The AI Clock

How we count this

These are estimates, not meter readings. No AI company publishes a live global count of prompts, tokens, power or water, so this clock builds one from the best public disclosures, anchored to Aug 8, 2026. Each counter is labelled by kind — a flow that accumulates through the year, a stock that stands right now, or a single record — because mixing those is the fastest way to mislead. The three levels bracket the honest range; the gap is wide because the real uncertainty is.

Growth here is compounding, not linear — that's the whole story of AI right now. Tokens, compute and capex are accelerating several-fold a year, so the totals keep climbing even as each token gets radically cheaper and more efficient. Those multipliers are themselves moving, so the model is meant to be re-anchored every quarter against fresh hyperscaler earnings, NVIDIA filings and IEA / Epoch updates — not extrapolated for years.

Prompts — ~9–15 billion a day worldwide · flow

The only official disclosure is still ChatGPT's ~2.5 billion a day (18 billion messages a week) — and that figure is from mid-2025; OpenAI hasn't refreshed it since. The worldwide total across every assistant is a bottom-up estimate carried forward at ~2.75× a year, so it is one of the softer counters despite resting on a real number. Sources: OpenAI "How People Use ChatGPT" (NBER, Sept 2025); TechCrunch (Jul 2025); Epoch AI (Aug 2025).

Tokens — ~8–15 quadrillion a month worldwide · flow

Google alone disclosed 1.3 quadrillion tokens/month in Oct 2025, up 7× to 3.2 quadrillion by May 2026; the worldwide total adds Microsoft, OpenAI, Anthropic and others. Reasoning and agentic workloads inflate tokens per request. Re-anchored upward in Aug 2026: this counter previously tracked Goldman Sachs' May 2026 estimate of ~5.6 quadrillion/month, which bottom-up run-rate estimates have since put closer to ~11 quadrillion — a level that left Google implausibly close to the entire world total. Compounding ~5× a year; note that longer-range forecasts imply a much slower ~2× a year, so treat this counter as re-anchored quarterly rather than projected. Sources: Google (Pichai, Oct 2025; Google I/O, May 2026); I/O Fund token run-rate (2026); Goldman Sachs (May 2026).

Electricity — ~160–206 TWh in 2026, AI data centres only · flow

AI-focused data centres drew ~155 TWh in 2025 and grew ~50% that year; for 2026 the IEA puts all data centres at ~565 TWh with AI-optimised servers about 31% of that, or ~175 TWh. Inference and training; excludes the wider grid. Sources: IEA "Energy and AI" (2025); IEA "Key Questions on Energy and AI" (2026).

Water — ~312–765 billion liters in 2025 · flow

On-site cooling plus the water used to generate the electricity, derived from the power figure via water-use-efficiency assumptions — one of the softest counters. ~30–35% a year. Sources: Cell Reports Sustainability (2025); Google 2025 Environmental Report; Microsoft (2025).

Inference spend — ~$106 billion in 2025 · flow

Operational spend running models, distinct from buildout capex. Inference is now roughly half of all AI compute, heading to two-thirds in 2026. ~1.2–1.5× a year. Sources: MarketsandMarkets (2025); IDC AI Infrastructure Tracker (2025); Deloitte TMT Predictions 2026.

CO₂ — ~37–89 Mt in 2026, AI only · flow

Grid carbon intensity applied to AI electricity; all data centres emitted ~180 Mt in 2024. Isolating AI's slice needs contested assumptions. The carbon intensity is part of what the three levels bracket, not a fixed input: they work out to roughly 230, 310 and 430 g CO₂/kWh respectively, spanning a heavily-contracted clean-power fleet at the low end and something near the ~395–400 g/kWh world grid average at the high end. That assumption is a judgement, not a measurement, and it is the single biggest reason this counter's range is so wide. Sources: IEA (2025); Cell Reports Sustainability / arXiv 2601.06063 (2025).

Buildout capex — ~$410B in 2025 → ~$725B guided for 2026 · flow

Big-four hyperscaler capital spend on data centres and chips, up ~77% year over year — the firmest and largest number here. Amazon ~$200B, Microsoft ~$190B, Alphabet $175–185B, Meta $115–135B on 2026 guidance. Corrected downward in Aug 2026: this counter had drifted to an implied ~$978B for calendar 2026, well above what the four have actually guided. Sources: company Q2 2026 earnings and guidance; CNBC (Jul 2026); Goldman Sachs "Tracking Trillions" (2026).

People using AI — ~1.1–2.0 billion across all tools · stock

ChatGPT passed ~900 million weekly users by June 2026, with reports of roughly a billion by July — the billion figure is press reporting of internal data, not an OpenAI disclosure. Gemini's app reached ~900M monthly at I/O 2026. Overlap between them is heavy, so the worldwide unique total is well below the sum. Growth at the leaders has slowed to ~1.3× a year (800M in Dec 2025 → ~900M in Jun 2026), down sharply from the ~2× of a year ago as rich markets saturate; this counter now uses 1.35×. Sources: Reuters (Jun 2026); The Information (Jul 2026, unofficial); Google I/O (May 2026); Stanford HAI AI Index 2026.

AI compute installed — ~27–37 million H100-equivalents · stock

The global AI data centre fleet reached ~30 GW of power capacity by Q4 2025, up from ~13.5 GW a year earlier. Capacity growth of ~2.3× a year understates compute growth, since each generation delivers more work per watt; this counter uses ~3.4× a year for H100-equivalents. Source: Epoch AI datacenter power tracker (2026).

Largest training run — ~1×10²⁷ FLOP · record

Projected forward from the biggest confirmed single training run (Grok 4, xAI, mid-2025) at ~4.5× a year. No 2026 record run has been publicly confirmed — labs stopped disclosing training compute for their latest frontier models, so this is the one counter where the honest answer is that nobody outside those labs knows. Treat it as the weakest number on this page. Source: Epoch AI Trends dashboard (2026); frontier growth 5×/yr, CI 4–6×.

The per-unit panel — derived, not measured

"Cheaper every year" divides each total by the token total: energy per 1,000 tokens is AI electricity ÷ tokens, and so on. So those figures fall at the gap between the two growth rates — spend growing ~1.35× against tokens growing ~5× gives a ~73%/yr drop in price per token. Nobody measures this directly and we are not claiming they do. It describes blended real-world spend across a shifting mix of models, which is a different thing from the widely-cited ~40× a year decline you may have seen — that one tracks the price of a fixed capability level (GPT-4-quality output), and ranges from 9× to 900× a year depending which milestone is chosen. Both are real; they answer different questions. We use the derived one because it matches what this panel is actually about. Fixed-capability comparison: Epoch AI, LLM inference price trends (2026).

Rates of change

Tokens ~5×/yr; installed compute ~3.4×/yr; frontier training ~4.5×/yr; capex ~1.77×/yr; AI electricity ~1.4×/yr; users ~1.35×/yr — most decelerating as power, advanced packaging and HBM supply tighten. Per-token energy, cost and water fall fast because volume outruns the totals. Sources: Epoch AI; IEA; Goldman Sachs; a16z.