This week’s chart is from Epoch AI’s brand new Chip Owners Explorer. Google dominates global AI compute ownership with roughly 5 million H100-equivalents, more than Microsoft, and far more than Amazon, Meta or Oracle (and the whole of China). But look closely at the colour. That enormous pink block is TPUs, Google’s custom chips. Strip those out and Google’s Nvidia GPU fleet is relatively modest.
This matters because we’re entering the GB200 NVL72 era. Nvidia’s rack-scale Blackwell systems are delivering big gains for Anthropic. Google’s TPU v6 Trillium chips are strong on price-performance and power efficiency, drawing 300W versus Blackwell’s 1,000W+ per chip. Google claims 4x better cost efficiency than H100 for large language model workloads. But the GB200 NVL72 isn’t an H100. It’s a different beast entirely, and Google’s Ironwood (TPU v7) response is only just arriving.
The chart shows Google owns the most AI compute, but compute isn’t static. If Blackwell’s rack-scale architecture delivers for others as it has for Anthropic, then Microsoft, Meta, Amazon and Oracle, all overwhelmingly Nvidia customers, could see their effective compute surge past Google’s installed base.
Owning the most chips isn’t the same as owning the best ones.