How to read the tracker
This tracker compares AI data-center capacity across operators. The default series is IT power in megawatts, summed from Epoch AI's public campus database (updated 24 Sep 2026). It covers large AI campuses Epoch has built from satellite imagery, permits, and company statements. It is not a company's entire cloud: ordinary regions, older halls, and unnamed colocation stay out.
H100-equivalents are Epoch's modeled conversion of those same campuses into Nvidia H100-class compute. Google TPUs, Amazon Trainium, and Nvidia GB200 or GB300 systems are scaled by Epoch's chip factors. The number is an estimate. It is the comparable compute series. It is not a count of boxes on a floor.
Disclosed chips are raw accelerator counts an operator states, with the chip mix in the note. SpaceXAI's 25 Sep 2026 inventory is 780,000 chips across Colossus 1 and Colossus 2. Those chips are not added into the H100-equivalent series, and a planned delivery is not plotted as operating capacity. Elon Musk's following-week, November, and conditional December GB300 tranches stay in the note.
A campus is attributed to its hardware owner. A lab renting the cluster (OpenAI at Abilene, Anthropic at New Carlisle) is not given a second row, or the same power would be counted twice. Colocation shells with no named AI owner, including several QTS and DayOne sites, are left out. Empty cells mean no comparable disclosure, not zero. Epoch estimates its campus set covers about 44% of global deployed AI compute.