It is a model, not a measurement. Nothing here is a live feed. This page says exactly which parts rest on published data and which rest on our assumptions.
Roughly 100,000 AI agent turns per second worldwide, where a turn means one model step inside an agent loop rather than one human message. Order of magnitude 105. The honest range is 30,000 to 300,000.
Two independent routes land in the same place. Top down: Google reported 3.2 quadrillion tokens a month in May 2026, so all providers together are plausibly around 8 quadrillion, or about 3 billion tokens a second; agent loops are perhaps 30% of that, and a turn costs 10,000 to 40,000 tokens once you count the context it re-reads. Bottom up: around 15 million developers running coding agents at a few hundred turns a day, plus support, research and workflow agents.
The largest single uncertainty is tokens per turn, and it is an ambiguity rather than a margin: whether providers count cached prefix reads in "tokens processed" moves the answer three to five times on its own.
Each country's total is its national population times a per-capita factor built from consumer LLM adoption, developer population and per-capita usage intensity, weighted about 60% developer-driven because coding agents emit tens of turns per session and chat emits one. About 400 cities then distribute that total inside each country.
No coastlines are drawn. The geography appears because a wide kernel around each city traces where people actually live, and people live on coasts.
Chinese activity is modelled rather than observed. It runs on domestic platforms that appear in no public dataset, so building the weighting from Western telemetry alone would render China as a dark hole beside a blazing India. That is a visible factual error; an estimate is not. The same applies at smaller scale to Russia and Iran.
Local time is computed per column from longitude, using a sharp daytime curve for coding agents and a broader consumer curve with an evening bump. The shapes are assumed, not measured. The world total swings only about 1.5 times across a day because the Americas, Europe and Asia fill in each other's nights, while any single region swings seven to eight times.
The cool points are 89 datacentre sites weighted by approximate AI-relevant capacity. They never dim, because a datacentre does not sleep when its local sun goes down. Capacities mix operating, contracted and under-construction sites, so treat them as relative weights rather than an inventory.
| Input | Source | Confidence |
|---|---|---|
| 3.2 quadrillion tokens/month | Google I/O, May 2026 | high |
| 900M weekly actives, Feb 2026 | OpenAI | high |
| 180M developers, US 28M, India 21.9M | GitHub Octoverse | high |
| Per-capita usage by country | Anthropic Economic Index | high |
| Northern Virginia 3.8 GW | CBRE, Q1 2026 | high |
| US ~45% of AI datacentre capacity | CBRE 2026 trends | medium |
| ChatGPT country shares | aggregators citing Similarweb | medium |
| Google's share of world tokens | our assumption | low |
| Agentic share of tokens | our assumption | low |
| 60/40 developer-to-consumer weighting | our assumption | low |
| Time-of-day curve shapes | our assumption | low |
| China correction factor | our estimate, not publicly verifiable | low |
The low-confidence rows drive a large part of the image. They are exposed as live controls in the piece rather than buried as constants, so you can move them and see what changes.
Replacing the assumed time-of-day curves with measured ones. Public hourly traffic data per country, and developer activity timing from public commit records, would turn two of the low-confidence rows into sourced ones. That work is planned, not done.
Back to the map