Explainer · uses Where Is The Compute data

From compute to agents

Why the always-on agent wave is a compute story, in five steps. Figures come from the companion site with the same caveats.

1 · A chat is one call. An agent is a loop.

An agent plans, calls tools, reads results, checks itself and retries. The brief puts it at 10–20+ calls per task ⚑ reported. Gartner measures the cost side: agentic models need 5–30× the tokens per task ✓ verified, and inference cost per agentic workflow will rise more than fivefold through 2028 ✓ verified.

2 · Tokens per task

Chatbot reply 1×

Agent task (low) 5×

Agent task (high) 30×

Gartner, March 2026 ✓. Always-on agents add scheduled and background work on top.

3 · Inference becomes most of the compute

2023 — inference ~⅓

2025 — inference ~½

2026 — inference ~⅔

Deloitte TMT Predictions 2026 (a prediction, not a measurement) ✓. Separately, ExoBrain estimates frontier labs’ RL share at 53% and pre-training at 10% in Q3 2026 ◐. The brief’s “single-digit pre-training” doesn’t match that.

4 · Who uses the compute, and whose agent that is

Compute used by each lab, end-2025, Epoch AI medians with 90% intervals (amber). OpenAI and Google DeepMind are effectively tied. ✓ compute site

Compute owned by the parent company, Q1 2026 (preliminary). Owners rent much of this out, so it isn’t all for their own agents.

The largest single cluster, SpaceXAI’s Colossus 2 (~550K GPUs, operator-reported), partly serves Anthropic. The biggest cluster is not necessarily the biggest agent fleet.

5 · Pricing as rationing ◇ opinion

The brief reads free tiers, weekly meters and $100+ plans as ways of sharing out scarce chips ◇. Below: lab compute against entry price for five agents. Five points prove nothing. Hover for values.

No vendor publishes accelerator-hours per agent user, so nobody can rank compute actually assigned to Muse, Dots, Grok Bot, Claude or Spark. Prices as of 20 Oct 2026.

What happens 2026–30? (forecasts) →