The caveat is the story.
Epoch AI, an independent AI research organization, updated its Frontier Data Centers and GPU Clusters databases on June 1, 2026. The headline figure: global AI computing capacity is doubling approximately every 7 months, representing 3.3x annual growth since 2022. According to Epoch AI’s database update, NVIDIA holds more than 60 percent of global GPU cluster performance measured in the dataset.
Now the caveat: Epoch AI’s own documentation states that the GPU cluster dataset covers an estimated 10 to 20 percent of existing global aggregate GPU cluster performance. The data is incomplete by design, comprehensive reporting across every GPU cluster deployment globally isn’t feasible. What Epoch tracks is the visible frontier: the large, documented deployments that have enough public reporting to be included.
That’s not a reason to dismiss the figures. It’s a reason to read them correctly.
Evidence
The 7-month doubling rate and 3.3x annual growth figure describe the frontier of documented deployments. The actual total is larger. If the visible 10 to 20 percent is doubling every 7 months, the underlying infrastructure is almost certainly growing at a comparable or faster rate, the undocumented deployments tend to be smaller, less efficient, and catch up to frontier pace with a lag.
For practitioners, the directional conclusion holds: AI compute capacity is growing faster than most enterprise planning models assume. Teams that built three-year infrastructure roadmaps in 2023 are working with assumptions that have already been outpaced. The doubling rate also has direct implications for regulatory compute thresholds. EU AI Act Article 6 uses compute-based classification criteria. What counts as a “general-purpose AI model with systemic risk” today may describe a much larger set of deployments in 18 months if compute continues doubling at this pace.
The NVIDIA concentration figure, more than 60 percent of tracked GPU cluster performance, carries the same caveat. It’s a share of the measured dataset, not a certified global market share. That said, NVIDIA’s production ramp announcements this week, including Vera Rubin entering full production, reinforce why that share concentration has structural staying power. No competing architecture has demonstrated comparable deployment velocity in the documented frontier.
The part nobody mentions: the 10-20% coverage gap means that China’s GPU cluster deployments, which are less thoroughly reported in Western-accessible databases, are underrepresented in Epoch’s data. The actual global compute distribution may look meaningfully different from what the frontier dataset shows.
Analysis
The 10-20% coverage caveat cuts two ways. It means the absolute scale of global AI compute is larger than Epoch's numbers show. It also means the NVIDIA concentration figure (>60% of tracked deployments) is a share of the measurable frontier, not a certified global figure. China's deployments, which are underrepresented in Western-accessible databases, likely shift the actual global distribution. Use these numbers as directional intelligence from the best available independent source. Don't mistake them for a complete census.
What to watch
Epoch’s next database update. The 7-month doubling figure is a trailing measurement. If the next update shows the rate accelerating, 6 months, 5 months, the infrastructure investment thesis for AI tightens considerably. If the rate holds or slows, that’s a signal the buildout is stabilizing.
TJS synthesis
Use Epoch AI’s compute figures as directional intelligence, not precise measurement. The 7-month doubling rate is the strongest publicly available independent estimate of AI infrastructure acceleration, drawn from the most comprehensive frontier dataset that exists. The 10-20% coverage caveat doesn’t undermine the conclusion, it clarifies it. Act on the direction. Don’t over-index on the specific multiples.