Caveats

1. Prices and allowances change weekly. Everything is “as of 20 Oct 2026”. Spark’s entry price fell from $100 to $20 in about ten weeks ◐.
2. Most allowances aren’t published. Only Meta publishes token allowances, in its own “Muse tokens”. Everyone else meters weekly or by credits.
3. Lab compute is not agent compute. No vendor discloses accelerator-hours per agent user. Lab figures are Epoch AI end-2025 estimates with wide intervals, reused from Where Is The Compute.
4. 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.
5. 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.
6. 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.
7. Announced gigawatts are not installed gigawatts. Headline commitments usually land two to four years out and are often renegotiated or cancelable.
8. Different views use different dates (end-2025, Q1 2026, Oct 2026) and units (H100e, GPUs, GW). Don't add or compare them across charts.
9. Product scenarios and the Gemini Relay concept are speculation, clearly labeled. They are not company announcements.
10. Fit scores are opinion. The Picker’s personal/business/parallel/trust scores are our editorial judgement.
11. Forecasts are opinion. Phases, scenarios (40/40/20) and segment verdicts come from the brief author’s judgement ◇.
12. Vendor claims are vendor claims. Security designs (Sentinel, Auto Review, custom rules) are described as companies document them, not independently audited.
13. Missing section. Part of the source brief didn’t reach us. Claims from that part are only shown if found independently.
14. Illustrations are AI-generated, generic and not real facilities or products.