chip_cost | $36,586 | USD | derived; proposed | S-1 p.F-29, p.2 | Technology equipment, gross 9,146,575 thousand divided by more than 250,000 GPUs at Dec 31, 2024. A fleet average that includes servers and networking and excludes construction in progress. Compare with the model's capex per MW added (Outputs sheet), which is the same idea measured from cash flows. |
|---|
power_draw_kw | 1.44 | kW | derived; proposed | S-1 p.2 | More than 360 MW of active power divided by more than 250,000 GPUs at Dec 31, 2024. Facility-level power per GPU, so cooling and non-GPU load are already inside it. This is the ONE assumption that converts disclosed MW into GPU counts; every per-GPU line scales with it, the per-MW lines do not use it. |
|---|
pue | 1 | ratio | judgment; proposed | | Not used by the CoreWeave model: the per-GPU lines pin PUE at 1.0 in companies/coreweave.py because power_draw_kw is measured at the facility meter, so cooling is already inside it. Kept so the engine's inputs are complete. |
|---|
electricity_price_kwh | $0 | USD/kWh | judgment; proposed | | Not used by the CoreWeave model: pinned at 0 in companies/coreweave.py because each quarter's cash cost per GPU-hour is measured from revenue less adjusted EBITDA, which already contains energy. Kept so the engine's inputs are complete. |
|---|
utilization | 100.0% | share | judgment; proposed | | Not used by the CoreWeave model: pinned at 1.0 in companies/coreweave.py because revenue is measured per INSTALLED GPU-hour (quarterly revenue over average installed capacity), so idle and not-yet-billing GPUs are already inside that figure. |
|---|
tokens_per_sec | 0 | tokens/s | judgment; proposed | | Not applicable. CoreWeave sells GPU-hours and not tokens (S-1 p.95). |
|---|
price_per_m_tokens | $0 | USD/M tokens | judgment; proposed | | Not applicable. See tokens_per_sec. |
|---|
depreciation_years | 6 | years | disclosed; proposed; range 4 to 6 | S-1 p.F-16, p.F-15 | Accounting life of technology equipment. It was raised from five to six years on Jan 1, 2023. The economic life may be shorter if new GPU generations push rental prices down (risk factors p.33, p.44), which is why the low end is 4. |
|---|
financing_rate | 11.0% | decimal | disclosed; proposed; range 0.105 to 0.15 | S-1 p.F-34, p.F-41, p.F-39 | Cost of money applied to the GPUs already bought, for the per-GPU capital charge: the effective rate on DDTL 2.0, the facility that funded the 2024-25 fleet (S-1). Low is the actual average rate charged on DDTL 2.0 (10.53%); high is the effective rate on the older DDTL 1.0 (15%). Debt only; equity would want more. The forecast's interest line uses cost_of_debt instead, the rate on the marginal dollar, because the two questions differ: what did the installed fleet cost to finance, and what will the next GPU cost to finance. |
|---|
price_per_gpu_hour | $0 | USD/GPU-hour | judgment; proposed | | Not used: the model computes revenue per installed GPU-hour for each quarter from reported revenue and disclosed capacity (see the Outputs sheet). Kept at 0 so the engine's rental mode has a defined input. |
|---|
facility_cost_per_kw_month | $0 | USD/kW-month | judgment; proposed | | Not used by the CoreWeave model: pinned at 0 in companies/coreweave.py because data-centre leases are inside the cash cost measured from revenue less adjusted EBITDA. |
|---|
other_opex_per_gpu_hour | $0 | USD/GPU-hour | judgment; proposed | | Not used: each quarter's cash cost per GPU-hour is revenue less adjusted EBITDA over installed GPU-hours, computed by the model. Kept at 0. |
|---|
idle_power_share | 100.0% | share | judgment; proposed | | No effect while electricity_price_kwh is zero. |
|---|
residual_value_share | 0.0% | share | judgment; proposed; range 0 to 0.2 | | Cautious default. The S-1 gives no resale or redeployment value for retired GPUs. |
|---|
disclosed_cash_payback_years | 2.50 | years | disclosed; proposed | S-1 p.97 | Not an input. The company's own average cash payback on committed contracts at Dec 31, 2024, measured per GPU through adjusted EBITDA and net of customer prepayments. Shown beside the model's payback as a check. |
|---|
prepayment_share_of_tcv | 10.0% | share | derived; proposed; range 0.07 to 0.25 | S-1 p.96; 10-Q FY2026 Q2 and 10-K FY2025 cash flow statements (XBRL) with the contracted-power additions in the earnings releases | Two readings disagree. The S-1 says the weighted-average prepayment across active contracts was 15% to 25% of contract value at Dec 31 2024. The cash evidence since then is lower: prepayments received (change in deferred revenue plus prepaid revenue recognised) were 6,873 USD m over the six quarters to June 2026 against 2,400 MW of contracted power added, about 2.9 USD m per contracted MW, which is 8% of the 36 USD m of contract value per MW the model assumes (9 USD m a year for four years). The larger 2025-26 contracts appear to carry smaller prepayments, or the cash arrives later than signing. 10% leans on the cash evidence; the range spans both readings. Prepayments are generally credited against the final months of a contract, in some cases the first months after go-live. |
|---|
contract_years | 4 | years | disclosed; proposed; range 2 to 5 | S-1 p.96, p.152 | Committed contracts run two to five years; weighted-average duration was approximately four years at Dec 31, 2024. |
|---|
prepaid_revenue_recognised_share_per_quarter | 3.0% | share | derived; proposed; range 0.01 to 0.06 | 10-Q FY2026 Q2 (XBRL ContractWithCustomerLiabilityRevenueRecognized, ContractWithCustomerLiability) | Prepaid revenue recognised in a quarter as a share of the opening deferred-revenue balance: 1.2%, 2.3%, 3.7% and 3.3% over the four quarters to June 2026. Low because prepayments are credited in the final months of four-year contracts, so most of the balance has not started to unwind. A flat share understates recognition later in the forecast: the 4.1 USD bn balance of Dec 2024 is due to be credited during 2027-28, inside the window, which a flat 3% does not capture. A vintage-by-vintage schedule is the next refinement; the high end of the range approximates it. |
|---|
mw_added_per_quarter | 350 | MW | derived; proposed; range 200 to 600 | 8-K EX-99.1 2025-05-14 to 2026-08-11; 8-K EX-99.1 2026-05-07 (8 GW by 2030) | Active power going live per quarter. It was 60, 50, 120, 260, 150 and 500 MW over the six quarters to June 2026 (average of the last four: 258; last two: 325). Reaching the more than 8 GW by 2030 the company describes from 1.5 GW at June 2026 needs about 360 MW a quarter. Held flat across the forecast. Capacity is drawn from the contracted-but-not-active pipeline (2.2 GW at June 2026); the Drivers sheet shows that pipeline so a plan that outruns it is visible. |
|---|
mw_contracted_per_quarter | 400 | MW | derived; proposed; range 200 to 800 | 8-K EX-99.1 2025-05-14 to 2026-08-11 (total contracted power) | Contracted power signed per quarter: 300, 600, 700, 200, 400 and 200 MW over the six quarters to June 2026 (average 400). Drives the contracts signed, the backlog and the prepayments in the forecast; active power going live is a separate assumption and draws from the resulting pipeline. Held flat. |
|---|
revenue_per_mw_year_usd_m | $9,000,000 | USD m | derived; proposed; range 7 to 11 | Outputs sheet (revenue / average active power, 2025Q3 to 2026Q2) | Annualised revenue per MW of average active power was 10.3, 8.7, 9.0 and 8.2 USD m over the four quarters to June 2026 (mean 9.1). New capacity earns less per MW while it fills, so the trend is down; 9 assumes the fill lag stays as it is. |
|---|
adjusted_ebitda_margin | 58.0% | share | derived; proposed; range 0.52 to 0.65 | 8-K EX-99.1 2025-11-10 to 2026-08-11 | Adjusted EBITDA margin was 61.4%, 57.1%, 55.7% and 58.6% over the four quarters to June 2026 (mean 58.2%). Held flat. |
|---|
capex_per_mw_usd_m | $22,000,000 | USD m | derived; proposed; range 15 to 35 | Outputs sheet (cumulative capex / cumulative MW added since 2025Q1) | Cumulative capital expenditure per MW of active power added since 2025Q1 was 21.4 USD m at June 2026 (single quarters range from 13 to 51 because spending lands before the capacity it buys goes live). Applied to the MW added in the same quarter; a spend-to-live lag is not modelled. |
|---|
debt_share_of_capex | 80.0% | share | derived; proposed; range 0.5 to 0.9 | 10-Q FY2026 Q2 (XBRL ProceedsFromIssuanceOfLongTermDebt, RepaymentsOfLongTermDebt, PaymentsToAcquirePropertyPlantAndEquipment) | NET debt raised (proceeds less repayments) as a share of capital expenditure: 82% in both FY2025 (8,430 / 10,309 USD m) and H1 2026 (11,528 / 14,117). Gross proceeds were 115% and 119% of capex over the same periods because maturities were refinanced. The model applies the share net: gross draws are the net share of capex plus the scheduled repayments from the maturity ladder, so debt principal grows by 80% of capex. The DDTL facilities lend against the depreciated cost of GPUs backed by signed contracts, so the share holds while contracts keep coming. |
|---|
cost_of_debt | 7.5% | decimal | derived; proposed; range 0.065 to 0.1 | 10-Q FY2026 Q2 p.27 to p.28 | Interest expensed on the average debt balance, as a share of that balance: 558 USD m in Q2 2026 annualised over average principal of 30.4 USD bn is 7.4%, and the quarterly figure fell from 10.3% to 7.4% between Q2 2025 and Q2 2026 as cheaper facilities were added. The contractual rates on the June 2026 stack run from 7% (DDTL 4.0) to 15% (DDTL 1.0), with the 2026 facilities and notes at 9% to 10% (10-Q p.27 to p.28). The gap between contractual and expensed interest is interest capitalised into construction, which is already inside capex_per_mw_usd_m, so the expensed rate is the one that pairs with that capex figure without double counting. financing_rate, above, is the historical cost applied to the per-GPU capital charge. |
|---|
receivables_share_of_quarterly_revenue | 1 | ratio | derived; proposed; range 0.8 to 2 | 10-Q FY2026 Q2 and 10-K FY2025 balance sheets (XBRL AccountsReceivableNetCurrent) | Receivables at quarter end as a multiple of that quarter's revenue: 1.07, 1.60, 1.22, 2.02, 1.02 and 0.99 over the six quarters to June 2026. Receivables grew by 2,124 USD m over the period, most of the gap between reported cash from operations and adjusted EBITDA; invoiced prepayments sit here until they are paid. In the forecast the receivables balance grows with revenue at this multiple, so the cash drag fades as growth slows. |
|---|
other_operating_cash_share_of_revenue | -3.0% | share | derived; proposed; range -0.1 to 0 | 10-Q FY2026 Q2 and 10-K FY2025 cash flow statements (XBRL) | What reported cash from operations leaves unexplained after adjusted EBITDA, interest on debt, the change in deferred revenue and the receivables build: -324 USD m over the six quarters to June 2026, -3% of revenue (taxes paid, other working capital). Before the receivables line was separated, this residual was -2,448 USD m or -25% of revenue; receivables growth accounts for most of it. Single quarters swing widely, so only the multi-quarter figure means anything. |
|---|
minimum_cash_usd_bn | $2 | USD bn | judgment; proposed; range 1 to 4 | | Cash the company is assumed to keep on hand. Quarter-end cash was between 1.2 and 5.5 USD bn over the six quarters to June 2026 (1.4 at Dec 2024). When the forecast cash balance would fall below this, the model shows the shortfall as external funding required: equity-like money that bears no interest and is not repaid, so it does not enter debt or leverage. |
|---|