Jesse Hawkins · The global AI buildout

Today Oct 2026

17.7GW

drawn by the world's 93 largest AI campuses, about 43% of the world's AI computing capacity

The power of 14.4 million US homes.

H100-equivalents
14.4M
Online
73
Building
20
Data-centre power use, 2025
485 TWh
AI compute worldwide
~20.0M H100e
Capex, Q2 2026
$208.4B
Cloud regions
297 in 46 countries
  • AI campus, by power
  • AI supercomputer
  • Cloud region

Data: Epoch AI, CC BY 4.0 · IEA, CC BY 4.0 · company filings · cloud providers · by country · sources

Open the Buildout

Every chart, map and record below opens with a free account.

Sign up or sign in →

No password: a sign-in link comes by email.

7 of 18 series this site follows set a new high in their newest reading, and one a new low.

Every series →

  1. Machine · Producer prices: chips

    A new low

    28.5 in Aug 2026, preliminary

    U.S. Bureau of Labor Statistics

  2. Machine · Producer prices: circuit boards and modules

    A new high

    211.7 in Aug 2026, preliminary

    U.S. Bureau of Labor Statistics

  3. Buildout · US data-center construction

    A new high

    $85.0 bn a year in Aug 2026, preliminary

    U.S. Census Bureau

  4. Power · US home power price

    A new high

    17.99¢ per kWh in Jul 2026

    U.S. Energy Information Administration

Each against its own history on this site · Data: U.S. Bureau of Labor Statistics, public domain · U.S. Census Bureau, public domain · U.S. Energy Information Administration, public domain

+25% where AI data centers draw power. +29% everywhere else.

Power prices by state →

Home power prices since 2021, median state · brighter tiles, more AI capacity against demand · A comparison, not a cause.

  1. AK +19%
  2. ME +72%
  3. WI +29%
  4. VT +24%
  5. NH +34%
  6. WA +40%
  7. ID +25%
  8. MT +23%
  9. ND +14%
  10. MN +20%
  11. IL +43%
  12. MI +20%
  13. NY +46%
  14. MA +33%
  15. OR +36%
  16. NV +16%
  17. WY +25%
  18. SD +16%
  19. IA +11%
  20. IN +27%
  21. OH +43%
  22. PA +50%
  23. NJ +45%
  24. CT +28%
  25. RI +31%
  26. CA +45%
  27. UT +28%
  28. CO +27%
  29. NE +20%
  30. MO +22%
  31. KY +21%
  32. WV +27%
  33. VA +38%
  34. MD +60%
  35. DE +42%
  36. AZ +23%
  37. NM +13%
  38. KS +16%
  39. AR +19%
  40. TN +23%
  41. NC +30%
  42. SC +22%
  43. DC +85%
  44. OK +25%
  45. LA +18%
  46. MS +28%
  47. AL +27%
  48. GA +21%
  49. HI +30%
  50. TX +32%
  51. FL +29%

The Machine: one globe, three layers, and the sales and imports behind it

Open the globe →
  1. Data centers 86% of the tracked power is in the United States
  2. Supercomputers 75% of the listed computing power was in the United States
  3. Chip plants 52% of the listed chip plants are in Japan, the United States and China
  4. Sales +57% in a year: September 2026 sales of TSMC and 8 AI server makers. The other 859 companies on Taiwan’s main board: +42%.
  5. Imports 5.7× US imports of servers, chips and computer parts from Taiwan, Mexico and South Korea: the twelve months to August 2026 against four years earlier

Data: Epoch AI, CC BY 4.0 · Wikipedia's contributors, CC BY-SA 4.0 · Taiwan FSC via TWSE, OGDL 1.0 · U.S. Census Bureau, public domain · scoped and sourced on each page

56% of US workers are in occupations BLS rates high or very high for AI exposure

Find your job →
  1. Home health and personal care aidesModerate
  2. Retail salespersonsVery high
  3. Fast food and counter workersModerate
  4. General and operations managersHigh
  5. Registered nursesHigh
  6. CashiersHigh

The six largest jobs and their BLS AI exposure · 95.4 million of 170.3 million workers rated high or very high · Data: U.S. Bureau of Labor Statistics, public domain

120,136 job cuts announced with AI as the reason in 2026. 23 companies tied a cut to AI in an SEC filing.

Read the Ledger →
  1. 2023 · from May 4,247
  2. 2024 · derived 12,742
  3. 2025 54,836
  4. 2026 · to Sep 120,136

Announced job cuts, AI given as the reason · Data: Challenger, Gray & Christmas · New York State Department of Labor · SEC EDGAR · no company named

Open weights took a median of 127 days to match a closed leader on Arena’s Text board.

Open Frontier →

Text

Now: gemini-4-argon-high, 8 days and counting

WebDev

Now: claude-opus-5.5-max, 15 days and counting

Vision

Now: claude-fable-5-high, 11 days and counting

Text to image

Now: gpt-image-2.5-sunburst, 29 days and counting

proprietary open weights · Each block is one model’s time in first place. The board began in May 2023 for Text; the others later.

First place on four Arena leaderboards · Data: LMArena, CC BY 4.0