jun 2030 |
| 2031 9.00% senior notes | 1750 | 10 | fixed | 9.00% coupon | Feb 2031 |
| 2031 9.75% senior notes | 2750 | 10 | fixed | 9.75% coupon | Oct 2031 |
| 2032 9.625% senior notes | 1250 | 10 | fixed | 9.625% coupon | Jul 2032 |
| OEM and software financing (recourse) | 6891 | 10 | fixed (inferred) | Not given in this filing | Dec 2026 to Jul 2030 |
| DDTL 2.0 | 2796 | 11 | float (inferred) | Not given in this filing | Aug 2030 |
| DDTL 5.5 | 1555 | 11 | float | Term SOFR + 5.50% | Sep 2031 |
| Magnetar loan | 195 | 12 | fixed (inferred) | Accruing redemption premium | Jan 2029 |
| DDTL 1.0 | 1139 | 15 | float (inferred) | Not given in this filing | Mar 2028 |

## What an H100 rents for, November 20, 2026

| Price | $ per GPU-hour | Kind |
|---|---|---|
| CoreWeave, list less its full 60% discount | 2.46 | contract |
| Price index, 12-month reservation, GPUs since 2024 | 2.92 | contract |
| Amazon, three-year reservation | 2.97 | contract |
| Together AI, reserved 31 to 90 days | 3.45 | contract |
| Together AI, reserved 3 to 30 days | 3.69 | contract |
| On demand, middle of seven sellers | 4.59 | on demand |
| CoreWeave list price, on demand | 6.16 | on demand |

## On-demand list prices, $ per GPU-hour

| Provider | H100 | H200 | B200 | B300 | A100 | L40S |
|---|---|---|---|---|---|---|
| CoreWeave | 6.16 | 6.31 | 8.6 |  | 2.7 | 2.25 |
| Microsoft Azure | 6.98 |  |  |  |  |  |
| Amazon Web Services | 6.88 |  |  |  |  |  |
| Lambda | 4.59 |  | 7.69 |  | 3.19 |  |
| Nebius | 4.5 | 5.4 | 8.5 | 9.5 |  | 1.55 |
| RunPod | 4.39 | 5.29 | 7.99 | 8.99 | 2.19 | 1.09 |
| Together AI | 3.99 | 5.99 | 9.29 | 11.49 |  |  |

## Price history, on demand, $ per GPU-hour

| Series | Dated prices |
|---|---|
| H100, Amazon | 2023.5: $12.29; 2025.42: $6.88 |
| H100, CoreWeave | 2023.25: $6.20; 2025: $6.16 |
| H100, Lambda | 2023.75: $2.59; 2024: $3.49; 2024.75: $2.99; 2026.25: $3.99; 2026.89: $4.59 |
| A100 80GB, CoreWeave | 2022.75: $3.65; 2025: $2.70 |
| A100 40GB, Lambda | 2022.5: $1.10; 2024: $1.29; 2026.25: $1.99 |
| V100, Lambda | 2021: $1.50; 2022.5: $0.55; 2026.25: $0.79 |

## Three listed GPU clouds, quarter ended September 30, 2026

| Company | Rent, % of equipment a year | Cash costs, % of equipment | Debt, % of equipment | Average cost of debt (%) | GPU-backed loan rate (%) | Basis |
|---|---|---|---|---|---|---|
| CoreWeave | 36.2664 | 14.016 | 75 | 7.9 | 9.4 | Rent against technology equipment at cost |
| Nebius | 24.2 | 12.8 | 73 | 4 |  | Rent against all property, net: a floor |
| IREN | 15 | 10.9 | 77 | 3.5 | 10.6 | GPU rent against equipment and buildings in service, at cost |

## What the model does not do

- The whole company is squeezed into one average GPU. The mix of old and new machines is hidden.
- Rent is measured against equipment that grew by $13.2 billion in one quarter, so the 36% ratio lags a fleet still being switched on.
- One loan stands in for 19 debt tranches with different rates and maturities.
- Customer prepayments, which bring cash in early, are left out.
- A retired GPU is assumed to be worth nothing.
- The price sheet and filing figures were read by an automated page reader and have not all been checked by eye.
- The role-play tab's opening values are a prediction, not sourced figures.

## How to cite

Erik Bethke, "CoreWeave: A pro forma", version 1.0, 2026-11-21, https://erikbethke.com.
