Microsoft GPT-4 cluster

Compute

7.9K H100e

Chips

25,000

First operational

Apr 30, 2022

Microsoft GPT-4 cluster is a large GPU cluster operated by Microsoft, OpenAI, located in the United States, in the private sector. It came online around April 2022. At that moment it ranked number 1 among known clusters.

It is built on 25,000 NVIDIA A100. That amounts to about 7.9K H100e of AI compute. The cluster draws roughly 21.29 MW. Its hardware is estimated at $563M.

Epoch AI rates the certainty of this record as likely. Training finished August 2022. Training lasted ~3 months, meaning that it was operational late May at the latest

Full record

Owner: Microsoft, OpenAI
Users: OpenAI
Sector: Private
Status: Existing
Country: United States
Location: Des Moines, Iowa
Primary chip: NVIDIA A100
Primary chip count: 25,000
Total AI chips: 25,000
H100 equivalents: 7.9K
8-bit throughput: 1.6×10¹⁹ OP/s
Power capacity: 21.29 MW
Hardware cost: $563M
First operational: Apr 30, 2022
Rank when it came online: 1
Certainty: Likely