Frontier compute has an address.
A source-backed map of the physical campuses behind large-scale AI. It opens at ≥10,000 H100-equivalents or disclosed accelerators; lower the scale control to explore the 1,000-card inventory, with power, evidence quality and historical frontier systems preserved.
| Site | Status | Country | Scale | Mapped share | Power | Evidence |
|---|
10 large disclosures cannot honestly be placed on a campus map
Physical plus unlocated numeric H100e.
Epoch estimate across six major designers.
Scale check only: the low end counts mapped campuses alone, the high end adds unlocated disclosures that can duplicate campus estimates; the atlas is also newer than the shipment cutoff, and shipments are not necessarily online. Tracking limits →
Where the attributable H100e is concentrated.
Current numeric records, physical and unlocated; China-side possible Epoch overlaps suppressed, but unlocated disclosures can still partly duplicate whole-campus estimates. These are deployment-size bands—not chip-product bands or estimates of total sales. Consumer, workstation and gaming GPUs such as RTX PRO are outside both this atlas and the global shipment reference.
Physical evidence first; uncertainty left intact.
The backbone is Epoch AI’s satellite- and permit-derived campus dataset. Its current and projected H100-equivalent estimates remain separate. Named GPU systems and historically frontier machines come from Epoch’s GPU Clusters data. International gaps were filled only where a site, scale signal and credible operator or financing trail could be found.
Three different scale signals coexist. A numeric H100e estimate is preferred. Chinese domestic-card clusters now receive a visibly labeled central model plus a low/high range when either aggregate FP16 capacity or a defensible chip-performance bracket exists; raw card counts remain alongside it. A small set of dedicated AI campuses qualify through clearly frontier-scale power and committed capital, but receive no invented H100e number. In MW bubble mode, disclosed power always wins; a dashed outline marks a rough facility-load estimate from the disclosed chip mix using accelerator-generation-specific central kW assumptions, rather than a measured meter.
Epoch-derived and independent records can be viewed separately. Named Chinese additions that could correspond to Epoch’s anonymized systems remain visible, but are held out of combined H100e totals when both datasets are selected. This avoids asserting either a false match or a false additional cluster; selecting only the independent subset counts them within that subset.
C1 means direct operator/government disclosure or Epoch’s physical-evidence layer; C2 is corroborated but contains a material unknown; C3 is a real disclosure with incomplete location, schedule or scale evidence. Buildout is not a forecast: it is the project’s ultimate sourced plan, and may never be delivered.
Open Infrastructure Map links on each record help inspect nearby substations, lines and generators. Absence from this atlas is not evidence that a country lacks compute: secret clusters, multi-site cloud disclosures and China’s anonymized systems remain major blind spots.