How to read these numbers

Short version: these are careful estimates with wide error bars, measured in a yardstick that flatters some chips, across different dates. Here’s what that means in plain language.

1. No frontier lab discloses chip counts directly. Almost every figure here is a model built from power capacity, spending, cloud contracts and data-center analysis — that's why the ranges are wide.
2. H100e is a peak-throughput yardstick. It ignores real-world utilization, interconnect, memory bandwidth, and software differences between Nvidia GPUs, Google TPUs and AWS Trainium.
3. Used ≠ owned. OpenAI and Anthropic rent most of what they use; Google and Meta own huge fleets but their AI labs only get part of them.
4. Announced gigawatts are not installed gigawatts. Headline commitments usually land two to four years out and are often renegotiated or cancelable.
5. Different views use different dates (end-2025, Q1 2026, Oct 2026) and units (H100e, GPUs, GW). Don't add or compare them across charts.
6. Product scenarios and the Gemini Relay concept are speculation, clearly labeled. They are not company announcements.

What is an “H100-equivalent”?

H100e = H100-equivalents: a chip's peak dense 8-bit (FP8/INT8) throughput divided by an Nvidia H100's, multiplied by chip count (Epoch AI definition). Example from Epoch: a TPUv7 does ~2.3× an H100’s 8-bit operations, so a million TPUv7s count as 2.3M H100e.

What does a 90% confidence interval mean here?

Epoch builds a Monte Carlo model for each lab from power disclosures, spending, cloud contracts and data-center analysis. The bar is the median (50th percentile); the whiskers are the 5th and 95th percentiles. Epoch notes that some guesswork goes into the inputs, especially for Google DeepMind and Meta Superintelligence Labs.

Why does Google own the most but DeepMind not lead?

Google says just over half of its ML compute goes to Google Cloud in 2026; the rest is split between DeepMind and other internal uses like recommender systems. Epoch’s best guess is that DeepMind uses slightly under half of Google’s total.

What’s missing

• No public league table exists for inference capacity (tokens/sec, concurrent agents, latency, quotas).
• No lab discloses compute per product.
• Chinese labs aren’t in Epoch’s chip-users explorer yet.
• Inference that Microsoft or Amazon run on their own clouds for OpenAI/Anthropic models isn’t counted in those labs’ totals (Epoch: could add ~25%).

Sources: Epoch AI, “Data on AI chip users” (2026-09-09); Epoch AI, “AI chip users — methodology” (2026-09-09); Epoch AI (Gradient Updates, opinion), “Frontier labs don't use most AI compute (yet)” (2026-05-20)