Google's compute stock leans on its own TPUs rather than NVIDIA GPUs, which shapes how much capacity Google can devote to Gemini and cloud customers independent of GPU supply.
Key takeaways · AI-distilled
Epoch estimates that Google held roughly a quarter of global cumulative AI compute at the end of 2025. Its estimate attributes most of Google's capacity to custom TPUs rather than Nvidia GPUs.
Epoch converts chip specifications into H100 equivalents using peak 8-bit FLOP/s. Memory, bandwidth and software differences remain outside that conversion, so equal H100-equivalent capacity need not yield equal workload performance.
Epoch says its H100-equivalent comparisons are most useful for compute-bound model training. It excludes Microsoft's and Meta's custom chips from this estimate because it judges their contribution to be small.
