The Token Ceiling

The Thesis

Nobody arguing about the AI buildout is arguing about a number. They are arguing about a clock.

The internet’s infrastructure boom had one curve to run. A household paid a flat monthly fee for access, and the tenth web page cost the network the same as the thousandth. Revenue grew as households signed up and stopped growing when they ran out. That is a single dimension, and a single dimension saturates.

This buildout is metered. Every question, every agent run, every file a machine rewrites is billed in tokens, so demand has two dimensions instead of one: how many people use it, and how much each of them uses. The two multiply. The first is well past its steep stretch — you cannot ten-times a number that already sits near everyone. The second has barely started, and nothing in its shape says where it stops.

The coverage record shows the handoff in progress. It shows the chat frame — one person, one question, one answer — flat for four years, while the agent frame, which consumes tokens by the millions per task, compounds. It shows the products doing that consuming moving out of engineering and into ordinary office work. And it shows the capacity being financed against the second curve rather than the first, which is why the announced gigawatts look either reasonable or insane depending entirely on what year you let them pay off.

This page does not forecast the payoff. It sets out what the record already contains, what the arithmetic requires, and which of the two numbers has to move for the other to make sense.

The Evidence

Start with the frame, because the frame is what changed.

Counting archive records whose headline or summary carries a given phrase, over the same Jan 1–Aug 28 window each year: “chatbot” appears in 211 records in 2023 and 195 in 2026. Four years, no compounding. “AI agent” appears in 4 records in 2023, 28 in 2024, 149 in 2025 and 465 in 2026. In 2023 the archive held one agent story for every fifty-three chatbot stories. Today it holds two and a half agent stories for every one.

The archive’s own volume nearly doubled across that span, so raw counts flatter every phrase. Normalised, the direction is the same and starker: per 1,000 archive records, chatbot mentions fell from 27 to 13 while agent mentions rose from 0.5 to 32. The chat frame did not merely stop growing. It halved as a share of what gets written about, in the four years everyone spent calling this the chatbot era.

Now the two curves, one at a time.

Curve one is late. ChatGPT reported 900M weekly active users in February 2026 and became the fastest app ever to 1B monthly users in June, with app MAUs up 62% year over year. Inside companies the number is further along: BCG data reported in July has 74% of front-line employees using AI regularly, against 51% a year earlier. Both are real growth. Neither can repeat. A number at 74% has one doubling left in it at most, and then arithmetic closes the door.

Curve two is early, and it is where the tokens are. Codex went from 1.6M weekly users in February to 5M+ by June, a six-fold rise in four months, and OpenAI’s own report on that milestone said the plain thing: knowledge workers were already about 20% of Codex users, and the tool had stopped being only a coding tool. Inside OpenAI, non-developer Codex use rose 137x for individuals while staff adoption went from roughly 40% to 97.9%. By July the combined Codex and ChatGPT Work figure was 10M people. Anthropic’s side of it runs the same way: Claude Code went from $1B of annualised revenue in November 2025 to $2.5B+ by February 2026, and by then was authoring 4% of public GitHub commits.

That is the mechanism worth holding on to. Coding is not the destination; it is the format. A machine that can observe, act and check its own work needs a shell, a file and a test to do it, so a legal review, a compliance escalation and a slide deck all arrive at the same execution surface a build does. Anthropic shipped ten agents aimed at financial firms in May — pitch decks, statement review, compliance triage. The seat-based software industry noticed before the market did: by the end of 2025, 79 of 500 software companies, more than double the year before. A vendor changes its unit of sale when its customers’ consumption stops tracking their headcount.

The two curves separate cleanly if you divide revenue by audience. Anthropic’s run rate reached $65B by the end of July, sixteen times its level thirteen months earlier, on a consumer app audience Sensor Tower put at 56M monthly users. OpenAI is on track for $40B+ against a billion. That is roughly $1,160 of annual revenue per app user against roughly $40 — a 29-fold gap between two companies selling comparable models. The gap is not a quality judgment. It is the two curves, priced: one company’s revenue tracks how many people opened an app, the other’s tracks how much work machines did.

The arithmetic that has to close. Barclays counted 46GW of announced AI data-centre capacity in November 2025, and OpenAI alone had signed for 10GW by April 2026, three years ahead of its own target. The working assumption under those plans, stated plainly in SemiAnalysis’s case for SpaceX’s 10GW, is inference at about $100B per gigawatt per year — the same $/GW arithmetic this site keeps as its source of record. Hold that rate against the announced 46GW and it implies $4.6T of annual inference revenue. Against it, the actual figure: global AI sales excluding China ran $25B in the first quarter, roughly $100B annualised, and that was reported as good news because it finally exceeded the estimated $21B of data-centre and chip depreciation in the same quarter.

Two orders of magnitude sit between the assumption and the receipt. Neither number is contested. What is contested is the year they meet — and on a curve compounding at Anthropic’s recent rate the distance is a few years, while at the first curve’s rate it is never.

The counterweight is price. In one month, OpenAI cut GPT-5.6 Luna by about 80% after serving efficiencies, then cut Sol by more than 20% to $4 per million input tokens. It is also tuning the agent harness underneath Codex specifically to cut runaway token use. Every efficiency gain pulls twice: it makes the second curve affordable, and it means revenue only holds if volume climbs faster than price falls. Cheaper tokens are the reason intensity can grow, and the reason growing intensity does not automatically become growing revenue.

And the financing is dated even when the demand is not. Morgan Stanley expects AI-tied debt issuance near $570B, more than double the prior year, and has $1.5T of the $2.9T needed coming from outside the hyperscalers’ own cash. Lenders are stretching to underwrite to them. Bain’s estimate is that AI companies will need $2T of combined annual revenue and will fall about $800B short. A coupon has a date on it. A demand curve does not.

The Companies

Nvidia is where the second curve becomes a purchase order. In its August quarter it reported revenue up 106% to $96.22B with data centre up 117%, guided to $108B, and disclosed something more revealing than either: commitments to component suppliers of $279B, against $119B one quarter earlier and $50.3B three quarters before that. Those commitments are not a forecast of chip sales. They are a wager on token consumption two years out, placed with suppliers who need the order now.

Salesforce is the seat model meeting the meter, and in August it stopped fighting. Alongside a quarter with revenue up 11% to $11.35B, it unveiled Claudeforce with Anthropic — its data and workflows served inside Claude, starting with 37 prebuilt sales skills. The stock closed up 22.6%, its second-best day ever. What the market applauded was a company agreeing that the model, not the login, is where the work now happens. Whether that trades a per-seat line for a per-token one is the open question its own dossier tracks.

SpaceX is the purest bet on curve two among listed names, and it prices it in public. Its first quarter as a public company showed AI revenue up 247% to $2.56B against $15.8B of AI capital spending — roughly six dollars out for every dollar in, deliberately. Then it closed a $60B acquisition of Cursor, buying the token consumer to sit on top of the compute it is building. The power wall is what decides whether the gigawatts behind that arrive on schedule; this page is about whether anything will be running on them.

The neoclouds — CoreWeave, Nebius, Applied Digital — are the leveraged expression of the same question, since their contracts are written against demand that has to show up on a lender’s schedule rather than a technologist’s. The capex supercycle tracks who is funding the build; the inference layer tracks who collects when the tokens flow; the chokepoint rotation tracks who keeps the dollar inside the machine itself. This page sits under all three, on the only quantity any of them depend on.

The Lenses

Ray Dalio’s debt-cycle template turns on a mismatch that has nothing to do with whether an investment is wise: debts are fixed obligations with dates attached, and the income meant to service them is not. A boom becomes a bust not when the underlying idea fails but when the cash flow arrives later than the schedule assumed. That is exactly the shape here. The demand argument in this record is strong and the demand timing argument is unresolved, and with $570B of AI-linked issuance the second one is the one that decides outcomes. Dalio’s framework does not ask whether tokens will be consumed. It asks whether they will be consumed before the interest comes due, which is a different question with a different answer.

Benjamin Graham’s answer to an unfixable intrinsic value was never a better forecast. It was a wider margin. When the inputs are two curves whose crossing year nobody can name, the disciplined response is to widen the gap you demand between price and value, not to sharpen the projection. That is also why the two-orders-of-magnitude gap above is not, by itself, an accusation. A number that has grown sixteen-fold in thirteen months closes a 46x gap quickly; a number growing 62% a year never does. Graham’s point is that you do not get to assume which one you are holding.

What Moved

The cluster tells its own story. Every entry that measures how many people use AI sits in the first half of the year and is decelerating by construction. Every entry that measures how much gets used sits later and is still multiplying. What to watch is not whether the second curve exists — the record has settled that — but whether it keeps compounding through a full round of price cuts, which is the first time falling revenue per token and rising tokens per user will be tested against each other in the same quarter.

Sources

Phrase counts are drawn from TEXXR’s archive over matched Jan 1–Aug 28 windows in each year, counting records whose headline or summary contains the phrase; the archive held 7,872 records in the 2023 window and 14,470 in 2026, which is why ratios rather than counts carry the argument. Revenue-per-user figures divide disclosed run rates by reported app audiences and are approximations across mixed reporting dates, stated as such. Article IDs resolve at texxr.com/<id>.