AI Ledger Hepburn Advisory Hepburn Advisory

Overview/Usage Ledger

The questionHow much of this usage actually gets paid for?

The tokens are real. The revenue behind them isn't all there yet.

On 2025 figures, usage runs ~7.5× ahead of the cash it generates. This ledger tracks the gap between what's used and what's paid, and whether it's closing.

List-priced notional usage vs customer-paid revenue, 2025 · not revenue

Notional usage : Revenue, 2025

7.5×

List-priced notional usage against customer-paid AI revenue. Not revenue.

Activity · 2025

~387T/day

Token volume is real and growing.

Notional value · 2025

$177B

At list price, usage would be much larger than realised revenue.

Monetisation Modelled

~27%

The share of tokens paid for. The rest is free tiers, bundles, discounts and internal use.

Why this matters — usage is the best signal of demand, but not all demand is cash. The key question is how much activity converts into durable paid revenue as subsidies and free usage normalise. The Revenue Ledger tracks how much of this usage is collected as customer cash.

Token growth over the past years Modelled series

Total tokens processed per day, by quarter. Inference token volume grew +221% year-on-year in 2025, and the trend points toward roughly 620T/day by the end of 2026. That is the clearest evidence demand is real.

Solid = reported quarters (triangulated across OpenRouter, provider disclosures, AI Ledger consensus); Q1 2026 is provisional and under revision; dotted = projected to end-2026 from the Ledger's forward token model. Directional, not measured.

Who is buying the tokens

Demand traces to named companies running production workloads. The full per-company table is in the detail section below.

Source: AI Ledger consuming-company tracker, per-company confidence rated. See the full table →

The tokens and the money don't line up

Share of all tokens against share of AI revenue. A few labs capture most of the cash on a small slice of the volume. The rest of the tokens earn little or nothing as AI sales.

Exclude ad-platform tokens

Where the AI money sits — and how exposed it is to token prices Modelled read

Each pocket of AI revenue, by who pays and how they buy. Colour = how directly that revenue moves with token usage. Green is metered (rises and falls with tokens), red is bundled into seats or subscriptions (shielded from token prices). Two reads:

If you sell AI, the green pockets are where margin compounds as usage grows; the red pockets are stickier, seat-based revenue that doesn't.
If you buy AI, green is where your bill scales with how much you use (your exposure to future token-price moves); red is predictable, bundled spend.

The effective price has barely moved Indexed

List prices for a given model have collapsed — but as usage shifts to newer premium and reasoning models, the average price actually paid per token has barely fallen.

Like-for-like = average list price of leading frontier models (OpenAI / Anthropic / Google), indexed. Effective = total AI revenue ÷ total tokens, indexed. Both Q1 2024 = 100. Solid = actual through Q4 2025; dotted = forecast to end-2026 (trend continuation).

What to watch

The open question is how much of this activity becomes durable paid revenue. Three signals show whether demand is converting. Each is shown against where it stands now.

1Token volume growth

Needs≥80%/yr

→

Now~120%/yr

Token volume is compounding well above the rate that signals durable demand.

Watch: OpenRouter throughput · provider token disclosures each quarter

2Paid token share

Needs≥35%

→

Now~27%

Paid usage must outgrow free and subsidised tokens for activity to convert into revenue.

Watch: paid-vs-free token mix as free tiers and promotional credits normalise

3Effective price per token

NeedsStable

→

Now~Flat

Any single model gets much cheaper over time, but buyers keep moving to newer, pricier models, so the average price paid per token has stayed roughly flat. That is what lets revenue keep pace with usage.

Watch: the blended price paid per token, not list-price cuts

Sources: OpenRouter API, provider earnings calls, AI Ledger multi-model consensus. See methodology →

Detail & references

Model pricing & estimated volumeper-model list prices & token estimates

Top AI natives token & ARR detail by company

Site rebuilt: 02/07/2026

ARR vs collected: revenue figures on this page are provider ARR (trailing-quarter run-rate × 4). The Overview shows $24B of customer-paid AI revenue (2025, gross). ARR and customer-paid revenue measure different things. See Methodology and the Revenue Ledger.

Methodology caveat

Projections are directional arithmetic, not forecasts. Benchmarks are dated and sourced. Token volumes default to a multi-model consensus across fourteen providers; every assumption is editable. Where a number appears on a Ledger page, that page is the canonical source. Full method on the Methodology page.