The Global Token Meter — live estimate
Tokens are the metered unit of machine intelligence — every AI answer, every agent step, every generated line of code is billed in them. The meter below is an order-of-magnitude estimate built from disclosed figures, not a measurement. Every constant is listed in the methodology, with its source.
Tokens generated worldwide since you opened this page
at ≈ 1.9 billion tokens per second (central estimate)
Inference spend
≈ — per day at a blended $0.50 per million tokens.
Continuous power draw
Running around the clock — roughly two large nuclear reactors' worth of output, for text inference alone.
Cooling water
≈ — Olympic swimming pools drained per day.
For scale
The world generates a full English Wikipedia's worth of tokens (≈6 billion) every few seconds.
Why this looks like a debt clock
Electricity became a resource the moment it was metered: priced per unit, rationed by ability to pay, with a grid, utilities, and blackouts. Tokens crossed that line quietly. They are metered per million, sold in tiers, capped on free plans, and stockpiled by enterprises. Google alone reported 3.2 quadrillion tokens a month at I/O 2026 — a sevenfold increase in one year — and Goldman Sachs expects global consumption to multiply another 24× by 2030. The cost is real: electricity, water, capital, carbon. It is simply invisible to the bystander, because the meter is on someone else's wall.
The divide
Like any metered resource, access is unequal on three axes: volume (how many tokens you can afford), quality (which model generates them), and leverage (whether agents multiply your tokens while you sleep).
Published per-million-token prices span roughly $0.05 to $75 across the 2026 market. The gap between tiers is the quality divide.
Same money, 600× fewer tokens — but each one comes from a model that can do work the budget tier cannot. The rich buy quality and volume; everyone else picks one.
Global consumption spread over 8.2 billion people, against Google Cloud's disclosure that 375 customers each processed over a trillion tokens in a year.
A single trillion-token enterprise consumes about 137,000× the global per-person average — and there are hundreds of them.
The leverage axis is steeper still: Gartner finds agentic AI burns 5–30× more tokens per task than a chat exchange, and the efficiency spread between the lightest and heaviest models exceeds 200×. Those who can afford always-on agents aren't just consuming more tokens — they're converting money into unattended cognition, continuously. That is what "have and have-not" means in this economy.
Methodology
The clock multiplies elapsed time by a per-second rate derived from the constants below. Switch scenarios above to move every derived figure at once. Nothing here is precise; everything is anchored.
| Constant | Value used (central) | Anchor |
|---|---|---|
| Google's disclosed volume | 3.2 quadrillion tokens/mo | Sundar Pichai, Google I/O 2026 keynote, May 2026 (7× the 480T/mo of May 2025) |
| Global monthly volume | 5 quadrillion tokens/mo | Goldman Sachs Research, 2026 baseline (projects 24× to 120Q/mo by 2030). Conservative scenario: 4Q; high: 7Q. |
| Energy per token | 0.0003 Wh | Derived from Google's measured 0.24 Wh per median Gemini text prompt (Aug 2025), assuming ~700 tokens per prompt+response; OpenAI has cited ~0.34 Wh per query. Range 0.0002–0.0005. |
| Water per token | 0.0004 mL | Google: 0.26 mL per median prompt, same per-prompt token assumption. |
| Blended price per million tokens | $0.50 | Market prices span $0.05–$75/M in 2026; most volume runs on cheap and internal models, so the blend sits near the bottom. Range $0.25–$1.00. |
| Agentic multiplier | 5–30× tokens/task | Gartner analysis, March 2026. |
| Reference volumes | Wikipedia ≈ 6B tokens; reactor ≈ 1.1 GW; pool ≈ 2.5M L | English Wikipedia ≈4.5B words at ~1.33 tokens/word; typical large nuclear unit; FINA Olympic pool volume. |