Federation of American Scientists satellite imagery identified 37 new hyperscale AI data center construction sites globally. Google accounts for 7 projects, the most of any single company, per the report.
Key facts
- 37 new hyperscale AI data centers identified via satellite imagery
- 18 sites in the United States, 6 in China
- Total estimated power capacity exceeds 48 GW
- Google leads with 7 projects, Microsoft 5, Amazon 4
- Completion dates cluster Q3 2026 through Q2 2028
The Federation of American Scientists (FAS) published a new analysis using commercial satellite imagery to track hyperscale AI data center construction worldwide. The study identified 37 new sites under construction, spanning 12 countries across 4 continents.
The Geographic Distribution
The United States leads with 18 sites, followed by China with 6, Singapore with 3, and Germany with 2. The remaining sites are distributed across the UK, Japan, South Korea, Australia, India, Saudi Arabia, Ireland, and Finland. [According to the FAS report]
Power Capacity Estimates
FAS estimates the total power capacity across all 37 sites exceeds 48 GW. Individual site capacities range from 150 MW to 1.2 GW. The largest single site, located in northern Virginia, is projected at 1.2 GW. [per the satellite analysis]
Company Breakdown
Google leads with 7 identified projects, including the previously reported $5B+ Texas facility for Anthropic with 500 MW capacity. Microsoft has 5 projects, Amazon 4, and Meta 3. The remaining 18 sites could not be attributed to specific operators from satellite imagery alone. [FAS notes]
Construction Timeline
Completion dates cluster between Q3 2026 and Q2 2028. Ten sites are expected to come online in 2027 alone. The report notes that construction pace has accelerated 40% compared to 2024-2025 projects. [FAS analysis]
Unique Take
The satellite data reveals a structural shift: hyperscalers are building ahead of demand rather than reacting to it. This contrasts with 2023-2024, when data center construction lagged GPU availability. The 48 GW pipeline suggests operators expect inference workloads to dwarf training compute requirements within 18 months.
What to watch
Watch for Q3 2026 earnings calls from Google, Microsoft, Amazon, and Meta for updated capex guidance. If the 48 GW pipeline materializes, aggregate hyperscaler capex could exceed $200B annually by 2028. Also track power grid interconnection timelines in northern Virginia and Singapore.
