4 Billion Years On

AI Industry Data

AI models released, the data centres that train them, the electricity they use and the money behind them…

Tracking the AI industry - the notable models released each year and who builds them, the frontier data centres being built around the world and their combined compute, how much electricity they draw against whole countries, benchmark results, and private investment in AI. Data from Epoch AI, the Stanford AI Index and Our World in Data.

Key Facts

Notable AI Models
214
111 in 2025 · 103 in 2026 (Epoch AI)
AI Investment
$290B
2025 (AI Index Report)
Frontier AI Sites
96
76 Operational · 20 Planned
AI Energy Demand
132.5 TWh
0.42% of world electricity
Data from Epoch AI and the Stanford AI IndexEpoch AI and the Stanford AI Index

Epoch AI (models & energy) and the Stanford AI Index (investment).

Infrastructure

Data Center MapFrontier AI Data Center Locations

Show

96 frontier AI data centers tracked globally (77 in the US). Pin size shows power capacity.

Data from Epoch AIEpoch AI

Frontier Data CentersFrontier AI Data Centers

Tracked Sites
96
76 operational · 20 planned
Power
15,122MW
Operational sites
H100 Equivalents
15.8M
GPU compute, operational
Capital Cost
$572.9B
Operational sites

H100 Equivalents - all the GPU compute counted as NVIDIA H100 chips, so different hardware can be compared.

Tap a row to see it on the map.

Amazon1,037$39.3
SpaceXAI946$35.8
Oracle824$31.2
Microsoft636$24.1
—473$17.9
Meta471$17.8
—422$16.0
Microsoft369$14.0
Google368$13.9
SpaceXAI340$12.9
Google333$12.6
Google303$11.5
Amazon284$10.8
CoreWeave262$9.9
ByteDance243$9.2
Google238$9.0
—238$9.0
Google237$9.0
Google237$9.0
Google237$9.0
Amazon228$8.6
ByteDance221$8.4
Amazon213$8.1
AI XPV Platform210$8.0
Microsoft202$7.7
Huawei197$7.5
Microsoft190$7.2
Google183$6.9
Meta180$6.8
—180$6.8
Meta178$6.7
Meta178$6.7
Meta178$6.7
Huawei170$6.4
Alibaba169$6.4
Google161$6.1
Google154$5.8
Meta153$5.8
Meta152$5.8
Meta152$5.8
Meta146$5.5
—144$5.5
Google141$5.3
Google137$5.2
Meta133$5.0
CoreWeave132$5.0
Meta129$4.9
—118$4.5
Meta118$4.5
—113$4.3
Amazon108$4.1
Google105$4.0
Google103$3.9
Google88$3.3
Meta87$3.3
—85$3.2
CoreWeave82$3.1
Meta80$3.0
Google77$2.9
Google77$2.9
Nebius75$2.8
Microsoft75$2.8
Oracle72$2.7
CoreWeave68$2.6
CoreWeave65$2.5
—64$2.4
Core4258$2.2
Nebius50$1.9
Microsoft47$1.8
CoreWeave42$1.6
Firmus41$1.6
Nscale33$1.3
—32$1.2
CoreWeave28$1.1
SpaceXAI17$0.7
Nscale5$0.2
Cipher Mining——
CoreWeave——
CoreWeave——
Microsoft——
Mistral AI——
Google——
Google——
Google——
Meta——
Meta——
Nscale——
Softbank——
Oracle——
Softbank——
Oracle——
Oracle——
G42——
Oracle——
Meta——
Meta——

The world's largest known AI data centers, tracked from satellite images, building permits and public statements.

Data from Epoch AIEpoch AI

Epoch AI - Frontier Data Centers (CC BY). Updated Nov 2026.

AI ComputeAI Frontier Data Center Processing Power

Compute
15,595ExaFLOPS
FP16 · operational today
Per Person
1.90TFLOPS
Shared among 8.2B people
Phones Per Person
0.95
Smartphone equivalent
Vs All Phones
159.1%
Of ~4.9B smartphones' compute

Frontier Compute (ExaFLOPS, FP16)

After the Today line: sites planned or under construction, as Epoch AI projects them.

Compute per Person vs Devices (TFLOPS)

Smartwatch
0.10
Budget phone
0.50
AI per person
1.98.0 by 2029
Flagship phone
2.0
Laptop (M3 Pro)
7.0
PS5 console
10.3
Operational todayWith planned sites, 2029
Data from Epoch AIEpoch AI

Epoch AI (CC BY). People ~8.2B (UN 2025), smartphones ~4.9B (GSMA 2025).

Frontier data centers only, not AI on phones or laptops. TFLOPS - trillion operations a second (FP16). An H100 does 989. The figures and device comparison are today's operational capacity. The dotted bars and the chart show planned sites, taking compute to about 4.2 times today's by 2029.

Devices are approximate GPU peaks (flagship phone - A18 / Snapdragon 8 Gen 3 class, PS5 - 10.3 TFLOPS).

AI Electricity UseAI Share of World Electricity

World Electricity
0.42%
Used by frontier AI today
AI Demand
132.5TWh
Today's 15,122 MW, run all year

Frontier Data Centre Power

Power capacity across all tracked frontier data centres; after the Today line, Epoch AI's projection for planned sites.

AI vs Countries’ Electricity (TWh a year)

Philippines
123 TWh
Kazakhstan
124 TWh
⚡ Frontier AI
132 TWh440 by 2029
Netherlands
135 TWh
Iraq
150 TWh
United Kingdom
292 TWh
Iran
396 TWh
Saudi Arabia
455 TWh
Operational todayWith planned sites, 2029

If every planned site is built, frontier AI would use about 440 TWh a year by 2029.

Data from Epoch AI and Our World in DataEpoch AI and Our World in Data

Epoch AI (data centers) and Our World in Data (electricity generation).

Covers only tracked frontier facilities. Total AI-related electricity use (including cloud, inference and smaller facilities) is likely higher.

AI Models

Latest AI ModelsLatest AI Models Released

ModelOrganisationDate
Gemini 4 CarbonGoogle DeepMind2026-11-10
Claude Mythos 5.2Anthropic2026-11-09
Claude Fable 5.2Anthropic2026-11-09
Inkling 2 SmallThinking Machines2026-11-05
Ripple-65B PreviewRicursive Intelligence2026-11-04
MAI-Base-2Microsoft2026-10-15
Grok Voice 3.0xAI2026-10-14
Gemini 4 ArgonGoogle DeepMind2026-10-13
Muse Wonder FullMeta AI2026-10-12
Muse WonderMeta AI2026-10-12
Mistral Large 4Mistral AI2026-10-06
GPT-6.1 SolOpenAI2026-09-29
Claude Sonnet 5.5Anthropic2026-09-28
MiMo-V2.6 ProXiaomi Corp2026-09-22
GPT-6 SolOpenAI2026-09-22

Most recently released AI models.

Data from Epoch AIEpoch AI

Epoch AI. Continuously updated.

Models by OrganisationAI Models by Organisation (2025-2026)

Notable AI models released in 2025-2026, by organisation.

Data from Epoch AIEpoch AI

Epoch AI. Continuously updated.

AI Models Per YearAI Models Released Per Year

AI models released per year (2010-present).

Data from Epoch AIEpoch AI

Epoch AI. Continuously updated.

AI Systems Per YearAI Systems Released Per Year

Large-scale AI systems released each year, by type - language, vision, multimodal and more.

Data from Epoch AI and Our World in DataEpoch AI and Our World in Data

Epoch AI via Our World in Data. Updated monthly.

AI Systems by CountryCumulative AI Systems by Country

Large-scale AI systems built since 2017, running total by country.

Data from Epoch AI and Our World in DataEpoch AI and Our World in Data

Epoch AI via Our World in Data. Updated monthly.

FrontierMath Benchmark

Problems Solved (%) - Top 10 Models

Tiers 1-3 - research-level maths problems that can take expert mathematicians hours or days. Each model at its best setting. The current problem set, in use since Jun 2026, with models released up to Nov 2026.

Data from Epoch AIEpoch AI

Epoch AI Benchmarking Hub (CC BY). Updated daily.

Investment

Global AI Investment

Private investment in AI companies raising over $1.5M, in 2021 dollars (adjusted for inflation).Data through 2025. The AI Index Report adds a year each spring.

Data from Quid via the Stanford AI Index and Our World in DataQuid via the Stanford AI Index and Our World in Data

Quid via AI Index Report / Our World in Data.