What Is One Megawatt of AI Worth?
1 GW: investment vs SemiAnalysis lab revenue vs token math. Estimates, filings, and engine are labeled. The model never computes.
Capex to stand up 1 GW vs lab API revenue vs token math
Default capex $12M/MW times 1 GW (Dylan mid of $10-15M, all-in GPU + facility). SemiAnalysis lab range is $50M-$100M/MW, not the token engine. Rent is not added.
Labs at this scale
SemiAnalysis density for OpenAI and Anthropic. SPCX is the Q2 2026 filing, implied from 1.4 GW nameplate, not Dylan $100M/MW, not annualized. Type a $/MW to try an estimate. Hyperscalers and NVDA omitted (no AI $/MW split in this model).
| Name | Kind | At this scale | $/MW | Your $/MW |
|---|---|---|---|---|
| OpenAI | estimate | $50B-$100B / year | $50M-$100M | |
| Anthropic | estimate | $50B-$100B / year | $50M-$100M | |
| SPCX | filingQ2 2026, 8-K Exhibit 99.1 / Q2 10-Q, filed 4 Aug 2026 | $1.8B Q2 revenue · $11B Q2 capex | $1.8M rev · $11M capex |
The $11T path, 2024-2029estimate
$11TCumulative AI capex Dylan called "about $11 trillion." Funded $6T cash and $5T credit. SemiAnalysis prints $11.1T for the same window. Not a filing.
SemiAnalysis July 2026. Stock of debt, not the $5T of new credit in the stack above. Do not add $7.1T into $5T.
GW added (years he named)
Annual dollars he named
Only 2026 and 2028. No bar for 2024, 2025, or 2027. Those years were not given as a dollar figure.
a little bit over a trillion dollars of CapEx
more than $2 trillion
All-in callout $3T-$4T ($2.5T+ IT + $1-2T data center and energy).
Separate from the $11T cumulative: at 100 GW/year, IT about $5T; with plants and shells built ahead $7T-$10T that year. 2030-ish run-rate thought, not the 2024-2029 stack.
Dwarkesh Patel interview, 25 Aug 2026 (aV26V1UvkJw). SemiAnalysis "Nvidia GPU Debt Backstop", July 2026. Estimate, not a filing.
Sensitivity (revenue/MW/yr at current throughput)
| $/M \ util | 50% | 70% | 90% |
|---|---|---|---|
| $2.48/M | $109M/MW | $153M/MW | $197M/MW |
| $4.95/M (now) | $219M/MW | $306M/MW | $393M/MW |
| $9.90/M | $437M/MW | $612M/MW | $787M/MW |
Heat map: revenue/MW/yr
Rows are serving hours (20-100%). Columns are blended $/M tokens ($0.10-$10). Ring marks cells near the current annualized owned cost contour.
| util \ $/M | 0.10 | 0.25 | 0.50 | 1 | 2 | 3 | 5 | 7.5 | 10 |
|---|---|---|---|---|---|---|---|---|---|
| 20% | $1.8M | $4.4M | $8.8M | $18M | $35M | $53M | $88M | $132M | $177M |
| 30% | $2.6M | $6.6M | $13M | $26M | $53M | $79M | $132M | $199M | $265M |
| 40% | $3.5M | $8.8M | $18M | $35M | $71M | $106M | $177M | $265M | $353M |
| 50% | $4.4M | $11M | $22M | $44M | $88M | $132M | $221M | $331M | $442M |
| 60% | $5.3M | $13M | $26M | $53M | $106M | $159M | $265M | $397M | $530M |
| 70% | $6.2M | $15M | $31M | $62M | $124M | $185M | $309M | $464M | $618M |
| 80% | $7.1M | $18M | $35M | $71M | $141M | $212M | $353M | $530M | $706M |
| 90% | $7.9M | $20M | $40M | $79M | $159M | $238M | $397M | $596M | $795M |
| 100% | $8.8M | $22M | $44M | $88M | $177M | $265M | $442M | $662M | $883M |
Sources
- [1] tokensPerSecondPerMW (Base, Bull): 2.8M tokens/s per MW. SemiAnalysis InferenceMAX, 9 Oct 2025 (nightly 7 Oct). Hardware: B200. MW basis: all-in utility MW (not Critical IT). Workload: gpt-oss 120B FP4, 1k in / 8k out, 90 tok/s/user, cache OFF, processed in+out. Token mix: uncached input 11.1%, cached 0%, output 88.9% (matches 2.8M; not an agentic 90% cache mix). Limits: cache OFF. Do not pair this throughput with a cached/agentic mix.
- [2] tokensPerSecondPerMW (Bear): 0.9M tokens/s per MW. SemiAnalysis InferenceMAX (same snapshot), 9 Oct 2025 (nightly 7 Oct). Hardware: H100. MW basis: all-in utility MW (not Critical IT). Workload: same InferenceMAX snapshot as Base, H100 rather than B200. Bear tightens via throughput, utilization, and rent floor. Same DeepSeek mix as Base.
- [3] utilization: 0.70 serving hours (Base, Bull); 0.50 (Bear). Epoch TCO (~71%, rounded to 70% serving hours), 14 May 2026. 75% is NOT a SemiAnalysis number. Do not label 70% as InferenceMAX.
- [4] token mix and prices (Base, Bear): $1.35 / $5.40 per million (in / out), blended from mix (~$4.95). InferenceX v2, DeepSeek-class OpenRouter Crusoe list, 16 Feb 2026. Token mix: uncached input 11.1% at $1.35, cached 0%, output 88.9% at $5.40. Do not default $1.50. Blended = 0.111 * 1.35 + 0.889 * 5.40 (engine mix math, ~$4.95/M).
- [5] token prices (Bull): $5 / $25 per million (in / out), blended $22.78. Opus list on the same 11.1 / 0 / 88.9 mix, v1 Bull preset. Token mix: uncached input 11.1% at $5, cached 0%, output 88.9% at $25. Bull changes PRICE only. Throughput stays 2.8M (B200). Do not raise 2.8M for GB200 or Rubin.
- [6] annualRentPerMW: $12M/MW/yr Base and Bull (mid of $10-15M); $10M/MW/yr Bear. Dylan Patel / Dwarkesh, all-in GPU + facility rent, 25 Aug 2026. MW basis: all-in utility MW. Rented mode uses this as annual cost. Owned mode uses capex, not rent. Do not add PUE on top of 2.8M. Do not count annual rent and full capex in the same scenario.
- [7] capexPerMW: $12M per MW (mid of $10-15M). 1 GW capex = $12B.. Dylan Patel / Dwarkesh, all-in GPU + facility, 25 Aug 2026. MW basis: all-in utility MW. Hero payback is capex / annual token revenue. Annualized CRF cost is only for owned break-even, not the picture.
- [8] capacityMW: 1 GW default (1000 MW). Query mw=1 keeps the old 1 MW link.. model identity (one megawatt), scaled, v2. Revenue and investment scale linearly. GW is 1000 MW. Engine stays per-MW then times capacity.
- [9] SA lab API revenue: $50M-$100M per MW ($50B-$100B per GW). Anthropic cited as high as $50M/MW; path to $100M/MW; 1 GW produces $100B of revenue.. SemiAnalysis / Dylan Patel (Dwarkesh 25 Aug 2026; Fireside Alpha 11 Aug 2026), 25 Aug 2026. MW basis: all-in utility MW. Estimate, not a filing, not the InferenceMAX token engine. Do not replace this $100B with BASE token math (~$306B/GW).
- [10] SPCX AI segment: Q2 2026 AI revenue $2.561B, AI capex $15.828B, nameplate 1.4 GW.. 8-K Exhibit 99.1 / Q2 10-Q, 4 Aug 2026. Filing. Implied $/MW is revenue or capex divided by 1.4 GW nameplate, not Dylan $100M/MW. Quarter is not annualized. Hyperscalers and NVDA omitted: no AI $/MW split already in this model.
- [11] AI capex path 2024-2029: About $11T cumulative ($6T cash, $5T credit). SA note $11.1T. 2026 a little over $1T. 2028 more than $2T IT, $3-4T all-in. GW adds 30 / 50 / 70 / 90-100.. Dwarkesh 25 Aug 2026 (aV26V1UvkJw); SemiAnalysis Nvidia GPU Debt Backstop, July-Aug 2026. Estimate, not a filing. $7.1T is 2029 debt stock, not the $5T of new credit. 100 GW/year $5T-$10T is a 2030-ish run-rate, not the $11T cumulative. No yearly dollar bars for years he did not name.
Throughput 2.8M/s/MW is not multiplied by PUE. Hero uses $12M/MW capex, not rent.