Updated Jul. 25, 2026

Microsoft Project Osmium

Cumming, Iowa, United States

Microsoft Project Osmium is an operational AI data center with AI chips owned by Microsoft and used by OpenAI (speculative). It hosts an estimated 156k H100-equivalents of compute, supported by 190 MW of IT power at a capital cost of $7.2 B. We estimate that it uses Nvidia A100, Nvidia H100, and Nvidia B200 chips.

The timeline of IT power and compute capacity is based on satellite imagery, our cooling equipment power model, company disclosures, chip availability, and chip efficiency.

One building reportedly contained roughly 25,000 A100s used by OpenAI to train GPT-3.5.

OWNER Microsoft STATUS Operational USER OpenAI
Show compute as
Satellite view of Microsoft Project Osmium

Satellite imagery

Latest: Jun. 17, 2026 · 2 high-res images
View satellite explorer →

Scale & specifications

CURRENT COMPUTE

156k H100-eq

3.08e20 8-bit OP/s

CURRENT IT POWER

190 MW

PROJECTED IT POWER

190 MW

CURRENT COST

$7.2 B

PROJECTED COST

$7.2 B

HardwareNvidia A100, Nvidia H100, Nvidia B200

CoolingAir-cooled chiller

Buildout

Dec 18, 20180K H100e
Dec 18, 2018

Land clearing begins on the north part of the site

Jan 1, 2021

Building 1 and 1st building on the site are operational. Foundation work begins on Building 2

Compute
8kH100-eq
IT Power
48MW
Cost
$1.8B
Jun 30, 2023

Building 2 is operational. Estimated based on Microsoft H100 count 2023

Compute
41kH100-eq
IT Power
95MW
Cost
$3.6B
Mar 1, 2025

Building 3-4 operational, no visible construction.

Compute
156kH100-eq
IT Power
190MW
Cost
$7.2B
Oct 8, 2025

Site already complete.

Compute
156kH100-eq
IT Power
190MW
Cost
$7.2B
Jun 16, 2026

No notable changes since October.

Compute
156kH100-eq
IT Power
190MW
Cost
$7.2B

Sources

  1. SemiAnalysis article with IT power estimations and detailing usage by OpenAI for training GPT 3.5
  2. DCD Article on site expansion
  3. MISO Announcement reporting power for newest buildings, likely an overestimate
  4. Microsoft article on data centers in Iowa used by OpenAI

Read more about our methodology.