AI Data Centers Updates

A changelog of data centers added, and major changes to power, compute, cost, and chip counts, generated automatically from our dataset.

Aug. 24, 2026
Buildout progress
  • Google Pryor (North)
    • Compute increased from 26k to 763k H100-eq.
    • Power increased from 53 to 368 MW.
    • Cost increased from $2.0 to $13.9 B.
Revisions
Aug. 23, 2026
Revisions
  • CoreWeave Denton TX
    • Projected compute increased from 253k to 272k H100-eq.
    • Projected power increased from 262 to 282 MW.
    • Projected cost increased from $9.9 to $10.7 B.
Aug. 18, 2026
Newly added
Aug. 16, 2026
Buildout progress
  • Google New Albany
    • Compute increased from 352k to 616k H100-eq.
    • Power increased from 339 to 453 MW.
    • Cost increased from $12.8 to $17.2 B.
Aug. 15, 2026
Buildout progress
  • Google Mesa
    • Compute increased from 126k to 343k H100-eq.
    • Power increased from 90 to 183 MW.
    • Cost increased from $3.4 to $6.9 B.
Aug. 12, 2026
Newly added
Revisions
  • AWS Berwick
    • Trainium2 chips increased from 98.6k to 123.8k.
    • Projected Trainium3 chips increased from 34.2k to 50.1k.
  • Google Council Bluffs (East): Compute decreased from 339k to 289k H100-eq.
  • Meta Prometheus
    • Compute decreased from 763k to 677k H100-eq.
    • Power decreased from 631 to 562 MW.
    • Cost decreased from $23.9 to $21.3 B.
    • B200 chips decreased from 302k to 269k.
Aug. 9, 2026
Buildout progress
  • Southgate Melbourne became newly operational with 46k H100-eq of compute, 41 MW of power, and $1.6 B capital cost.
Aug. 6, 2026
Newly added
Revisions
  • Anthropic Lake Mariner: Projected TPU v7 chips increased from 25.1k to 226.1k.
  • Fluidstack Lake Mariner: Removed from the dataset.
Aug. 4, 2026
Newly added
Revisions
Aug. 3, 2026
Revisions
Aug. 2, 2026
Buildout progress
  • Meta Sarpy became newly operational with 92k H100-eq of compute, 80 MW of power, and $3.0 B capital cost.
Aug. 1, 2026
Buildout progress
  • Google Kansas City East became newly operational with 245k H100-eq of compute, 105 MW of power, and $4.0 B capital cost.
Jul. 31, 2026
Revisions
  • Crusoe Abilene Expansion
    • Projected compute increased from 1,481k to 1,708k H100-eq.
    • Projected Rubin chips increased from 167.4k to 192.9k.
  • Goodnight
    • Projected compute increased from 2,011k to 2,279k H100-eq.
    • Projected Rubin chips increased from 199.3k to 229.7k.
  • Meta Hyperion
    • Projected compute increased from 3,692k to 4,254k H100-eq.
    • Projected Rubin chips increased from 417.5k to 481k.
  • OpenAI Stargate New Mexico
    • Projected compute increased from 4,598k to 5,104k H100-eq.
    • Projected power increased from 1,578 to 1,750 MW.
    • Projected cost increased from $59.8 to $66.3 B.
  • OpenAI Stargate Shackelford
    • Projected compute increased from 3,962k to 4,563k H100-eq.
    • Projected Rubin chips increased from 174.4k to 201k.
  • QTS Cedar Rapids
    • Projected compute increased from 3,519k to 3,669k H100-eq.
    • Projected Rubin chips increased from 282.4k to 325.4k.
  • Start Campus Sines Data Campus: Projected compute increased from 472k to 541k H100-eq.
Jul. 30, 2026
Revisions
  • Colossus 2
    • Projected compute increased from 1,112k to 1,824k H100-eq.
    • Projected power increased from 946 to 1,531 MW.
    • Projected cost increased from $35.8 to $58.0 B.
    • Projected B300 chips increased from 330k to 612.2k.
  • Google Council Bluffs (East)
    • Compute increased from 335k to 339k H100-eq.
    • TPU v6e chips increased from 167.1k to 193.9k.
    • Added chip type: TPU v7 (36.6k).
Jul. 29, 2026
Revisions
Jul. 28, 2026
Newly added
Jul. 27, 2026
Revisions
  • Google Council Bluffs (East)
    • TPU v6e chips increased from 167.1k to 193.9k.
    • Added chip type: TPU v7 (36.6k).
  • OpenAI Stargate Lordstown: Removed from the dataset.
Jul. 25, 2026
Buildout progress
  • Meta Cheyenne became newly operational with 173k H100-eq of compute, 152 MW of power, and $5.8 B capital cost.
Revisions
Jul. 24, 2026
Newly added
Jul. 22, 2026

Note: We renamed the Nvidia ‘R200’ and ‘R300’ chip types to ‘Rubin’, because the latest Nvidia documentation just calls the GPU ‘Rubin’, and it’s unclear whether there will be an enhanced version of Rubin in the future.

Revisions
  • Amazon Ridgeland: Added projected chip type: Trainium3 (502.4k).
  • Crusoe Abilene Expansion
    • Added projected chip type: Rubin (167.4k).
    • Removed projected chip type: R200 (167.4k).
  • Goodnight
    • Added projected chip type: Rubin (199.3k).
    • Removed projected chip type: R200 (199.3k).
  • Google Cedar Rapids: Projected compute decreased from 854k to 816k H100-eq.
  • Google Council Bluffs (East): Compute decreased from 339k to 335k H100-eq.
  • Meta Hyperion
    • Projected compute decreased from 3,694k to 3,692k H100-eq.
    • Added projected chip type: Rubin (417.5k).
    • Removed projected chip type: R200 (309.2k).
  • Meta Montgomery: Projected B300 chips increased from 92.2k to 161.3k.
  • OpenAI Stargate Shackelford
    • Projected compute increased from 3,623k to 3,962k H100-eq.
    • Added projected chip type: Rubin (174.4k).
    • Removed projected chip type: R200 (174.4k).
    • Removed projected chip type: R300 (147.4k).
  • OpenAI Stargate UAE
    • Added projected chip type: Rubin (200k).
    • Removed projected chip type: R200 (300k).
  • QTS Cedar Rapids
    • Added projected chip type: Rubin (282.4k).
    • Removed projected chip type: R200 (111.6k).
    • Removed projected chip type: R300 (144.3k).
Jul. 17, 2026
Newly added
Revisions
Jul. 16, 2026
Newly added
Revisions
  • Fluidstack Lake Mariner
    • Compute decreased from 157k to 130k H100-eq.
    • Projected compute decreased from 1,071k to 1,046k H100-eq.
    • Added chip type: TPU v7 (25.1k).
Jul. 15, 2026
Buildout progress
  • Meta Aiken became newly operational with 162k H100-eq of compute, 142 MW of power, and $5.4 B capital cost.
Revisions
  • Amazon Madison Mega Site
    • Trainium2 chips decreased from 327k to 326.6k.
    • Projected Trainium2 chips decreased from 784k to 783.7k.
  • Anthropic-Amazon New Carlisle
    • Compute decreased from 687k to 686k H100-eq.
    • Projected compute decreased from 1,748k to 1,746k H100-eq.
    • Trainium2 chips increased from 1,040k to 1,045k.
    • Projected Trainium2 chips increased from 1,040k to 1,045k.
  • CoreWeave Chester VA
    • Projected compute decreased from 211k to 126k H100-eq.
    • Projected B200 chips increased from 26.2k to 41k.
    • Removed projected chip type: R200 (21.4k).
  • Fluidstack Lake Mariner
    • Removed chip type: TPU v6e (unknown count).
    • Removed chip type: TPU v7 (25.1k).
    • Removed projected chip type: TPU v6e (45,514.9k).
  • Microsoft Goodyear: Removed chip type: A100 (20k).
  • Microsoft Project Osmium: Compute increased from 152k to 156k H100-eq.
Jul. 14, 2026

Note: We corrected the 8-bit OP/s parameter for the TPU v6e chip. Previously, this used the FP8 specification, but the chip has an INT8 specification that is twice the speed. This resulted in much higher current compute estimates across many data centers that use Google TPUs. It also raised the compute extrapolation for future data centers using Google TPUs.

Revisions
  • Alibaba Zhangbei: Compute decreased from 133k to 32k H100-eq.
  • CoreWeave Chester VA
    • Compute decreased from 142k to 88k H100-eq.
    • Added chip type: B200 (26.2k).
    • Removed chip type: R200 (13.6k).
  • Fluidstack Lake Mariner
    • Compute increased from 110k to 157k H100-eq.
    • Projected compute increased from 469k to 1,071k H100-eq.
    • Projected TPU v6e chips increased from 360.4k to 45,514.9k.
  • Google Arcola: Compute increased from 64k to 157k H100-eq.
  • Google Cedar Rapids: Projected compute increased from 313k to 854k H100-eq.
  • Google Columbus
    • Compute increased from 215k to 409k H100-eq.
    • Projected compute increased from 321k to 707k H100-eq.
  • Google Council Bluffs (East): Compute increased from 163k to 339k H100-eq.
  • Google Fort Wayne: Projected compute increased from 1,425k to 4,704k H100-eq.
  • Google Kansas City East: Projected compute increased from 102k to 245k H100-eq.
  • Google Lancaster: Compute increased from 87k to 130k H100-eq.
  • Google Lincoln
    • Compute increased from 127k to 288k H100-eq.
    • Projected compute increased from 323k to 869k H100-eq.
  • Google Mesa
    • Compute increased from 63k to 126k H100-eq.
    • Projected compute increased from 153k to 343k H100-eq.
  • Google Midlothian: Compute increased from 87k to 191k H100-eq.
  • Google New Albany
    • Compute increased from 207k to 352k H100-eq.
    • Projected compute increased from 318k to 616k H100-eq.
  • Google Omaha
    • Compute increased from 136k to 268k H100-eq.
    • Projected compute increased from 398k to 1,137k H100-eq.
  • Google Pryor (North)
    • Compute increased from 17k to 26k H100-eq.
    • Projected compute increased from 326k to 763k H100-eq.
  • Google Red Oak
    • Compute increased from 47k to 108k H100-eq.
    • Projected compute increased from 129k to 316k H100-eq.
  • Google Storey County
    • Compute increased from 89k to 154k H100-eq.
    • Projected compute increased from 267k to 652k H100-eq.
  • Google The Dalles
    • Compute increased from 110k to 202k H100-eq.
    • TPU v5p chips decreased from 32.8k to 16.4k.
    • TPU v6e chips decreased from 186.5k to 119.5k.
    • Added chip type: TPU v7 (35.8k).
  • Google Waltham Cross: Compute increased from 77k to 124k H100-eq.
Jul. 13, 2026
Revisions
Jul. 10, 2026

Note: We lowered our confidence that Google DeepMind is using several Google data centers, given how much compute we expect Google Cloud to lease to customers.

Revisions
Jul. 9, 2026
Revisions
Jul. 8, 2026
Newly added
Revisions
  • CoreWeave Ellendale ND
    • B200 chips decreased from 37k to 36.9k.
    • Projected B200 chips decreased from 200.8k to 200.5k.
  • Google New Albany
    • Added capital cost of $12.8 B.
    • Projected compute decreased from 328k to 318k H100-eq.
  • Huawei Horinger
    • Compute increased from 128k to 134k H100-eq.
    • Ascend 910C chips increased from 65.4k to 69.6k.
  • VNET Bayin Ulanqab
    • Compute increased from 119k to 123k H100-eq.
    • Projected compute increased from 377k to 388k H100-eq.
    • Ascend 910C chips increased from 156.9k to 161.3k.
Jul. 7, 2026
Newly added
Revisions
  • Amazon Ridgeland
    • Projected compute increased from 633k to 788k H100-eq.
    • Projected Trainium2 chips decreased from 964.6k to 763.6k.
  • Colossus 2
    • Compute increased from 1,112k to 1,390k H100-eq.
    • B300 chips increased from 220k to 330k.
  • OpenAI Stargate Shackelford: Projected compute decreased from 4,164k to 3,623k H100-eq.
Jul. 6, 2026
Newly added
Revisions
Jul. 3, 2026
Revisions
  • Huawei Horinger
    • Compute decreased from 129k to 128k H100-eq.
    • Ascend 910C chips decreased from 66.5k to 65.4k.
Jul. 2, 2026
Newly added
Revisions
Jul. 1, 2026
Buildout progress
  • Colossus 2
    • Compute increased from 556k to 1,112k H100-eq.
    • Power increased from 490 to 946 MW.
    • Cost increased from $18.6 to $35.8 B.
    • B300 chips increased from 110k to 220k.
  • Google Midlothian
    • Compute increased from 56k to 87k H100-eq.
    • Power increased from 69 to 103 MW.
    • Cost increased from $2.6 to $3.9 B.
Jun. 30, 2026
Newly added
Buildout progress
  • CoreWeave Denton TX
    • Compute increased from 126k to 253k H100-eq.
    • Power increased from 131 to 262 MW.
    • Cost increased from $5.0 to $9.9 B.
    • B200 chips increased from 50k to 100k.
  • Coreweave Helios became newly operational with 160k H100-eq of compute, 132 MW of power, and $5.0 B capital cost.
  • Google Omaha
    • Compute increased from 65k to 136k H100-eq.
    • Power increased from 158 to 237 MW.
    • Cost increased from $6.0 to $9.0 B.
Revisions
Jun. 26, 2026
Newly added
Jun. 25, 2026
Revisions
  • Meta Rosemount
    • Compute increased from 203k to 216k H100-eq.
    • Added chip type: B200 (85.4k).
    • Removed chip type: B300 (80.5k).
Jun. 24, 2026
Revisions
  • Colossus 2: Users order changed from SpaceXAI, Cursor, Anthropic to Anthropic, Cursor, SpaceXAI.