Chapter 5 · The physical AI economy

Compute, concrete and copper.

Upstream of every chatbot are models, chips, buildings and power. Epoch AI tracks the first three; Census tracks construction. These are inputs. Capacity is not utilization, announced is not operational, and global is not U.S.

Training compute of notable AI models

Each dot is a model. Blue: models Epoch flags as frontier.

Modeled estimate

474 notable models with compute estimates since 2012. Training compute has grown by many orders of magnitude.

1e141e161e181e201e221e241e262014201620182020202220242026ChatGPT

Caveat. Most compute figures are Epoch estimates from papers, hardware and training time; confidence varies by model. Training compute is not inference capacity.

UnitFloating-point operations (log scale)GeographyGlobal (developer country varies)Observed2012 to 2026-11-05SourceEpoch AIRetrieved2026-11-18
Inspect the evidence
What this chart does not show
  • Inference compute serving users.
  • Economic value of any model.
  • Models without a compute estimate.
Lineage: source to transformation to chart
  1. notable_ai_models.csv (CC BY 4.0)
  2. Rows with publication date and training compute; no imputation
  3. Our World in Data's training-compute charts adapt this Epoch table; they are the same evidence family, not independent confirmation.

Text description. 474 notable models with compute estimates since 2012. Training compute has grown by many orders of magnitude.

Link

AI model releases per year, by modality

Observed

Language and multimodal models dominate; video and audio models grew after 2023.

Language and multimodalImageVideoAudio, music, speechScience and medicineRobotics and drivingOther
02004006008001000201820192020202120222023202420252026*

Caveat. Counts reflect Epoch's inclusion criteria and collection lag; recent years are under-collected (science models especially). Fewer listed models is not fewer models.

UnitCount of models in Epoch's all-models tableGeographyGlobalObserved2018 to 2026 (2026 partial)SourceEpoch AIRetrieved2026-11-18
Inspect the evidence

Text description. Language and multimodal models dominate; video and audio models grew after 2023.

Link

Release tracker: frontier, video, music and audio models

Observed

Table of recent model releases filterable by modality.

DateModelOrganizationModalityEpoch domains
2026-11-17GPT-6.2 AstraOpenAILanguage and multimodalLanguage,Vision,Multimodal
2026-11-17GPT-6.2 SolOpenAILanguage and multimodalLanguage,Multimodal,Vision
2026-11-12Gemini 4 HydrogenGoogle DeepMindVideoLanguage,Multimodal,Vision,Audio,Video
2026-11-10Gemini 4 CarbonGoogle DeepMindLanguage and multimodalLanguage,Multimodal,Vision
2026-11-09Claude Fable 5.2AnthropicLanguage and multimodalLanguage,Multimodal,Vision
2026-11-09Muse EmberMeta AIVideoLanguage,Multimodal,Vision,Audio,Video
2026-11-09Claude Mythos 5.2AnthropicLanguage and multimodalLanguage,Multimodal,Vision
2026-11-05Inkling 2 SmallThinking MachinesLanguage and multimodalLanguage,Vision,Multimodal
2026-11-04SWE-2 FlashCognitionLanguage and multimodalLanguage
2026-11-04Ripple-65B PreviewRicursive IntelligenceLanguage and multimodalLanguage
2026-10-28Particle-1ParticleLanguage and multimodalLanguage
2026-10-27MAI-Image-2.7MicrosoftImageImage generation
2026-10-27Genie 4Google DeepMindVideo3D modeling,Language,Audio,Multimodal,Video
2026-10-27MAI-Cyber-1.1-FlashMicrosoftLanguage and multimodalCybersecurity,Language
2026-10-26GPT-6.1 TerraOpenAILanguage and multimodalLanguage,Multimodal,Vision
2026-10-23Kimi K3.1 ProMoonshotVideoLanguage,Multimodal,Vision,Video
2026-10-22BobaHumans& (Humans and)Language and multimodalLanguage
2026-10-22Canva CraftCanvaImageImage generation
2026-10-21Mirendil-BaseMirendilLanguage and multimodalLanguage
2026-10-20AmpersandHumans& (Humans and)Language and multimodalLanguage
2026-10-16Mercury 3Inception LabsLanguage and multimodalLanguage
2026-10-15Muse Spark OpenMeta AIVideoLanguage,Multimodal,Vision,Audio,Video
2026-10-15MAI-Base-2MicrosoftLanguage and multimodalLanguage,Vision,Multimodal
2026-10-15Kling 4.0Kuaishou TechnologyVideoImage generation,Video,Audio
2026-10-15GPT-Live-1.1 QuipOpenAIAudio, music, speechAudio
2026-10-15GPT-Live-1.1OpenAIAudio, music, speechAudio
2026-10-15Bonsai 2 9BPrismMLLanguage and multimodalLanguage,Vision,Multimodal
2026-10-14Grok Voice 3.0xAIAudio, music, speechLanguage,Audio,Speech
2026-10-13Gemini 4 ArgonGoogle DeepMindLanguage and multimodalLanguage,Multimodal,Vision
2026-10-12Muse WonderMeta AIVideoLanguage,Vision,Video,Multimodal
2026-10-12Muse Wonder FullMeta AIVideoLanguage,Vision,Video,Multimodal
2026-10-09Qwen-Image-2.1-ProAlibabaImageImage generation
2026-10-07Claude Haiku 5.5AnthropicLanguage and multimodalLanguage,Vision
2026-10-06Mistral Large 4Mistral AILanguage and multimodalLanguage,Multimodal,Vision
2026-10-06Gemini 3.6 Flash Image (Nano Banana 2.1)Google DeepMindImageImage generation,Language
2026-10-05Beam 1.0ReflectionLanguage and multimodalLanguage
2026-10-01Grok Imagine Video 1.5 LitexAIVideoImage generation,Video,Vision,Audio
2026-10-01MAI-Voice-2.1-FlashMicrosoftAudio, music, speechSpeech
2026-10-01MAI-Voice-2.1MicrosoftAudio, music, speechSpeech
2026-10-01MAI-Transcribe-2-StreamingMicrosoftAudio, music, speechSpeech,Language,Audio
2026-09-30Mercury DecideInception LabsLanguage and multimodalLanguage
2026-09-29GPT-6.1 SolOpenAILanguage and multimodalMultimodal,Language,Vision
2026-09-29Mercury VoiceInception LabsLanguage and multimodalLanguage
2026-09-28Claude Sonnet 5.5AnthropicLanguage and multimodalLanguage,Multimodal,Vision
2026-09-23Nemotron 3.5 Bolt E4BNvidiaAudio, music, speechLanguage,Vision,Multimodal,Audio,Speech
2026-09-22MiMo-V2.6 ProXiaomi CorpLanguage and multimodalLanguage
2026-09-22MiMo-V2.6 Flash Xiaomi CorpLanguage and multimodalLanguage
2026-09-22Claude Opus 5.5AnthropicLanguage and multimodalLanguage
2026-09-22GPT-6 SolOpenAILanguage and multimodalMultimodal,Language,Vision
2026-09-22GPT-6 LunaOpenAILanguage and multimodalMultimodal,Language,Vision
2026-09-22GPT-6 Luna DecisionsOpenAILanguage and multimodalLanguage,Vision,Multimodal
2026-09-22FLUX 3 ActionBlack Forest LabsVideoRobotics,Multimodal,Video,Vision,Language
2026-09-21Grok 4.7xAILanguage and multimodalLanguage
2026-09-20Qwen-Image-2.1AlibabaImageImage generation
2026-09-18Step 5 PreviewStepFunLanguage and multimodalLanguage,Vision,Multimodal
2026-09-15Periodic NeonPeriodic LabsLanguage and multimodalLanguage,Materials science
2026-09-15JevTypeSafe.aiLanguage and multimodalLanguage
2026-09-10SWE-2CognitionLanguage and multimodalLanguage
2026-09-10PersimmonHumans& (Humans and)Language and multimodalLanguage
2026-09-09DeepSeek V4.1 FlashDeepSeekLanguage and multimodalLanguage,Vision,Multimodal

60 releases since 1 Jul 2026 in Epoch's all-models table. Refresh with npm run data.

Caveat. A release is not adoption, and adoption is not impact. Modality is derived from Epoch's domain tags (first match: video, audio, image, language).

UnitModel releasesGeographyGlobalObserved1 Jul 2026 to latestSourceEpoch AIRetrieved2026-11-18
Inspect the evidence

Text description. Table of recent model releases filterable by modality.

Link

Facilities

Frontier AI data centers in the United States.

Epoch tracks 77 large U.S. AI data centers with an estimated 12,640 MW of current power between them. Power is estimated from satellite imagery, permits and filings, not metered. State totals below aggregate Epoch's facility estimates; states without a tracked site are hatched.

Estimated current power of tracked AI data centers, by state

Modeled estimate

Largest tracked capacity: TX 1730 MW, TN 1373 MW, OH 1244 MW, IN 1215 MW.

Caveat. Capacity, not utilization. Only sites Epoch tracks; many smaller data centers are absent. Hosting a facility is not the same as local benefit.

UnitMegawatts (sum of facility estimates)GeographyUnited States, state totalsObservedLatest Epoch observation per siteSourceEpoch AIRetrieved2026-11-18
Inspect the evidence

Text description. Largest tracked capacity: TX 1730 MW, TN 1373 MW, OH 1244 MW, IN 1215 MW.

Link
Epoch AI Frontier Data Centers (CC BY 4.0). Owner tags cleaned of Epoch's confidence markers. Retrieved 2026-11-18.
SiteStatePower (MW)H100eOwner
Anthropic-Amazon New Carlisle
Users: Anthropic
IN1,037818,428Amazon
Colossus 2
Users: Anthropic, Cursor, SpaceXAI
TN9461,111,673SpaceXAI
OpenAI Stargate Abilene
Users: OpenAI
TX8241,018,696Oracle
Microsoft Fairwater Atlanta
Users: OpenAI, Microsoft
GA636768,769Microsoft
Meta Prometheus
Users: Meta
OH471536,343Meta
Microsoft Fairwater Wisconsin
Users: OpenAI, Microsoft
WI369445,680Microsoft
Google Pryor (North)
Users: Google DeepMind
OK368636,685Google
Colossus 1
Users: Anthropic
TN340275,796SpaceXAI
Google New Albany
Users: Google DeepMind
OH333323,901Google
Google Columbus
Users: Google DeepMind
OH303331,986Google
Amazon Madison Mega Site
Users: Anthropic
n/a284214,348Amazon
CoreWeave Denton TX
Users: OpenAI
TX262252,653CoreWeave
QTS Richmond 1
VA238242,547n/a
Google Bristow
Users: Google DeepMind
VA238209,861Google
Google Council Bluffs (East)
Users: Google DeepMind
IA237255,685Google
Google Omaha
Users: Google DeepMind
NE237198,585Google
Google Papillion
Users: Google DeepMind
n/a237100,556Google
Amazon Ridgeland
Users: Anthropic
n/a228171,299Amazon
AWS New Albany
n/a213201,112Amazon
Anthropic Lake Mariner
Users: Anthropic
NY210292,830AI XPV Platform
Microsoft Goodyear
Users: OpenAI, Microsoft
AZ202205,154Microsoft
Microsoft Project Osmium
Users: OpenAI
IA190155,634Microsoft
Google Mesa
n/a183289,540Google
QTS Richmond 2
VA180182,921n/a
Meta-QTS Hillsboro 2
Users: Meta
OR180123,839Meta
Meta Rosemount
Users: Meta
MN178215,675Meta
Meta Jeffersonville
Users: Meta
IN178203,133Meta
Meta Temple
Users: Meta
TX178202,473Meta
Google Storey County
Users: Google DeepMind
n/a161133,906Google
Google The Dalles
n/a154175,341Google
Meta Montgomery
Users: Meta
AL153174,517Meta
Meta Kuna
Users: Meta
n/a152173,248Meta
Meta Cheyenne
Users: Meta
WY152172,815Meta
Meta Huntsville
Users: Meta
n/a146166,761Meta
QTS Richmond 3
VA144174,836n/a
Google Lincoln
NE141197,575Google
Google Lancaster
OH137130,369Google
Meta Eagle Mountain
Users: Meta
UT133151,354Meta
Coreweave Helios
Users: OpenAI
TX132159,552CoreWeave
Meta Aiken
Users: Meta
SC129147,044Meta
The bridge from compute to prosperity runs: capability and cost, then adoption, workflow, output and welfare. Rising compute is strong evidence for the first link only. See the evidence chain on the home page.