SUBSTRATE WIRE
SupportRSS

Fully AI-automated. No human reviews items before they're published. Every item links to its source. AI checks can miss errors. Report corrections.

Fully AI-automated. No human reviews items before they’re published. Every item links to its source; AI checks can miss errors - report corrections here. How evidence is checked.

Compute Intel

Compare three intelligence rankings and dated compute estimates: The AI race →

China: tracked clusters and coverage limits →

Who owns the chips?

ESTIMATED CAPACITY · H100e

Epoch AI estimates ↗ · generated Mar 10, 2026 - Apr 22, 2026 where supplied.

Source fetched Oct 21, 2026 PT. Checked weekly; the underlying estimates update less often.

Epoch’s owner row “xAI” is shown as SpaceXAI (formerly xAI), following the 2026 acquisition by SpaceX. The historical estimates are unchanged.

Latest estimates by chip maker. Includes incomplete reporting periods. Coverage dates are shown on each row.

01

Google

Through Mar 31, 2026 · incomplete

ESTIMATE≈5.47MH100e
range not additive+8.4% QoQ · Q1 2026 · incomplete

Google TPU 76.3% · Nvidia GPU 23.7%

Largest publicly documented cluster: Magic G4 Google Cloud Rental · ≈8K H100e estimate · Existing

Chip mix, ranges & sources

5,472,735 H100e · Epoch median estimate (summed total). Medians are added; percentile ranges are not.

  • Google TPU4,173,802 H100e · Epoch median estimate

    5th - 95th percentile: 3,383,807 - 4,960,594 H100e · 76.3% of this total

    Through Mar 31, 2026 · +9.5% QoQ · incomplete

    Estimates generated on: Mar 27, 2026

    Incomplete: based on Broadcom fiscal quarters ending 2/1/2026. Estimates generated on: 03-27-2026 18:13

  • Nvidia GPU1,298,933 H100e · Epoch median estimate

    5th - 95th percentile: 1,005,404 - 1,669,709 H100e · 23.7% of this total

    Through Mar 31, 2026 · +5.0% QoQ · incomplete

    Estimates generated on: Apr 15, 2026

    Incomplete: based on Nvidia fiscal quarters ending 1/25/2026. Estimates generated on: 04-15-2026 18:24

Owner QoQ compares all chip makers in Q1 2026 with Q4 2025. This may precede the latest total.

Largest publicly documented cluster

Companies do not disclose most cluster sizes. Published examples cover only part of each company's infrastructure.

Magic G4 Google Cloud Rental ↗

8,000 H100e · 11.2 MW capacity · ESTIMATE

Status: Existing · Source certainty: Confirmed

Used by: Magic

Owner(s) in source: Google

Existing/operational clusters are preferred. Coverage is incomplete; a fleet total is not one cluster. Cluster capacity is separate from the ownership estimate above.

Epoch AI · AI Supercomputers ↗
Epoch AI · AI Chip Owners ↗
02

Microsoft

Through Dec 31, 2025 - Mar 31, 2026 · incomplete

ESTIMATE≈3.59MH100e
range not additive+23.9% QoQ · Q4 2025

Nvidia GPU 92.6% · AMD GPU 7.4%

Largest publicly documented cluster: OpenAI/Microsoft Goodyear Arizona · ≈100K H100e estimate · Existing

Chip mix, ranges & sources

3,587,962 H100e · Epoch median estimate (summed total). Medians are added; percentile ranges are not. This total combines different reporting periods.

  • Nvidia GPU3,323,440 H100e · Epoch median estimate

    5th - 95th percentile: 2,558,193 - 4,304,925 H100e · 92.6% of this total

    Through Mar 31, 2026 · +5.5% QoQ · incomplete

    Estimates generated on: Apr 15, 2026

    Incomplete: based on Nvidia fiscal quarters ending 1/25/2026. Estimates generated on: 04-15-2026 18:24

  • AMD GPU264,522 H100e · Epoch median estimate

    5th - 95th percentile: 195,584 - 358,133 H100e · 7.4% of this total

    Through Dec 31, 2025 · +22.5% QoQ

    Estimates generated on: Apr 3, 2026

Owner QoQ compares all chip makers in Q4 2025 with Q3 2025. This may precede the latest total.

Largest publicly documented cluster

Companies do not disclose most cluster sizes. Published examples cover only part of each company's infrastructure.

OpenAI/Microsoft Goodyear Arizona ↗

100,000 H100e · 142.7 MW capacity · ESTIMATE

Status: Existing · Source certainty: Likely

Used by: OpenAI

Owner(s) in source: Microsoft,OpenAI

First operational: Oct 2, 2024

Existing/operational clusters are preferred. Coverage is incomplete; a fleet total is not one cluster. Cluster capacity is separate from the ownership estimate above.

Epoch AI · AI Supercomputers ↗
Epoch AI · AI Chip Owners ↗
03

Amazon

Through Dec 31, 2025 - Mar 31, 2026 · incomplete

ESTIMATE≈2.52MH100e
range not additive+24.8% QoQ · Q4 2025

Nvidia GPU 59.5% · Amazon Trainium 40.5%

Largest publicly documented cluster: Amazon Titan training cluster · ≈4.34K H100e estimate · Existing

Chip mix, ranges & sources

2,520,154 H100e · Epoch median estimate (summed total). Medians are added; percentile ranges are not. This total combines different reporting periods.

  • Nvidia GPU1,499,427 H100e · Epoch median estimate

    5th - 95th percentile: 1,160,907 - 1,933,974 H100e · 59.5% of this total

    Through Mar 31, 2026 · +5.0% QoQ · incomplete

    Estimates generated on: Apr 15, 2026

    Incomplete: based on Nvidia fiscal quarters ending 1/25/2026. Estimates generated on: 04-15-2026 18:24

  • Amazon Trainium1,020,727 H100e · Epoch median estimate

    5th - 95th percentile: 907,525 - 1,164,343 H100e · 40.5% of this total

    Through Dec 31, 2025 · +27.7% QoQ

    Estimates generated on: Mar 10, 2026

Owner QoQ compares all chip makers in Q4 2025 with Q3 2025. This may precede the latest total.

Largest publicly documented cluster

Companies do not disclose most cluster sizes. Published examples cover only part of each company's infrastructure.

Amazon Titan training cluster ↗

4,339 H100e · 11.4 MW capacity · ESTIMATE

Status: Existing · Source certainty: Likely

Used by: Amazon

Owner(s) in source: Amazon

First operational: Jan 15, 2023

Existing/operational clusters are preferred. Coverage is incomplete; a fleet total is not one cluster. Cluster capacity is separate from the ownership estimate above.

Epoch AI · AI Supercomputers ↗
Epoch AI · AI Chip Owners ↗
04

Meta

Through Dec 31, 2025 - Mar 31, 2026 · incomplete

ESTIMATE≈2.41MH100e
range not additive+25.8% QoQ · Q4 2025

Nvidia GPU 80.9% · AMD GPU 19.1%

Largest publicly documented cluster: Meta 100k · ≈100K H100e estimate · Existing

Chip mix, ranges & sources

2,410,152 H100e · Epoch median estimate (summed total). Medians are added; percentile ranges are not. This total combines different reporting periods.

  • Nvidia GPU1,950,294 H100e · Epoch median estimate

    5th - 95th percentile: 1,509,815 - 2,521,685 H100e · 80.9% of this total

    Through Mar 31, 2026 · +5.9% QoQ · incomplete

    Estimates generated on: Apr 15, 2026

    Incomplete: based on Nvidia fiscal quarters ending 1/25/2026. Estimates generated on: 04-15-2026 18:24

  • AMD GPU459,858 H100e · Epoch median estimate

    5th - 95th percentile: 344,991 - 612,180 H100e · 19.1% of this total

    Through Dec 31, 2025 · +24.6% QoQ

    Estimates generated on: Apr 3, 2026

Owner QoQ compares all chip makers in Q4 2025 with Q3 2025. This may precede the latest total.

Largest publicly documented cluster

Companies do not disclose most cluster sizes. Published examples cover only part of each company's infrastructure.

Meta 100k ↗

100,000 H100e · 142.7 MW capacity · ESTIMATE

Status: Existing · Source certainty: Likely

Used by: Meta

Owner(s) in source: Meta AI

First operational: Oct 30, 2024

Existing/operational clusters are preferred. Coverage is incomplete; a fleet total is not one cluster. Cluster capacity is separate from the ownership estimate above.

Epoch AI · AI Supercomputers ↗
Epoch AI · AI Chip Owners ↗
05

Oracle

Through Dec 31, 2025 - Mar 31, 2026 · incomplete

ESTIMATE≈1.21MH100e
range not additive+27.9% QoQ · Q4 2025

Nvidia GPU 90.6% · AMD GPU 9.4%

Largest publicly documented cluster: Oracle OCI Supercluster H200s · ≈65.5K H100e estimate · Existing

Chip mix, ranges & sources

1,207,395 H100e · Epoch median estimate (summed total). Medians are added; percentile ranges are not. This total combines different reporting periods.

  • Nvidia GPU1,094,216 H100e · Epoch median estimate

    5th - 95th percentile: 815,384 - 1,471,064 H100e · 90.6% of this total

    Through Mar 31, 2026 · +6.1% QoQ · incomplete

    Estimates generated on: Apr 15, 2026

    Incomplete: based on Nvidia fiscal quarters ending 1/25/2026. Estimates generated on: 04-15-2026 18:24

  • AMD GPU113,179 H100e · Epoch median estimate

    5th - 95th percentile: 74,703 - 175,251 H100e · 9.4% of this total

    Through Dec 31, 2025 · +24.0% QoQ

    Estimates generated on: Apr 3, 2026

Owner QoQ compares all chip makers in Q4 2025 with Q3 2025. This may precede the latest total.

Largest publicly documented cluster

Companies do not disclose most cluster sizes. Published examples cover only part of each company's infrastructure.

Oracle OCI Supercluster H200s ↗

65,536 H100e · 93.5 MW capacity · ESTIMATE

Status: Existing · Source certainty: Likely

Used by: Cloud

Owner(s) in source: Oracle

First operational: Nov 21, 2024

Existing/operational clusters are preferred. Coverage is incomplete; a fleet total is not one cluster. Cluster capacity is separate from the ownership estimate above.

Epoch AI · AI Supercomputers ↗
Epoch AI · AI Chip Owners ↗
06

CoreWeave

Through Dec 31, 2025

ESTIMATE≈832KH100e
≈785K - ≈882K H100e · 5th - 95th percentile+77.6% QoQ · Q4 2025

Nvidia GPU 100.0%

Largest publicly documented cluster: CoreWeave H200s · ≈42K H100e estimate · Existing

Chip mix, ranges & sources

832,068 H100e · Epoch median estimate (summed total). Medians are added; percentile ranges are not.

  • Nvidia GPU832,068 H100e · Epoch median estimate

    5th - 95th percentile: 785,173 - 882,226 H100e · 100% of this total

    Through Dec 31, 2025 · +77.6% QoQ

    Estimates generated on: Apr 22, 2026

Owner QoQ compares all chip makers in Q4 2025 with Q3 2025. This may precede the latest total.

Largest publicly documented cluster

Companies do not disclose most cluster sizes. Published examples cover only part of each company's infrastructure.

CoreWeave H200s ↗

42,000 H100e · 59.9 MW capacity · ESTIMATE

Status: Existing · Source certainty: Likely

Used by: CoreWeave

Owner(s) in source: CoreWeave

First operational: Aug 28, 2024

Existing/operational clusters are preferred. Coverage is incomplete; a fleet total is not one cluster. Cluster capacity is separate from the ownership estimate above.

Epoch AI · AI Supercomputers ↗
Epoch AI · AI Chip Owners ↗
07

SpaceXAI (formerly xAI)

Through Dec 31, 2025

ESTIMATE≈554KH100e
range n/a0.0% QoQ · Q4 2025

Nvidia GPU 100.0%

Largest publicly documented cluster: xAI Colossus Memphis Phase 3 · ≈276K H100e estimate · Existing

Chip mix, ranges & sources

553,714 H100e · Epoch median estimate (summed total). Medians are added; percentile ranges are not.

  • Nvidia GPU553,714 H100e · Epoch median estimate

    range n/a · 100% of this total

    Through Dec 31, 2025 · 0.0% QoQ

    Estimates generated on: not supplied

Owner QoQ compares all chip makers in Q4 2025 with Q3 2025. This may precede the latest total.

Largest publicly documented cluster

Companies do not disclose most cluster sizes. Published examples cover only part of each company's infrastructure.

xAI Colossus Memphis Phase 3 ↗

275,796 H100e · 352.4 MW capacity · ESTIMATE

Status: Existing · Source certainty: Confirmed

Used by: xAI

Owner(s) in source: xAI

First operational: Jul 22, 2025

Existing/operational clusters are preferred. Coverage is incomplete; a fleet total is not one cluster. Cluster capacity is separate from the ownership estimate above.

Epoch AI · AI Supercomputers ↗
Epoch AI · AI Chip Owners ↗

Bars share one scale across companies. Chip-mix shares use estimated H100e, not chip counts. QoQ uses the latest comparable pair of quarters, dated on each row.

Aggregate bucketsUNRANKED · 3 GROUPS

Country totals and “Other” pool many owners. They are separate from the company ranking; their bars use a separate scale. China (smuggled) is a speculative estimate.

-

All other owners

Through Dec 31, 2025 - Mar 31, 2026 · incomplete

ESTIMATE≈3.73MH100e
range not additive+30.8% QoQ · Q4 2025

Nvidia GPU 89.6% · AMD GPU 10.4%

Chip mix, ranges & sources

3,727,334 H100e · Epoch median estimate (summed total). Medians are added; percentile ranges are not. This total combines different reporting periods.

  • Nvidia GPU3,338,748 H100e · Epoch median estimate

    5th - 95th percentile: 2,299,624 - 4,704,061 H100e · 89.6% of this total

    Through Mar 31, 2026 · +12.3% QoQ · incomplete

    Estimates generated on: Apr 22, 2026

    Excludes Microsoft, Meta, Amazon, Google, Oracle, China, xAI, CoreWeave. Incomplete: based on Nvidia fiscal quarters ending 1/25/2026. Estimates generated on: 04-22-2026 11:57

  • AMD GPU388,586 H100e · Epoch median estimate

    5th - 95th percentile: 228,136 - 538,546 H100e · 10.4% of this total

    Through Dec 31, 2025 · +14.0% QoQ

    Estimates generated on: Apr 3, 2026

Owner QoQ compares all chip makers in Q4 2025 with Q3 2025. This may precede the latest total.

Epoch AI · AI Chip Owners ↗
-

China: all owners, domestic and legally imported chips

Through Dec 31, 2025 - Mar 31, 2026 · incomplete

ESTIMATE≈1.16MH100e
range not additive - QoQ

Huawei Ascend 61.3% · Nvidia GPU 34.8% · Cambricon 3.0% · AMD GPU 0.9%

Chip mix, ranges & sources

1,155,114 H100e · Epoch median estimate (summed total). Medians are added; percentile ranges are not. This total combines different reporting periods.

  • Huawei Ascend707,968 H100e · Epoch median estimate

    range n/a · 61.3% of this total

    Through Dec 31, 2025 · +22.9% QoQ

    Estimates generated on: not supplied

  • Nvidia GPU401,698 H100e · Epoch median estimate

    5th - 95th percentile: 324,871 - 496,849 H100e · 34.8% of this total

    Through Mar 31, 2026 · 0.0% QoQ · incomplete

    Estimates generated on: Apr 15, 2026

    Incomplete: based on Nvidia fiscal quarters ending 1/25/2026. Estimates generated on: 04-15-2026 18:24

  • Cambricon34,925 H100e · Epoch median estimate

    5th - 95th percentile: 32,079 - 45,788 H100e · 3% of this total

    Through Dec 31, 2025 · +31.1% QoQ

    Estimates generated on: not supplied

  • AMD GPU10,523 H100e · Epoch median estimate

    5th - 95th percentile: 8,605 - 12,837 H100e · 0.9% of this total

    Through Dec 31, 2025 · - QoQ

    Estimates generated on: Apr 3, 2026

Owner QoQ is unavailable: there is no common pair of adjacent quarters covering every chip maker.

Epoch AI · AI Chip Owners ↗
-

China: smuggled chips (estimate)

Through Dec 31, 2025

ESTIMATE≈662KH100e
≈294K - ≈1.59M H100e · 5th - 95th percentile+21.0% QoQ · Q4 2025

Nvidia GPU 100.0%

Chip mix, ranges & sources

662,048 H100e · Epoch median estimate (summed total). Medians are added; percentile ranges are not.

  • Nvidia GPU662,048 H100e · Epoch median estimate

    5th - 95th percentile: 294,087 - 1,592,327 H100e · 100% of this total

    Through Dec 31, 2025 · +21.0% QoQ

    Estimates generated on: not supplied

Owner QoQ compares all chip makers in Q4 2025 with Q3 2025. This may precede the latest total.

Epoch AI · AI Chip Owners ↗

How to read the board

An H100-equivalent (H100e) compares a chip’s peak computing power with one Nvidia H100 GPU, using dense 8-bit operations (or 16-bit where 8-bit is unsupported). Peak operations are not measured workload performance.

Labs such as OpenAI and Anthropic mostly rent their compute, so it is counted under the companies that own the chips (for example Microsoft, Oracle, Google, Amazon or CoreWeave). A cloud fleet can also serve other customers and internal workloads.

OpenAI is listed as a user of Microsoft clusters in the supercomputer dataset. This is not a complete rental map.

Each total adds the median of the latest row for each owner and chip maker. Percentile ranges are shown only for single-maker totals; open a row for the individual ranges. A missing range or comparison stays “ - ” or “range n/a”.

Cluster examples use Epoch AI’s AI Supercomputers dataset, fetched Oct 21, 2026 PT, with excluded entries, possible duplicates and non-single-cluster rows removed. Existing/operational entries take priority. These are the largest entries in that dataset, not a claim of exhaustive coverage today.

Ownership CSV ↗Clusters CSV ↗CC BY 4.0 ↗

THE RECORD

Compute event ledger

Model releases, public filings, export controls and price observations. Events are not summed into the ownership scoreboard.

Snapshot assembled Oct 24, 2026 PT · each row keeps its event date and provenance.

254 entries

Ledger sources & last successful checks

A successful fetch can return no new events. Unavailable sources do not imply that no events occurred. AWS spot prices are omitted when credentials are unavailable.