Tokens Per Day Index
The AI inference demand index · July 16, 2026

AI tokens processed today

Tokens are the “barrels per day” of artificial intelligence. There is no official meter for global token consumption, so this is a transparent, fully-sourced estimate, modeled live from the latest disclosures.

~334.7trillion
334.7 trillion tokens so far today · a modeled count, not a meter
= 300.1T disclosed floor + 34.6T implied for non-disclosers, grown to today · method →
300.1T
Disclosed floor →
400–432T
Independent est. →
~691T
Compute capacity →
Tokens per day over time
Trillions/day. Solid = disclosed floor; long-dashed = estimate (the estimated total, and revenue-derived companies like Anthropic); faint dashed = last figure held to today. Hover for values; switch total / country / company.
Global floorEstimated total
Trillions of tokens / day Feb '24AprJunAugOctDecFeb '25AprJunAugOctDecFeb '26AprJun 0 100 200 300 400 0 50 100 150 0 50 100 150 200
Why the curve is this steep, and why 2024 looks near-zero. The growth is genuine: China reports a ~1,400× rise in two years (0.1→140T/day), Google ~330×. The solid floor is a disclosed lower bound that was tiny in 2024 because almost no one reported then; the dashed estimate corrects for that thin coverage (a larger uplift early, when China and others were not yet counted), lifting 2024 to ~1–7T/day. But on a linear axis even that is ~1% of today, so it still looks near-zero. That is real exponential growth, not missing data.

By country is an estimate. Only China publishes national token consumption (National Data Bureau), so China is measured. Other regions split the non-China provider tokens (which are served worldwide) by OpenRouter billing-location shares (US ~47% / Europe ~18% / Asia ex-China ~13% / Rest ~22%). A ~1% sample skewed to Western developers, so treat the non-China regions as low-confidence. Sources →

A day in tokens

When are those tokens processed? Here is the 24-hour shape, each line as a percent of its own peak hour.

Usage over a 24-hour cycle (UTC)
Percent of each series' peak hour. Hover for values.
GlobalUnited StatesChina
0 25 50 75 100 % of peak 00:0004:0008:0012:0016:0020:00
There is no global night. Western providers peak 12–18 UTC (Anthropic's disclosed window: 8am–2pm ET), but China, roughly half of all tokens, peaks while the US sleeps (~00–10 UTC, China's business day). The two largely cancel, so the global line swings only about half (~50–100%) while the US and China each swing about 4×. The world's busiest hour is ~13 UTC, when US morning, European afternoon, and Chinese evening overlap. This is a stylized model, not measured hourly tokens: the peak windows are sourced (Anthropic, Cloudflare Radar), but the curve shapes are inferred from regional business hours and blended by token share, so real data would be noisier. Sources →

Tokens, all time

The running total since reporting began in January 2024, integrated from the estimated daily total (the same basis as the headline), so the daily and all-time figures stay consistent.

~89.6quadrillion
89.6 quadrillion tokens processed since January 2024 · counting
Cumulative tokens processed (all time)
Quadrillions of tokens (1 quadrillion = 1,000 trillion), integrated from the estimated daily total.
0 25 50 75 100 Quadrillion tokens 20252026 ~89.6Q

How our numbers compare to independent estimates

Our reported floor (300.1 T/day) only counts disclosures, so it sits below what independent analysts estimate for the true global total. Those estimates span the paid API market (~50T/day) up to all-surface throughput (~430T/day). Our own estimated total (~334.7 T/day) sits below the all-surface figures too, because it only imputes providers we can explicitly bound — the gap to ~430 is surfaces nobody has disclosed and we refuse to guess.

SourceScopeT/dayAs ofConf.
Tokens Per Day (this index) Reported floor (disclosures only) 300.1 2026-07-15 measured
Tokens Per Day (this index) Estimated total (floor + implied estimates) ~334.7 2026-07-15 modeled
Epoch AI / Exponential View All providers (global) ~432 mid-2026 medium
OpenRouter (1% extrapolation) Global inference ~400 late-2025 low
Tomasz Tunguz (The Token Race) Global (all providers) ~88 Sep-2025 medium
a16z / OpenRouter LLM API market ~50 late-2025 medium
Epoch AI (compute-derived) Frontier lab (OpenAI) 10–100 late-2025 high

These are independent third-party estimates, included for comparison and not summed into our figures. Definitions differ: “all providers” includes multimodal/all-surface tokens; “API market” excludes captive first-party traffic. Full sources →

How the number is built

Three independent layers, each more uncertain than the last. They constrain each other; agreement across all three is what makes the headline credible.

Layer 1 · Floor

Reported floor

Only public disclosures, normalized to trillions/day at each disclosure’s period midpoint. Overlaps are removed so nothing is double-counted. A hard lower bound.

See the data →
Layer 2 · Estimate

Implied estimates

For big names that don’t disclose usable counts, we estimate by the method that fits: revenue-implied for usage-billed providers, usage-implied for free consumer products.

See the estimates →
Layer 3 · Check

Compute sanity check

An independent cross-check from the hardware side: accelerators × throughput × utilization. Demand sits at ~48% of mid-case capacity, physically comfortable.

Read the method →

What this is, and what it isn’t

It is a transparent floor-and-estimate for global AI token throughput, where every disclosure links to its source and an archived snapshot.

It isn’t an official total. Real usage is higher than the reported floor; the estimates are explicitly uncertain. We publish the lower bound and show our work rather than a single confident number.

How it’s calculated Browse the data

Sources & references

Every figure on this index is traceable to a primary source. The full citation list, with archived snapshots, lives on the Sources page.

Independent estimates

Company disclosures

Research & academic

Analyst & industry

Compute & hardware

Sources do not all measure the same thing (input vs output tokens, one provider vs all surfaces, marketplace samples vs global totals, multimodal vs text). We preserve what each reported and flag the differences. See the methodology.