General assumptions
Shared by every business in the chain.
Applied to all four businesses.
GPU capacity per MW
About 14.3kW maximum system power for an 8-GPU HGX B200 server per NVIDIA.
PUE of 1.2 to 1.4 per Crusoe; portfolio PUE of 1.25 per NBIS.
Per HGX B200 server (Mercatus). $300–460k for an 8-GPU server when buying a 512-GPU rack.
Assumes 10% additional hardware and equipment.
8 GPUs per HGX B200 server.
Sources as of October 2026 and subject to change.
Capacity per megawatt
How many servers and GPUs one MW of power supports, and what they cost. Throughout, “$mn/MW” means $ millions per megawatt per year and “$/GPU-hr” spreads the same figure over 8,760 hours and the GPUs per MW here.
| per MW | |
|---|---|
| Theoretical servers per MW | 69.9 |
| Servers per MW after PUERounded down to whole servers. | 58 |
| GPUs per MW | 464 |
| Compute CapEx per MW | $28.7mn |
| IREN: $29mn per MW ($5.8bn of CapEx for 200MW). | |
Assumptions
Buy land, build and power the shell, sell capacity on a triple-net lease.
$11–13mn per MW per APLD, GLXY and others.
$2.4mn per MW per Riot (Aug 2026) and WULF (Jul 2026); $1.9mn per MW per HUT and CLSK (Jul 2026).
Triple-net lease, project level: APLD guidance 85%+, GLXY 90%+, CIFR 85%+.
30-year useful life; estimates range from 25 to 40 years.
70–90% loan-to-value is typical.
SOFR +200–300bps is typical with a tenant guarantee; outliers as high as +500bps (Galaxy Helios).
Sources as of October 2026 and subject to change.
Unit economics
Sells capacity to the compute owner. Per MW per year, and per GPU hour using the GPUs per MW from General Assumptions.
| $/GPU-hr | $mn/MW | |
|---|---|---|
| Revenue | $0.54 | $2.20 |
| EBITDA | $0.46 | $1.87 |
| EBITDA margin | 85% | |
| Less: depreciation | 0.10 | 0.40 |
| EBIT | $0.36 | $1.47 |
| EBIT margin | 67% | |
| Less: interest | 0.15 | 0.62 |
| Less: taxes | 0.04 | 0.17 |
| Net income | $0.17 | $0.68 |
| Unlevered payback (years)1 | 8.0 | |
| Levered payback (years)2 | 2.2 |
- 1Unlevered payback: CapEx ÷ after-tax EBITDA.
- 2Levered payback: equity share of CapEx ÷ (EBITDA − interest − taxes).
Assumptions
Rent land and power from the datacenter, buy GPUs, sell GPU hours. The neocloud.
IREN recently priced at $20mn per MW, with discussion at $25mn. NBIS: $35mn per MW implied by recent PAYG price increases. CRWV: a recent 3–6 month contract priced at $40mn per MW.
$0.0813 per kWh, 2025 national industrial average per EIA.
Assumes 2.5% for customer support and similar costs.
5 years per NBIS, 6 years per CRWV; evidence suggests the useful life may be longer.
5-year amortization, in line with the assumed useful life. Used for the levered payback only.
70% debt financed is typical per CRWV.
SOFR +400–600bps is typical; CRWV +450–550bps in recent deals.
Sources as of October 2026 and subject to change.
Unit economics
Buys capacity from the datacenter, sells GPU hours to the token factory.
| $/GPU-hr | $mn/MW | |
|---|---|---|
| Revenue | $4.92 | $20.0 |
| Less: datacenter leaseEquals the datacenter’s revenue. | 0.54 | 2.2 |
| Less: electricity | 0.18 | 0.7 |
| Less: other cost to service | 0.12 | 0.5 |
| EBITDA | $4.08 | $16.6 |
| EBITDA margin | 83% | |
| IREN’s MSFT deal: 85% at project level. | ||
| Less: depreciation | 1.41 | 5.7 |
| EBIT | $2.67 | $10.8 |
| EBIT margin | 54% | |
| Less: interest | 0.42 | 1.7 |
| Less: taxes | 0.45 | 1.8 |
| Net income | $1.80 | $7.3 |
| Unlevered payback (years)1 | 2.3 | |
| NBIS and IREN guide to roughly 2 years. | ||
| Levered payback (years)2 | 1.2 | |
- 1Unlevered payback: revenue ÷ after-tax EBIT.
- 2Levered payback: equity share of CapEx ÷ (EBITDA − interest − taxes − loan amortization).
Assumptions
Rent GPUs from the compute owner, serve a model, sell tokens. Kimi K3 on B200 as the example.
OpenRouter reported weighted-average API price (Kimi K3 as the example).
OpenRouter reported weighted-average API price.
Meta: roughly 1:10 output to input for summaries, 1:100 for coding.
SemiAnalysis: Kimi K3 sustains 4.6k tokens/s per GPU on B200 on its agentic coding workload at 50 tokens/s per user.
Meta: about 50% is “easy to reach”. Utilization becomes problematic around 65% as memory latency increases.
Assumes 5% for payments, customer support and similar costs.
Sources as of October 2026 and subject to change.
Unit economics
Rents GPUs from the compute owner, sells tokens.
| $/GPU-hr | $mn/MW | |
|---|---|---|
| Revenue | $10.88 | $44.2 |
| Less: GPU rentalPaid to the neocloud; equals its revenue. | 4.92 | 20.0 |
| Less: other cost to service | 0.54 | 2.2 |
| EBITDA | $5.42 | $22.0 |
| EBITDA margin | 50% | |
| OpenAI compute margin about 70% (Oct 2025, The Information). Anthropic inference margin about 65% (May 2026, SemiAnalysis). | ||
| Less: taxes | 1.08 | 4.4 |
| Net income | $4.33 | $17.6 |
Assumptions
The hyperscaler owns all three businesses, so every assumption feeds it. Change any of them here or on its own tab.
DatacenterCapEx $12.0mn, revenue $2.2mn, 85% EBITDA margin
$11–13mn per MW per APLD, GLXY and others.
$2.4mn per MW per Riot (Aug 2026) and WULF (Jul 2026); $1.9mn per MW per HUT and CLSK (Jul 2026).
Triple-net lease, project level: APLD guidance 85%+, GLXY 90%+, CIFR 85%+.
30-year useful life; estimates range from 25 to 40 years.
70–90% loan-to-value is typical.
SOFR +200–300bps is typical with a tenant guarantee; outliers as high as +500bps (Galaxy Helios).
Compute OwnerRevenue $20.0mn ($4.92/GPU-hr), CapEx $28.7mn
IREN recently priced at $20mn per MW, with discussion at $25mn. NBIS: $35mn per MW implied by recent PAYG price increases. CRWV: a recent 3–6 month contract priced at $40mn per MW.
$0.0813 per kWh, 2025 national industrial average per EIA.
Assumes 2.5% for customer support and similar costs.
5 years per NBIS, 6 years per CRWV; evidence suggests the useful life may be longer.
5-year amortization, in line with the assumed useful life. Used for the levered payback only.
70% debt financed is typical per CRWV.
SOFR +400–600bps is typical; CRWV +450–550bps in recent deals.
Token Factory$13.00 out and $0.50 in per M tokens, 4,600 tokens/s, 60% utilization
OpenRouter reported weighted-average API price (Kimi K3 as the example).
OpenRouter reported weighted-average API price.
Meta: roughly 1:10 output to input for summaries, 1:100 for coding.
SemiAnalysis: Kimi K3 sustains 4.6k tokens/s per GPU on B200 on its agentic coding workload at 50 tokens/s per user.
Meta: about 50% is “easy to reach”. Utilization becomes problematic around 65% as memory latency increases.
Assumes 5% for payments, customer support and similar costs.
Sources as of October 2026 and subject to change.
Unit economics
Datacenter, compute owner and token factory combined on one megawatt.
| $mn/MW | |
|---|---|
| CapExDatacenter plus compute. | $40.7 |
| RevenueToken factory revenue. | $44.2 |
| EBITDADatacenter, compute owner and token factory combined. | $40.5 |
| EBITDA margin | 92% |
| Less: depreciation | 6.1 |
| EBIT | $34.3 |
| EBIT margin | 78% |
| Less: interest | 2.3 |
| Less: taxes | 6.4 |
| Net income | $25.6 |
| Unlevered payback (years)1 | 1.5 |
| Levered payback (years)2 | 0.4 |
- 1Unlevered payback: CapEx ÷ after-tax EBIT.
- 2Levered payback: equity share of CapEx, at the compute owner’s debt share, ÷ (EBITDA − interest − taxes).
Every GW of energy capacity added is worth $44.2bn of revenue today.
Token factory revenue of $44.2mn per MW per year, across the 1,000 MW in a gigawatt. The datacenter lease ($2.2mn) and the GPU rental ($20.0mn) are paid out of it rather than added to it. The bar below shows where it ends up.
- Electricity $0.7mn
- Cost to service $3.0mn
- Depreciation $6.1mn
- Datacenter EBIT $1.5mn, 4% of EBIT
- Neocloud EBIT $10.8mn, 32% of EBIT
- Frontier lab EBITDA $22.0mn, 64% of EBIT
- CapEx per MW
- $40.7mn
- EBITDA
- $40.5mn
- EBIT margin
- 78%
- Levered payback, consolidated
- 0.7years
- GPUs per MW
- 464
| $mn/MW | Datacenter | Compute Owner | Token Factory | Hyperscaler | CombinedSum of the three businesses |
|---|---|---|---|---|---|
| CapEx | $12.0 | $28.7 | – | $40.7 | $40.7 |
| Revenue | $2.2 | $20.0 | $44.2 | $44.2 | $66.4 |
| EBITDA | $1.9 | $16.6 | $22.0 | $40.5 | $40.5 |
| EBITDA margin | 85% | 83% | 50% | 92% | 61% |
| Net income | $0.7 | $7.3 | $17.6 | $25.6 | $25.6 |
| Share of chain EBIT | 4% | 32% | 64% | 100% | 100% |
| Unlevered payback (years) | 8.0 | 2.3 | – | 1.5 | 1.5 |
| Levered payback (years) | 2.2 | 1.2 | – | 0.4 | 0.7 |
Combined adds the three businesses as separate companies, so its revenue counts the datacenter lease and the GPU rental that the hyperscaler internalizes. Its levered payback divides the debt raised by after-tax EBITDA; the other payback definitions are footnoted on each business’s tab.
Sensitivity: token prices against GPU rental
Rows scale both token prices from the current inputs; columns set the compute owner’s revenue per MW, which is the token factory’s rental cost. Everything else stays as set. The outlined cell is the current scenario.
| Compute owner revenue (GPU rental), $mn/MW | |||||||
|---|---|---|---|---|---|---|---|
| Token prices | $10.0$2.46/GPU-hr | $15.0$3.69/GPU-hr | $20.0$4.92/GPU-hr | $25.0$6.15/GPU-hr | $30.0$7.38/GPU-hr | $35.0$8.61/GPU-hr | $40.0$9.84/GPU-hr |
| −60% | 38% | 10% | (18%) | (46%) | (75%) | (103%) | (131%) |
| −40% | 57% | 38% | 20% | 1% | (18%) | (37%) | (56%) |
| −20% | 67% | 53% | 38% | 24% | 10% | (4%) | (18%) |
| Current | 72% | 61% | 50% | 38% | 27% | 16% | 5% |
| +20% | 76% | 67% | 57% | 48% | 38% | 29% | 20% |
| +50% | 80% | 72% | 65% | 57% | 50% | 42% | 35% |