168.7× power→333.8× compute
1.98× compute/MWAll three are indexed from the same global quarterly series.
DECISION VIEW / C × E
Compute, power and the energy envelope
Connects estimated global accelerator capacity to nominal board power and IEA context, with all assumptions visible and adjustable.
Compute & energy efficiency
Theoretical compute per nominal MW1,566.5 H100e/MW2025 Q4
Equivalent dense 8-bit capacity3.1 TOPS/W
01,0002,000
2022 Q12025 Q4
Annual electricity model
Sensitivity model84.3 TWh / year17.4% of 2025 all-data-centre electricity (485 TWh)
Model, not measured consumption: 13.36 GW nominal power × 8,760 h × utilisation × facility overhead. The comparison is not an observed AI electricity share. Useful work per kWh remains unmeasured.
Analysis summary Assessed 2025-12-31
INTERPRETATION
1.98×
Theoretical compute per nominal MWIncrease from 791.7 to 1,566.5 H100e/MW since 2022-Q1.
Use this view to distinguish capacity growth from theoretical efficiency gains, and to stress-test energy implications without turning a scenario into an observation.
117.0 TWh→485 TWh
24.1%Nameplate sensitivity against all data centre electricity, not a measured AI share.
What drives the growth?
Compute holdings increased 333.8× while nominal accelerator power increased 168.7×.
Compute, power and efficiency since 2022-Q1 · Theoretical compute / nominal board power / compute per nominal MW.
Does the fleet provide more theoretical compute per watt?
Yes, in this peak-performance proxy: 1.98× H100e per nominal MW since 2022-Q1.
Theoretical fleet efficiency · 3.10 theoretical dense INT8 TOPS/W; nominal board power, not measured useful work.
What does today's nameplate power mean on an annual basis?
117.0 TWh/year at 100% and PUE 1.0; use the calculator for explicit assumptions.
Annualised nameplate envelope · Nominal accelerator power × 8,760 hours at 100% utilisation and PUE 1.0; not measured consumption.
Nameplate envelope versus all data centre electricity · Sensitivity ratio against the IEA's 485 TWh/year for all data centres in 2025.
How much useful AI work do we get per kWh?
Not established. This requires workload-specific wall-power and performance measurements.
Theoretical fleet efficiency · 3.10 theoretical dense INT8 TOPS/W; nominal board power, not measured useful work.
- Retrieved
- 2026-08-10
- Data
- 2024-12-31
- Licence
- CC BY 4.0
Used for energy context, not to derive the H100-equivalent compute series.
Open original source ↗- Retrieved
- 2026-08-12
- Data
- 2025-12-31
- Licence
- CC BY 4.0
Updated IEA observation for 2025 and central 2030 projection. Total data-centre electricity, not AI-only demand.
Open original source ↗- Retrieved
- 2026-08-10
- Data
- 2025-12-31
- Licence
- CC BY 4.0
Ownership estimates, not physical location, utilisation, or access.
Open original source ↗- Retrieved
- 2026-08-11
- Data
- 2026-08-11
- Licence
- CC BY 4.0
Defines H100e as peak dense 8-bit operations per second relative to an Nvidia H100; it is a comparison model, not a chip count.
Open original source ↗VERIFIABLE EVIDENCE / 5 CLAIMSOpen evidence +
Claims, calculations, sources and known limitations for this view. Source data and dashboard interpretation are kept separate.
Theoretical fleet efficiency increased from 791.7 to 1,566.5 H100e per nominal MW from 2022-Q1 to 2025-Q4, approximately 1.98 times.
CALCULATION
efficiency = H100e median / nominal accelerator power MWEpoch AI: Data on AI Chip Owners ↗Epoch AI: What does H100e compute capacity mean? ↗16 measurements32 source facts
- H100e is a theoretical dense 8-bit comparison model.
- Nominal board power is not actual electricity consumption, and the result is not useful work per joule.
compute-energy/claim/efficiency-trajectoryFrom 2022-Q1 to 2025-Q4, estimated global compute holdings increased approximately 333.8 times, nominal accelerator power 168.7 times and theoretical compute per nominal MW 1.98 times.
CALCULATION
latest / first for compute, nominal MW and (H100e/MW)Epoch AI: Data on AI Chip Owners ↗Epoch AI: What does H100e compute capacity mean? ↗1 measurement4 source facts
- The scaling index describes theoretical installed/estimated accelerator capacity, not actual AI work.
compute-energy/claim/scale-versus-efficiencyThe world series' nominal accelerator power in 2025-Q4 is equivalent to 117.0 TWh per year if nameplate power is sustained continuously at PUE 1.0.
CALCULATION
13359.1 MW × 8760 h / 1,000,000 = 117.0257 TWh/year- This is an upper nameplate sensitivity, not measured energy use.
- It excludes CPU, memory, networking and cooling, and does not automatically account for PUE above 1.0.
compute-energy/claim/nameplate-energy-envelopeThe nameplate envelope of 117.0 TWh/year is 24.1 percent of the IEA's estimate of 485 TWh for all data centre electricity in 2025.
CALCULATION
117.0257 / 485 × 100 = 24.129%Epoch AI: Data on AI Chip Owners ↗International Energy Agency: Key Questions on Energy and AI ↗2 measurements2 source facts
- The denominator includes all data centres; the numerator includes only nominal AI accelerator power.
- The ratio is an order-of-magnitude sensitivity, not an observed share.
compute-energy/claim/data-center-contextCompute per nominal watt is a theoretical hardware proxy. Useful AI work per kWh requires workload, precision, system and wall-power measurements that this module does not yet contain.
Epoch AI: Data on AI Chip Owners ↗Epoch AI: What does H100e compute capacity mean? ↗1 measurement2 source facts
- Do not use the curve as a benchmark for model quality or actual energy efficiency.
compute-energy/claim/model-boundaryTRACEABLE FACT LOG / 33 SOURCE FACTSOpen log +
compute-energy/fact/compute-stock/2022-03-31Global estimated AI accelerator stock
Epoch AI · evidence-catalog:evidence/compute/world/2022-03-31compute-energy/fact/nominal-power/2022-03-31Global nominal accelerator board power
Epoch AI · evidence-catalog:evidence/compute/world/2022-03-31compute-energy/fact/compute-stock/2022-06-30Global estimated AI accelerator stock
Epoch AI · evidence-catalog:evidence/compute/world/2022-06-30compute-energy/fact/nominal-power/2022-06-30Global nominal accelerator board power
Epoch AI · evidence-catalog:evidence/compute/world/2022-06-30compute-energy/fact/compute-stock/2022-09-30Global estimated AI accelerator stock
Epoch AI · evidence-catalog:evidence/compute/world/2022-09-30compute-energy/fact/nominal-power/2022-09-30Global nominal accelerator board power
Epoch AI · evidence-catalog:evidence/compute/world/2022-09-30compute-energy/fact/compute-stock/2022-12-31Global estimated AI accelerator stock
Epoch AI · evidence-catalog:evidence/compute/world/2022-12-31compute-energy/fact/nominal-power/2022-12-31Global nominal accelerator board power
Epoch AI · evidence-catalog:evidence/compute/world/2022-12-31compute-energy/fact/compute-stock/2023-03-31Global estimated AI accelerator stock
Epoch AI · evidence-catalog:evidence/compute/world/2023-03-31compute-energy/fact/nominal-power/2023-03-31Global nominal accelerator board power
Epoch AI · evidence-catalog:evidence/compute/world/2023-03-31compute-energy/fact/compute-stock/2023-06-30Global estimated AI accelerator stock
Epoch AI · evidence-catalog:evidence/compute/world/2023-06-30compute-energy/fact/nominal-power/2023-06-30Global nominal accelerator board power
Epoch AI · evidence-catalog:evidence/compute/world/2023-06-30compute-energy/fact/compute-stock/2023-09-30Global estimated AI accelerator stock
Epoch AI · evidence-catalog:evidence/compute/world/2023-09-30compute-energy/fact/nominal-power/2023-09-30Global nominal accelerator board power
Epoch AI · evidence-catalog:evidence/compute/world/2023-09-30compute-energy/fact/compute-stock/2023-12-31Global estimated AI accelerator stock
Epoch AI · evidence-catalog:evidence/compute/world/2023-12-31compute-energy/fact/nominal-power/2023-12-31Global nominal accelerator board power
Epoch AI · evidence-catalog:evidence/compute/world/2023-12-31compute-energy/fact/compute-stock/2024-03-31Global estimated AI accelerator stock
Epoch AI · evidence-catalog:evidence/compute/world/2024-03-31compute-energy/fact/nominal-power/2024-03-31Global nominal accelerator board power
Epoch AI · evidence-catalog:evidence/compute/world/2024-03-31compute-energy/fact/compute-stock/2024-06-30Global estimated AI accelerator stock
Epoch AI · evidence-catalog:evidence/compute/world/2024-06-30compute-energy/fact/nominal-power/2024-06-30Global nominal accelerator board power
Epoch AI · evidence-catalog:evidence/compute/world/2024-06-30compute-energy/fact/compute-stock/2024-09-30Global estimated AI accelerator stock
Epoch AI · evidence-catalog:evidence/compute/world/2024-09-30compute-energy/fact/nominal-power/2024-09-30Global nominal accelerator board power
Epoch AI · evidence-catalog:evidence/compute/world/2024-09-30compute-energy/fact/compute-stock/2024-12-31Global estimated AI accelerator stock
Epoch AI · evidence-catalog:evidence/compute/world/2024-12-31compute-energy/fact/nominal-power/2024-12-31Global nominal accelerator board power
Epoch AI · evidence-catalog:evidence/compute/world/2024-12-31compute-energy/fact/compute-stock/2025-03-31Global estimated AI accelerator stock
Epoch AI · evidence-catalog:evidence/compute/world/2025-03-31compute-energy/fact/nominal-power/2025-03-31Global nominal accelerator board power
Epoch AI · evidence-catalog:evidence/compute/world/2025-03-31compute-energy/fact/compute-stock/2025-06-30Global estimated AI accelerator stock
Epoch AI · evidence-catalog:evidence/compute/world/2025-06-30compute-energy/fact/nominal-power/2025-06-30Global nominal accelerator board power
Epoch AI · evidence-catalog:evidence/compute/world/2025-06-30compute-energy/fact/compute-stock/2025-09-30Global estimated AI accelerator stock
Epoch AI · evidence-catalog:evidence/compute/world/2025-09-30compute-energy/fact/nominal-power/2025-09-30Global nominal accelerator board power
Epoch AI · evidence-catalog:evidence/compute/world/2025-09-30compute-energy/fact/compute-stock/2025-12-31Global estimated AI accelerator stock
Epoch AI · evidence-catalog:evidence/compute/world/2025-12-31compute-energy/fact/nominal-power/2025-12-31Global nominal accelerator board power
Epoch AI · evidence-catalog:evidence/compute/world/2025-12-31compute-energy/fact/data-center-electricity/2025-12-31Global data-centre electricity
International Energy Agency · evidence-catalog:evidence/energy/data-centers/2025-electricityINTERPRETATION LIMIT
The Compute series estimates ownership and nominal accelerator power, not physical location or use.
H100e and TOPS/W describe theoretical dense 8-bit peak performance, not useful model work per joule.
The IEA figure covers all data centres; regional energy production and grid access have not yet been normalised here.
MODULE STATUSWORKING
1.98× theoretical compute/MW since 2022-Q1; 117.0 TWh/year is only a nameplate sensitivity