This is a model, not a forecast. It does not say what will happen. You choose the assumptions and it works out what follows from them. The settings can be set to give almost any outcome, so a result is only as good as the assumptions behind it; the starting settings are one view, not a prediction.
Anthropic and OpenAI have promised to pay for far more compute than their revenue covers today. The model starts from what each lab has committed to, taken from filings and leaks, and from your assumptions: how fast AI improves, how much more customers pay as it does, how much new compute each lab buys, and what could go wrong. From those it works out each lab's revenue, compute bill, profit and cash, quarter by quarter from October 2026 to the end of 2032, and, for Anthropic, the share price that business supports.
The charts show reported figures up to September 2026 and the model's result after that. A good way to use it is to ask what you would have to believe: for the labs to pay their bills, or for Anthropic's shares to be worth what you pay for them.
On the starting settings, Anthropic earns $155B in 2027 and $405B in 2028. The forecast stops at the end of September 2029, when the best AI model passes ECI 250: past that point the page does not guess (see Where the forecast stops). Its 2028 profit supports a share price of about $42, against the $70 to $78 range in its IPO filing. The ready-made scenarios give anywhere from about $0 to about $42, so read the share price as a test of the settings, not a target.
Whose each number is. Every number is read from a filing (Anthropic's IPO filing with the US securities regulator), taken from a press report of leaked figures, published by a named organisation, taken from the AI Futures Project's model, or assumed. Each assumption says whose it is: mine (I'm Lucas, and this is my site), or the AI agent's that built this page for me. The agent's choices have not been checked unless the page says so. Every input and its source is listed under Every number and where it comes from. Inputs were last updated on 2026-11-17.
Not investment advice. A personal research model: several of its inputs are leaks or assumptions, and it can be wrong.
On the starting settings the model is a loop:
Each link can be loosened: revenue can be set by hand, market shares can be fixed, and the settings under What could go wrong bend the loop where critics say it breaks.
The Who pays chart is separate from the loop: it shows what is published about who pays each lab and for what. Neither lab publishes revenue by industry, so the chart has none. The forecast does not use it; its sources are at the end of the table of inputs below.
A scenario sets several settings at once, starting from the starting settings and changing only what is listed for it. A scenario named after a person or a group is inspired by what they have said in public. The name is only a label: they did not make it, check it or endorse it, and this page has no link to them. Where a setting needed a number they did not give, the number is the agent's reading and is not confirmed. "Lucas's settings" is my own view, built by the agent from my description of November 17, 2026; the numbers the agent picked are not confirmed either.
| Scenario | Forecast stops | Anthropic's revenue, 2027 | 2028 | 2031 | Money raised, to 2032 or the stop: Anthropic | OpenAI | All labs, at the end of the forecast | Share price supported |
|---|---|---|---|---|---|---|---|---|
| Starting settings | End of September 2029: best model above ECI 250 | $155B | $405B | not forecast | $285B | $234B | $2.68T | about $42 (from 2028) |
| Lucas's settings | End of 2029: best model above ECI 250 | $181B | $365B | not forecast | $312B | $289B | $2.47T | about $8 (from 2029) |
| No speed-up: the old pace | End of September 2031: OpenAI runs out of money | $113B | $204B | not forecast | $446B | $305B; runs out Sep 2031 | $1.94T | about $4 (from 2030) |
| The measured speed-up keeps growing | End of June 2028: best model above ECI 250 | $183B | not forecast | not forecast | $81B | $105B | $1.72T | about $32 (from 2027) |
| OpenAI races for compute | End of September 2029: best model above ECI 250 | $155B | $405B | not forecast | $287B | $290B | $2.69T | about $40 (from 2028) |
| Both labs race for compute | End of September 2029: best model above ECI 250 | $155B | $396B | not forecast | $97B | $352B | $2.73T | about $35 (from 2028) |
| AI Futures Project | End of June 2027: all labs above $5.00T a year | not forecast | not forecast | not forecast | $81B | $51B | $7.22T | none: no full year is forecast |
| Dario Amodei | End of September 2027: best model above ECI 250 | not forecast | not forecast | not forecast | $81B | $75B | $1.46T | none: no full year is forecast |
| Ramez Naam | End of March 2031: OpenAI runs out of money | $111B | $192B | not forecast | $225B | $278B; runs out Mar 2031 | $1.30T | about $3 (from 2030) |
| Ed Zitron | End of March 2028: OpenAI runs out of money | $79B | not forecast | not forecast | $81B | none; runs out Mar 2028 | $223B | about $9 (from 2027) |
| Capability stalls | End of 2031: OpenAI runs out of money | $139B | $267B | $525B | $493B | $272B; runs out Dec 2031 | $1.50T | about $0 |
| Revenue stops following capability | End of September 2029: best model above ECI 250 | $135B | $181B | not forecast | $179B | $177B | $646B | about $6 (from 2028) |
| Slow adoption | End of September 2029: best model above ECI 250 | $87B | $113B | not forecast | $183B | $213B | $457B | about $5 (from 2028) |
| The bubble pops | End of September 2029: best model above ECI 250 | $155B | $342B | not forecast | $251B | $195B | $1.24T | about $10 (from 2028) |
| Open models take the market | End of September 2029: best model above ECI 250 | $125B | $267B | not forecast | $185B | $258B | $1.53T | about $8 (from 2028) |
| Data-centre moratorium | End of September 2029: best model above ECI 250 | $155B | $405B | not forecast | $288B | $236B | $2.68T | about $39 (from 2028) |
| Cap on training runs | does not stop | $155B | $275B | $327B | $81B | $717B | $3.11T | about $13 |
| Cap on training runs, US only | End of March 2030: best model above ECI 250 | $155B | $274B | not forecast | $81B | $300B | $1.79T | about $9 (from 2029) |
| Moratorium and training cap | does not stop | $155B | $275B | $327B | $81B | $718B | $3.11T | about $13 |
All labs: the yearly revenue pace of every AI lab together, at the end of 2032 or at the stop. Money raised: the listed raises and what was raised when cash ran low, to then (see Money raised). A scenario whose forecast stops shows "not forecast" for every later year; its share price is worked from its last full forecast year, which the table names. OpenAI's figures use Anthropic's cost shares and an out-of-date cash figure, so read them as a size, not a forecast. When the priced year shows no profit, the supported price rests on cash alone: about $8 for the cash Anthropic holds after its IPO, worked from 2031, and less where it has spent that cash or owes investors who came in later.
The top chart shows how capable Anthropic's and OpenAI's best AI models are, on the ECI score (Epoch AI's Capabilities Index). Up to today the lines are Epoch AI's measured scores, stepping up with each release. After that each lab's line follows the research compute it buys. The dashed line is a forecast for the best AI model anywhere, which the labs are measured against.
Effective compute is turned into ECI in a straight line, the way the AI Futures Project's model does it: about 14.2 points for each tenfold rise on the starting line, so that it rises 15 points a year at the old pace (their figure). The line is moved to pass through the best AI model measured on November 16, 2026 (173.4), which puts it about 6 points above the AI Futures line. Scores far above today's are that line carried on, not something a benchmark has measured.
They differ in how fast the biggest training runs grow and how fast training methods improve.
| ECI score of the best AI model | Jan 2027 | End of 2027 | End of 2028 | End of 2029 | End of 2030 | End of 2031 | End of 2032 | Effective compute, 2027 through 2032 |
|---|---|---|---|---|---|---|---|---|
| Epoch AI's trends, no takeoff | 170 | 185 | 200 | 215 | 230 | 240 | 250 | ×420,000 |
| AI Futures model: slower test case | 169 | 186 | 208 | 231 | 251 | 267 | 281 | ×837 billion |
| AI Futures model: default settings | 169 | 187 | 257 | 353 | 393 | 418 | 436 | ×1029 |
METR, an independent evaluator, reported on November 5, 2026 that Anthropic's models have improved about ×1.9 as fast since October 2025 as from March 2024 to October 2025 (its range: ×1.5 to ×2.8). Before, the best AI model gained about 15 ECI points a year. Epoch AI's public scores show less: ×1.6 for Anthropic's best released model and ×1.3 for the best released model of any lab. The page opens on ×1.9 (my choice); the setting runs from ×1, the old pace, to ×3. At ×1.9 the best AI model ends 2027 at ECI 205, against 190 at the old pace. A second setting says how that pace changes each year from today: 0 holds it (the start), above 0 it keeps speeding up (+100% doubles it every 1 year, the agent's choice for the scenario where it does), and below 0 it slows down and capability levels off. METR says its newest evidence makes "a permanent slope change look more plausible than a one-time jump".
METR guesses that most of the speed-up, perhaps 80%, is due to AI, and its measure counts Anthropic's unreleased models. On a pace that keeps rising it says "we do not see clear evidence of this" in the scores, and that other evidence, which Anthropic removed from the public report, "makes this seem more plausible". Its forecast for a future model was removed too.
An AI Futures run started from that pace. Their runs have a speed-up of their own, later on: at its default settings their run first goes ×1.9 the old pace in February 2028. I asked that for the scenarios named after forecasters the takeoff "should have already started just now", so with the speed-up above ×1 the page starts a run today from the point where it first goes that fast. At ×1.9, everything on the default run comes 1.3 years sooner. This is the agent's construction, not the AI Futures Project's, and their own forecast after METR's report is not in their public code. One more setting slows the run's clock from today: at ×1.5, what the run does in two years takes three, and the run is entered further along, so that slowed down it still goes at the speed-up set.
The prediction market as a floor. One setting lifts the line to pass through the Kalshi market's point as it stood when the page was published, ECI 185 in January 2027. The lift grows in a straight line from nothing today to its full size on that date, and stays; a line already above the point is not changed. I asked that no scenario named after a forecaster predict less progress than the market. On the default run started at ×1.9 the lift is 5 ECI points. The point is a snapshot: the marker on the chart follows the live market, the point does not.
Kalshi runs a market on when any AI model will first score above 185 on ECI. Read on Nov 17, 2026, 10:00 pm Central, it gives even odds by Jan 23, 2027, one chance in four by Dec 20, 2026 and three in four by Apr 9, 2027. The capability chart marks it with a green diamond and a bar, re-read once a minute while the page is open. On the starting settings this page's best AI model passes 185 in April 2027, about 3 months later. Part of that gap is the shape of the lines: a new release lifts the measured score in one step, while the page's line is smooth. The market is thin (4,848 contracts traded in all, with buying and selling prices up to 8 points apart), so on the starting settings it is shown and not used.
The forecast lines assume the leading lab's research compute grows by a set factor each year. On the starting settings that assumption is dropped: only Anthropic and OpenAI push the frontier, so the best AI model follows what those two actually buy. If they buy more than the forecast assumes, the best model runs ahead of the line; if less, behind. Research compute works in two ways.
Compute bought in one year shows in the next year's models, so a lab's research compute first moves its position at the start of 2029. Every other lab follows the best AI model and stays as far behind it in time as it is today, so it gets closer when the frontier slows. Each lab can also be given extra progress of its own, in ECI points a year. That starts at none for both, because history shows no lasting edge (below).
Where each lab starts. At the ECI score Epoch AI measured for its best released model on November 16, 2026 (my choice): Anthropic 173.4, OpenAI 169.0, and the best other lab (Google) 166.5. Anthropic's lead is about 3.5 months of progress. It dates from one release on November 9 and is inside Epoch AI's own range for Anthropic's score (169.9 to 177.5); on October 1 OpenAI was ahead, 169.0 against 168.5. From 2024 to the third quarter of 2026 Anthropic's best model gained 17.3 points a year and OpenAI's 17.1. Today's market shares already reflect today's models, so a lead at the start does not by itself move market share: shares move only when a lab's position changes from here. That is this page's choice. The starting scores do decide which lab has the best AI model, and the market's growth follows the best model.
The AI Futures model's rule: ten times less compute for experiments makes methods research 2.8 times slower, and how good methods are is all the research done so far, raised to the power 3.19. Chips get cheaper, so the same dollars buy more each year: the model assumes 49% more (Epoch AI's estimate for chips since 2023). Holding its position takes research compute that grows as fast as the forecast assumes; for Anthropic, in dollars, that path starts from its research compute over its first four quarters and runs $53 billion in 2027 and $141 billion in 2028.
The Revenue settings offer three ways to set it: follows capability, market share follows research compute (the starting setting, and the whole loop); follows capability, shares fixed (the market grows with capability and each lab keeps today's market share); and set by hand (each year's growth is a setting and capability changes nothing).
All AI labs together earn $208 billion a year at today's pace: Anthropic $71 billion (filed, for September 30, 2026), OpenAI $50 billion (a leak), and every other lab $87 billion. The last figure is an assumption. It comes from one market-research estimate, read second-hand, that the two labs held about 60% of model revenue in early 2026. A survey of what companies spend on model APIs alone would put every other lab at about $20 billion.
Anthropic's filing also gives $81 billion for October 31. One month in, the model on its starting settings is at $75.9 billion, below the filed figure: the starting multiplier gives slower growth than Anthropic actually had in October 2026.
The market grows with the best AI model's effective compute, by the AI Futures Project's revenue rule: revenue multiplies by a set number for each year's worth of progress at the mid-2026 pace. Their middle setting is ×5 and their range 2 to 12.5. The page starts at their low end, ×2 (my choice). Their rule is written for the leading lab alone, from Anthropic's revenue in May 2026; here it is applied to all labs together, from today's revenue.
History gives more, but only as an upper limit. A fit made for this page puts revenue against the ECI of the best AI model over six periods from 2024 to the third quarter of 2026: Anthropic's and OpenAI's revenue together grew about ×3.8 per year of progress, and Anthropic's alone about ×9.6. But the best model's score and the calendar rose together almost exactly (a correlation of 0.999), so growth from better models cannot be told apart from customers arriving over time.
A ceiling. A rule like that never stops, so revenue bends flat as it nears a ceiling: the most all labs together can earn in a year. The ceiling is a placeholder, $10 trillion a year, about 8% of world output in 2026 and about 1.6 times what the world spends on all information technology. No published estimate of this number was found.
The rule says what customers would pay for the best AI model. They do not pay it at once: firms take time to change how they work, and the more of the market AI has already won, the harder the rest is to win. So each market chases the rule's figure: the time it needs to double is the time the rule alone would take, plus an adoption delay. The delay is about 1 month at today's $208B a year, 0.40 years (about 5 months) at $1 trillion a year, and 1.2 years at $3 trillion: it gets 2 times as long each time the market doubles. A fall in demand is not delayed. A lab that plans ahead knows the delay and plans with it.
No source gives either number. Both are my choices. The delay is set so that Anthropic's 2028 revenue on the starting settings equals the middle of Manifold's prediction market on it, $419B as read on November 17, 2026 (play money, 8 traders), which I trust more than Anthropic's own projection. The doubling with size is not checked by anything, and it sets the years after 2028. Without the delay the starting settings would give $169B in 2027 and $512B in 2028, against $155B and $405B with it.
ECI is not usefulness. If what customers pay for grows more slowly than the benchmark score, the revenue rule should count less progress than the score shows. One setting does that: at 90%, a gain of 10 ECI points counts as 9. It starts at all of it. No source gives the number.
Price. While labs keep their markup over serving cost steady, price does not change the result: the cost of a task can fall as fast as it likes, as long as the price falls with it. Two settings cover price moving against serving cost: how much the price of a unit of work changes each year (starting at no change), and how strongly demand responds to price (starting at 1, where a price change leaves revenue the same). The second could not be measured: Anthropic's prices barely moved in its filed history. Its filing puts the effect of price cuts at about $600 million in one quarter and $70 million in another.
For scale, the $10 trillion ceiling is about 15% of all wages (wages are 52.6% of world output of $128 trillion; information technology is $6.4 trillion). The AI Futures model guesses the leading lab alone earns $1 trillion a year when AI can do all of a lab's coding work. A proper model of demand is meant to replace the ceiling. Their middle setting of the revenue rule is "based on recent revenue growth, adjusted downward due to expectations of slowing soon". Anthropic's own revenue has grown at a pace of about ×3.3 a year since the May 2026 figure their rule starts from. Against its own best model, rather than the best model anywhere, Anthropic's revenue grew about ×7.8 per year of progress. The starting ×2 is below both fits; the AI Futures middle, ×5, sits between them.
The adoption delay is a snapshot: it does not move when Manifold's market moves (tools/fit_adoption_delay.py in the project finds it again). Other large markets grew more slowly as they got bigger: above about $300 billion a year, growth fell by about a third to a half for each doubling. That is the agent's reading of a handful of cases, not a fit.
On price: one study of a routing service (OpenRouter, December 2025) finds almost no response, a 10% cut going with 0.5 to 0.7% more usage. Anthropic's filing says the opposite in words: "Lower costs per task drive greater usage". No token counts are filed. Read as a demand curve, the work customers buy is a level set by the best AI model, times the price to the power of minus the price sensitivity; a better model shifts the whole curve out, and that shift is the revenue rule.
Businesses and consumers are separate markets on the starting settings (my choice). Business revenue is what is paid by use, plus company plans; consumer revenue is subscriptions. Each has its own revenue rule and its own part of the ceiling, and a lab's market share is worked out in each. Business revenue has followed capability more strongly: about ×5.1 per year of progress for the two labs, against ×2.8 for consumers. The page starts business at the AI Futures low end, ×2, and consumers at ×1.5, which keeps the fitted relation between the two and is this page's choice. Starting mix: Anthropic 22% consumer (filed), OpenAI 50% (a second-hand press report), every other lab 15% (a guess). The consumer ceiling is a placeholder: $1 trillion of the $10 trillion (about 4 billion adults paying $20 a month).
| Period | Business revenue, both labs, yearly pace | Anthropic's part | Consumer revenue, both labs, yearly pace | Anthropic's part |
|---|---|---|---|---|
| 2024 | $1.3B | 22% | $2.8B | 4% |
| 2025 | $8.4B | 45% | $9.3B | 8% |
| the first half of 2026 | $35.8B | 73% | $21.6B | 30% |
| the third quarter of 2026 | $72.8B | 73% | $34.9B | 43% |
This project's fit (tools/fit_competition.py). Anthropic: paid by use as business and subscriptions as consumer, both filed. OpenAI: its total from press reports, split by the share of its revenue from business customers: 27% in 2024, 35% in 2025, 39% in the first half of 2026, 50% in the third quarter of 2026.
Consumer revenue followed capability about 0.6 times as strongly as business revenue; the starting consumer figure keeps that relation, not the fitted ×2.8, which would make consumer revenue grow faster than business revenue, the opposite of the history. These are upper limits: better models and the passing of time cannot be told apart in four periods.
Not found: OpenAI's revenue in dollars by kind of customer for any period after 2024, only rough shares from remarks to investors and second-hand reports, and for 2025 nothing at all (35% business is assumed, between 30% in early 2025 and 40% at the start of 2026; 30% to 40% moves the two growth numbers by less than 0.1). Anthropic's revenue from individuals: its subscriptions include business seats, so its consumer share here is too high. Any other lab's revenue by kind of customer. Whether consumers switch lab faster than businesses, and what it costs to serve each kind: the same is used for both.
The Revenue and bill chart shows, as green diamonds, where prediction markets put each lab's revenue: the middle guess, with a bar from the market's 10% point to its 90% point where its answers give one. They are there for comparison only; nothing in the model uses them, apart from the one 2028 market the adoption delay is pinned to. Every one of them is on Manifold, where people bet play money, with 6 to 75 traders each, so read them as a few people's guesses, not a market price. No market on either lab's revenue was found on Polymarket or Kalshi (searched on November 17, 2026). The odds were read on Nov 18, 2026, and again each time the page is published. The middle and range are worked out here from each market's answers; a point that falls in an open-ended answer such as "above $300B" is left out, not guessed.
| Market | What it asks | Traders | What the market says | This page, on the starting settings |
|---|---|---|---|---|
| Anthropic 2027 Q1 Revenue | Anthropic's revenue in January to March 2027 | 7 | middle $35B; 10% to 90%: $18B to $49B | $25B |
| Anthropic Q2 2027 revenue | Anthropic's revenue in April to June 2027 | 11 | middle $50B; one chance in ten below $33B; the top is open-ended | $33B |
| Anthropic 2027 GAAP revenue | Anthropic's revenue in 2027 | 9 | middle $240B; one chance in ten below $87B; the top is open-ended | $155B |
| Anthropic Q4 2027 revenue | Anthropic's revenue in October to December 2027 | 6 | middle $75B; no range: both ends are open-ended | $55B |
| Anthropic 2028 revenue | Anthropic's revenue in 2028 | 8 | middle $419B; 10% to 90%: $130B to $942B | $405B |
| Anthropic annualized revenue run rate by end of 2026? | The highest yearly pace of revenue Anthropic has reported by the end of 2026 | 56 | middle $98B; 10% to 90%: $84B to $119B | $87B |
| OpenAI 2026 revenue | OpenAI's revenue in 2026 | 35 | middle $38B; 10% to 90%: $35B to $42B | $36B |
| OpenAI annualized revenue run rate by end of 2026? | The highest yearly pace of revenue OpenAI has reported by the end of 2026 | 33 | middle $68B; 10% to 90%: $61B to $97B | $60B |
| Anthropic 2026 Revenue > $50B? | The chance of Anthropic 2026 revenue above $50B | 15 | 87% | $53B |
| Anthropic 2026 Revenue > $60B? | The chance of Anthropic 2026 revenue above $60B | 22 | 25% | $53B |
| Anthropic 2026 Revenue > $70B? | The chance of Anthropic 2026 revenue above $70B | 12 | 10% | $53B |
| Will Anthropic Annual Recurring Revenue be under $90bn end of 2026? | The chance of Anthropic having a yearly pace of $90B or more at the end of 2026 | 75 | 69% | $87B |
| OpenAI reaches $100B revenue in 2028? | The chance of OpenAI reaching $100B in 2028 (the market leaves open whether that is the year's revenue or the yearly pace) | 41 | 83% | $237B |
Also found and not drawn: Anthropic 2026 revenue (no middle can be read from it: most of the odds sit in an open-ended answer); Anthropic 2027 revenue (no middle can be read from it: most of the odds sit in an open-ended answer); Anthropic 2028 revenue [old buckets] (no middle can be read from it: most of the odds sit in an open-ended answer); How much revenue will OpenAI earn in 2026? (a second market on the same question with fewer traders (middle $38B)); OpenAI 2027 revenue (no middle can be read from it: most of the odds sit in an open-ended answer); OpenAI 2028 revenue (no middle can be read from it: most of the odds sit in an open-ended answer).
Past some point nobody can say what happens next, and a chart that kept going would only show the model's placeholders (my rule). So the forecast ends at the first quarter's end where either of two levels is passed, or where a lab runs out of money. From there the charts are empty, with a label that says why, and no number for a later date is shown anywhere.
On the starting settings the forecast stops at the end of September 2029: the best AI model passes ECI 250. All labs' revenue is then $2.68T a year. The share price then rests on the last full forecast year's profit (my choice): the shares are held that many fewer years, and the company is taken to be worth at the stop what a company with that profit is worth now. That is a strong assumption, since nothing after the stop is forecast. A forecast that stops before its first full year is over has no share price. Both levels can be moved or switched off under "Where the forecast stops". When revenue is set by hand there is no stop.
Each lab starts with today's market share of each market; across both that is Anthropic 34%, OpenAI 24%, every other lab 42%. A lab's pull on customers follows how its model's position changes from where it stands today: with half the effective compute that holding its position takes, it has half the pull, and its market share is its pull divided by everyone's. A lab's own model never gets worse; it only fails to keep up. Customers move gradually: each year a lab's market share covers 70% of the distance to where its pull puts it, which closes half the gap in about 7 months. Both numbers rest on a fit to six periods of history, which is suggestive, not a measurement.
Open models (ones anyone can download and run) compete by the same rule, from 4% of what customers pay today (Menlo Ventures: US business spending on models, September 2026, though they process 19% of the words). What they win pays the labs nothing. They follow the best model 8 ECI points behind (Epoch AI: four to six months) and keep that lag in time.
Margins fall as rivals catch up. Each lab's margin over what serving costs it is 86% today (Anthropic's filing; OpenAI is given the same). It holds while the lab is at least as far ahead of its best alternative, every other lab or open models, as it is today, and falls in a straight line to 50% as that lead goes to nothing: the margin of Fireworks AI, a company that hosts open models (a report of a report; none publishes its margin). Only the margin moves: customers buy about as much more as prices fall (NBER: about 1.1), so revenue stays as it was. No source ties a lab's prices to its lead; the straight line is my agent's choice.
The same fit sets Anthropic's revenue over OpenAI's against the gap between their best AI models.
| Period | Anthropic, yearly pace | OpenAI, yearly pace | Anthropic ÷ OpenAI | ECI gap (Anthropic − OpenAI) | Longest task, Anthropic ÷ OpenAI |
|---|---|---|---|---|---|
| 2024 | $0.4B | $3.7B | 0.10 | −2.0 | 0.77 |
| H1 2025 | $2.3B | $8.6B | 0.27 | −4.5 | 0.79 |
| H2 2025 | $6.8B | $17.5B | 0.39 | −4.0 | 0.70 |
| Q1 2026 | $18.9B | $22.8B | 0.83 | −3.5 | 1.44 |
| Q2 2026 | $46.2B | $26.8B | 1.72 | −2.9 | 2.89 |
| Q3 2026 | $67.8B | $40.0B | 1.70 | +0.7 | 2.97 |
Anthropic's revenue is filed. OpenAI's is from press reports of leaked figures, some second-hand; nothing by quarter was found before 2026. Scores are the average over each period of the best model released so far, from Epoch AI's files. "Longest task" is METR's measure of how long a task a model can finish; its measurements stop on 2026-04-07, so the last row lacks OpenAI's newer models.
The simple rule says revenue multiplies by a set number for each year of progress, up to a ceiling. Five other shapes are built in, three limits a government could set, and no new money. All are off to start, and they can be combined. None comes with numbers from the people who argue it, so every number in them is yours to set.
What a lab expects when these are on. A lab that plans ahead still expects the simple rule, slowed by the adoption delay and held to any adoption limit already in force. It lowers what it expects by how far the market's growth over the last year fell short of the rule's, and it never sees a fall coming. Compute committed before a fall is still owed, which is what turns a fall in demand into a need for new money, or, with no new money, into a failure.
No revenue forecast for 2030 was found from any of the people named, so the scenarios built on their arguments are shapes with this page's numbers, not their forecasts. Ramez Naam, sometimes counted a bear, is not one on revenue: he argues against an intelligence explosion, and says he can see "a trillion, a couple trillion, three trillion in revenue going to AI companies".
Serving takes its share of revenue first; research is what is left of the bill. In the first nine months of 2026 Anthropic spent about 14.5 cents of each $1 of revenue on serving customers (its "cost of revenue", which it says is almost all compute) and about 42 cents on all other compute. The filing does not split compute by use, so research here is total compute less cost of revenue. If serving ever needs more than the bill, the lab buys the extra at short notice, up to the growth limit below. When revenue follows capability, research compute is never cut below the quarter before to make room for serving.
Revenue needs compute to serve it. Demand is what customers would pay; revenue is the part the lab has the compute to serve. If there is too little, research is cut back first, down to the previous quarter's level, and then customers go unserved. That revenue is lost for the quarter; the customers stay, and do not go to a rival in the meantime.
Anthropic's IPO filing lists $900.9 billion of compute commitments, $451.6 billion of it before 2031, but no full year-by-year schedule. The model builds one from three pieces.
On top of this, the filing estimates Anthropic will pay $150 to $175 billion under compute deals with no fixed dollar amount, mostly Google chips to be delivered from 2028 to 2031, with no payment dates. The model counts the middle, $162.5 billion, paid evenly over 4 years from 2028. Both are settings.
| Year | Dated in the filing | SpaceX (short-term) | Undated, spread by the model | Agreed, no fixed price | Signed or agreed |
|---|---|---|---|---|---|
| 2027 | 29.2 | 37.2 | 29.5 | 0.0 | 68.0 |
| 2028 | 44.4 | 37.2 | 63.5 | 40.6 | 148.5 |
| 2029 | 47.8 | 33.6 | 80.5 | 40.6 | 168.8 |
| 2030 | 45.1 | 0.0 | 80.5 | 40.6 | 166.2 |
| 2031 | not dated | 0.0 | 125.5 (assumed) | 40.6 | 166.2 |
| 2032 | not dated | 0.0 | 125.5 (assumed) | 0.0 | 125.5 |
$ billions, starting settings. Only the first column is read straight from the filing.
After 2030 the filing gives no years: $449.3 billion is left for 2031 to 2045 (this page's subtraction; the filing says substantially all of it falls in those years). The model assumes payments carry on at the 2030 rate in 2031 and 2032. That is an assumption: an even spread over the 15 years would be about $30 billion a year.
The model runs by quarter from October 1, 2026, and the filing dates payments by year. Each year's dated payments are split evenly over its four quarters. 2027 is the exception: the filing's total for the first 12 months fixes how much of 2027 falls in January to September (a setting), and that part is split evenly over three quarters. The undated payments follow the rising line month by month. OpenAI's leaked plan is by year; inside a year its spending is taken to grow each quarter at the pace it grows from one year to the next. None of these splits is in a filing.
The filing says the Google chips may be bought with five-year financing, which would spread the payments over more years, and that the leases that have not started yet run about 10 years on average.
OpenAI files nothing. Its bill is a leaked plan for total compute spending through 2030, spread over the years in the shape of its leaked training-compute plan and held at its 2030 level in 2031 and 2032. It is a plan, not signed commitments.
Everything above is already signed or agreed. How much more a lab signs is its own choice, between profit now and more compute to earn more later, so the page lets you pick a strategy.
Compute added in one year stays on the bill in later years. A last year for the strategy can be set, to look at "spend now, profit later", and a lab whose cash runs low raises money (see Money raised). When revenue is set by hand, extra compute only costs more. When revenue follows capability, research compute pays back a year later through a better model, a larger market share and, because only the two labs push the frontier, faster progress for the whole market.
Why the plan rests on the serving cost. A lab's operating margin is whatever is left after it decides how much research to buy, so it is the wrong number to hold steady. What a lab can count on is the gap between what customers pay and what serving them costs. Today $1 of revenue costs 14.5 cents to serve (customers pay about 6.9 times the cost) and leaves 38 cents after serving and all other costs. In the first nine months of 2026 Anthropic spent about 110% of that on research and other compute, which is why its operating margin was −4%. The serving cost has not been steady: 71 cents per $1 in the first nine months of 2025 and 48 cents in the last quarter of 2025. The Margins settings let it drift.
Prices above the usual one. The usual price is what Anthropic pays under its multi-year deals. SpaceX compute costs ×3 that, and compute bought at short notice, beyond the committed bill, costs ×2 (both my choices). SemiAnalysis puts SpaceX's price at up to $50 billion a gigawatt a year, against about $12 billion for multi-year rental; cloud price lists put short notice at about ×1.5 to ×2.5 the price of committed compute. What is paid above the usual price buys no compute. Anthropic's 2026 costs per $1 of revenue mix the two prices, so the model turns them into compute at the usual price: about 23% less than they cost in 2026.
How fast it can grow. Data centres, chips and power take time to build, so all the compute a lab pays for can grow at most 3 times a year, applied quarter by quarter (at most 1.32 times the quarter before). Compute already signed is always delivered. Nothing fixes the figure: Anthropic's own compute spending grew 2.9 times in 2025 and 4.3 times in the first nine months of 2026, the second with short-term compute its filing says "costs significantly more".
How much the world builds. A rough yardstick, and the weakest data on the page: no public source gives the world's AI compute year by year. The page counts gigawatts of AI data centres: 15, 29 and 62 at the end of 2025, 2026 and 2027, from the AI Futures Project's forecast of April 2025 (a forecast, not a count). From 2028 it grows ×1.36 a year, which reaches McKinsey's middle case for 2030, 156 gigawatts (read second-hand), and 289 by the end of 2032. A gigawatt is taken to cost $12.5 billion a year to rent, the middle of the range Dario Amodei gave in February 2026; that figure is the agent's choice and is not confirmed. On the starting settings Anthropic's bill is 11% of the world's AI compute today and 20% in 2028; OpenAI's is 17% and 14%.
The price of new compute. Compute a lab has already signed for keeps its price. New compute is bought at the going price, which rises when buyers want more than the world builds (my idea: scarce compute should cost more through supply and demand; the rule is the agent's own). Each quarter the model adds up what is wanted. Each lab wants what its strategy asks for. Everyone else spends a set number of dollars whatever the price: 67% of what all the world's AI compute costs at the usual price (the agent's choice). If together they want more than the compute not already signed for, the price rises until the two match, and a lab gets less compute for its money. Research gives way first, then customers go unserved, except that a lab never drops customers it served the quarter before to keep its research going (the agent's rule). On the starting settings the price is above the usual price in 17 of the 25 quarters, and ×1.38 at its highest (in 2030 Q4).
The two sources do not sit well together: growth drops from ×2.1 a year to ×1.36 at 2028. Buttons give McKinsey's low and high cases (110 and 237 gigawatts in 2030, ×1.21 and ×1.56 a year, derived here from the gigawatts each case adds) and the AI Futures pace carried on. McKinsey's page could not be fetched. Other reports put a gigawatt-year at $8.5 billion (Epoch AI, cost of ownership) to $20 to $25 billion. For comparison, Epoch AI put OpenAI at 10% to 15% of the world's working AI compute at the end of 2025, and Anthropic at about two thirds of OpenAI. The model does not check that all labs together stay within what the world can build.
The share for everyone else: the two labs hold 28% of the world's AI compute today on this page's count, and the AI Futures Project puts them at 34% at the end of 2027, which leaves about two thirds for everyone else. Four things are left out: everyone else spends the same dollars whatever the price; the price never goes below the usual one, so there is no cheap compute in a glut; a higher price does not make the world build more; and a lab does not raise its spending to keep its customers served. The last two make scarce compute look worse than it would be. The secure-a-share strategy sets its dollars at the usual price, so when the price is high it gets less than the share it aimed for.
Each lab lives on its cash, its revenue and the money it raises. If that still cannot pay the compute it owes, it runs out of money and the model stops there (see Where the forecast stops). Anthropic starts with its cash at September 30, 2026, and its IPO is its first raise. The raises each lab is taken to make at the start, all changeable under Money raised:
| Lab | Raise | When | Amount | Where it comes from |
|---|---|---|---|---|
| Anthropic | IPO and private placement | November 2026 | $81B | Filed: the filing's cash after the IPO at $74 a share, less its cash at September 30. Pricing was planned for November 18, 2026. |
| OpenAI | Round now closing | November 2026 | $15B | Reported (CNBC, November 12, 2026): about $15 billion at $1.25 trillion. It aimed for about $20 billion; the rest is not counted. |
| OpenAI | IPO | July 2027 | $60B | The date: where Polymarket and Kalshi put even odds. The amount is my choice: no outlet reports one. |
Raising when cash runs low (under Strategy). When a quarter would leave a lab's cash below 6 months of its spending, it raises what brings it back up, up to 1× its yearly revenue in any 12 months, the listed raises included, and only while its revenue is above the same quarter's a year before. When growth stops, so does the money. The multiple is my choice; the floor is the agent's choice, not confirmed. For scale: OpenAI raised about 11 times its revenue in 2025 and about 9 times in March 2026, Anthropic about 1.4 times in May 2026, and OpenAI's round now closing is about 0.4 times. A lab counts on what it could raise when it decides how much new compute to sign.
No new money (under What could go wrong) keeps only money already raised, Anthropic's IPO, and turns off raising when cash runs low. Ed Zitron's scenario uses it.
What I could not find. OpenAI's cash after March 2026: the model starts it at $73 billion, the "more than $73 billion" reported for the end of March, so what it spent since and the later payments of its March round are not in it. The size of OpenAI's IPO: no outlet reports one, and Polymarket's market on it is too thin to use. Any plan by Anthropic to raise after its IPO: none was found; its $15 billion credit line and up to $15 billion more from Amazon are not counted.
Anthropic's profit is close to zero and swings with how much research it buys, so valuing it on today's profit does not work. The page works backward from a share price instead.
The "share price supported" runs the same steps the other way: the price at which a forecast's profit earns the return wanted. Three inputs are placeholders, not estimates: the return you want (10% a year), the value per $1 of operating profit (23), and the margin cap (30%). The compute counted is everything the strategy has committed for that year, at the price actually paid (my choice). A setting can count instead only what it takes to stand still: serving that revenue plus research compute held at the year before's level, or the signed bill if that is more. So the page shows what a price assumes about growth, not whether that growth will happen.
The forecast's last full year on the starting settings is 2028 (see Where the forecast stops), so that is the year priced. By its end the best AI model is Anthropic's, at ECI 233 against 234 on the forecast's line, with OpenAI's about 3 months behind. Anthropic's market share has gone from 34% to 34% and OpenAI's from 24% to 20%, of a market at a pace of $1.7 trillion a year. Part of that shift is the kinds of customer, not the models: OpenAI sells half to consumers, the slower market (with one kind of customer the same run ends at 33% and 24%). In 2028 Anthropic buys $134 billion of research compute and OpenAI $102 billion; the limit of 3 times a year holds Anthropic's bill back in 4 quarters, from Q3 2028 to Q3 2029.
On these settings the share price is unstable. Both labs plan one year ahead and put 100% of what expected revenue leaves into research compute, which plans to break even. So Anthropic's 2028 profit is not a measure of how well it did: it is the gap between the revenue that came in and the revenue it planned for when it committed the compute, a small difference between two large numbers. Anthropic planned for $344B of revenue in 2028 and gets $405B; its compute bill is $184 billion and its operating margin 7%, which supports about $42 a share. Putting 80% of what is left into research instead gives about $56, and 110%, as today, about $27. Counting only what it takes to stand still gives about $80.
It follows that on these settings buying less compute raises the share price. If Anthropic signs nothing more at all, its 2028 revenue is the same, $405B (research compute moves the models only a year later), and the supported price is about $75 against $42; on compute first, about $77. Success moves the price much less than the plan does: give Anthropic 3 ECI points a year of extra progress and its 2028 revenue goes to $680B and its market share from 34% to 48%, while the supported price goes to about $157. These follow from valuing the company on one year's profit while its plan aims at no profit. They are not findings about the labs. For a share price that means something on this basis, the plan has to keep a profit (research below 100% of what is left), or the strategy has to stop before 2028.
Each row changes one thing from the starting settings and shows Anthropic in 2028, the year the starting settings are priced from, so the rows compare. These are "if the rule holds" numbers, not predictions. The compute bill includes anything paid above the usual price: for SpaceX, at short notice, and when compute is scarce. A row whose forecast stops before the end of 2028 shows no numbers.
| What is changed | Revenue in 2028 | Market share, end of 2028 | Best AI model's ECI, end of 2028 | Compute bill, 2028 | Operating margin, 2028 | Share price supported |
|---|---|---|---|---|---|---|
| Starting settings | $405B | 34% | 233 | $184B | +7% | about $42 |
| Research gets 80% of what is left | $405B | 34% | 233 | $172B | +10% | about $56 |
| Research gets 110% of what is left, as today | $405B | 34% | 233 | $198B | +4% | about $27 |
| Share price counts only what it takes to stand still | $405B | 34% | 233 | $184B | +7% | about $80 |
| Compute can grow at most ×2 a year | $405B | 34% | 233 | $166B | +12% | about $63 |
| No limit on compute growth | $405B | 34% | 233 | $191B | +6% | about $35 |
| Anthropic gains 3 ECI points a year of its own | $680B | 48% | 245 | $226B | +20% | about $157 |
| OpenAI gains 3 ECI points a year of its own | $299B | 21% | 237 | $180B | −8% | about $9 |
| No adoption delay | $512B | 34% | 232 | $201B | +13% | about $86 |
| Adoption delay the same at every size | $347B | 34% | 233 | $178B | +1% | about $15 |
| SpaceX kept to its end, in 2029 | $406B | 35% | 233 | $208B | +2% | about $16 |
| One kind of customer | $438B | 33% | 232 | $203B | +6% | about $41 |
| Consumer revenue rule ×2, like business | $433B | 33% | 232 | $201B | +6% | about $41 |
| Consumer revenue rule ×2.8 (the fitted number) | $470B | 33% | 232 | $219B | +6% | about $42 |
| Anthropic commits compute the same quarter | $409B | 36% | 234 | $216B | +0% | about $10 |
| Anthropic commits compute two years ahead | $405B | 34% | 233 | $154B | +15% | about $75 |
| Anthropic signs nothing more | $405B | 34% | 233 | $154B | +15% | about $75 |
| Anthropic follows the forecast's compute path | $406B | 34% | 233 | $190B | +6% | about $36 |
| Anthropic on compute first | $406B | 34% | 233 | $153B | +15% | about $77 |
| Business revenue rule ×5 (the AI Futures middle), consumer ×2.8 | $803B | 34% | 233 | $242B | +23% | about $212 |
| The same, with the forecast never stopped (what the model gives past the stop: not to be relied on) | $803B | 34% | 233 | $242B | +23% | about $212 |
| Business revenue rule ×3.8 (the two labs' history) | $711B | 34% | 232 | $234B | +20% | about $166 |
| Business revenue rule ×9.6 (Anthropic's history) | $892B | 35% | 233 | $258B | +24% | about $248 |
| Ceiling $5 trillion | $376B | 34% | 233 | $181B | +5% | about $30 |
| Prices fall 10% a year, price sensitivity 2 | $445B | 34% | 232 | $207B | +6% | about $41 |
| Everyone else spends 100% of what the world's compute costs: compute is always scarce | $405B | 34% | 233 | $191B | +6% | about $35 |
| Revenue set by hand | $216B | not modelled | not modelled | $148B | −16% | about $8 |
| What is changed | Revenue | Market share | Best AI model's ECI | Compute bill | Operating margin | Share price supported |
|---|---|---|---|---|---|---|
| The two labs start level | $405B | 34% | 233 | $184B | +7% | about $43 |
| Start as measured on October 1 (OpenAI just ahead) | $406B | 34% | 229 | $184B | +7% | about $43 |
| SpaceX cancelled: nothing paid from April 2027 | $405B | 34% | 233 | $184B | +7% | about $43 |
| SpaceX and short notice at the usual price | $404B | 34% | 232 | $185B | +7% | about $40 |
| Consumer ceiling $3 trillion | $401B | 34% | 233 | $184B | +7% | about $41 |
| Customers move at once | $399B | 31% | 233 | $184B | +7% | about $39 |
| Customers move 40% of the way each year | $406B | 35% | 233 | $184B | +7% | about $43 |
| Share follows capability less: 0.3 | $407B | 35% | 233 | $184B | +7% | about $43 |
| Share follows capability more: 2 | $403B | 33% | 233 | $184B | +7% | about $41 |
| OpenAI on compute first | $405B | 34% | 233 | $184B | +7% | about $42 |
| Ceiling $20 trillion | $418B | 34% | 233 | $188B | +8% | about $45 |
| Prices fall 10% a year, price sensitivity 1 | $405B | 34% | 232 | $186B | +7% | about $41 |
| Research compute does not speed up methods (×1) | $405B | 34% | 233 | $184B | +7% | about $43 |
| Other labs push the frontier too: the forecast holds | $405B | 34% | 233 | $184B | +7% | about $43 |
| Everyone else spends 40% of what the world's compute costs | $405B | 34% | 233 | $181B | +8% | about $46 |
| Shares fixed | $408B | 35% | 234 | $184B | +8% | about $44 |
Filed: read directly in an SEC filing. Leak: a press report of leaked or investor documents, not checked against a primary source. Assumed: a modelling choice, adjustable on the page. Reported: a figure published by a named organisation, often read second-hand; the source note says which. Their model: a setting or an output of the AI Futures Project's model, which is a forecast, not a measurement. Reference: shown for context, not used in the arithmetic. Line numbers refer to a text conversion of Anthropic's Form S-1/A, Amendment No. 2, filed November 10, 2026.
| Input | Value | Kind | Source |
|---|---|---|---|
| Anthropic | |||
| Anthropic scheduled payments, Oct-Dec 2026 | $5.231B | filed | S-1/A No. 2, contractual obligations table, lines 10343-10350 |
| Anthropic scheduled payments, 2027 | $29.165B | filed | S-1/A No. 2, contractual obligations table, lines 10343-10350 |
| Anthropic scheduled payments, 2028 | $44.357B | filed | S-1/A No. 2, contractual obligations table, lines 10343-10350 |
| Anthropic scheduled payments, 2029 | $47.752B | filed | S-1/A No. 2, contractual obligations table, lines 10343-10350 |
| Anthropic scheduled payments, 2030 | $45.07B | filed | S-1/A No. 2, contractual obligations table, lines 10343-10350 |
| Anthropic short-term commitments at Sept 30, 2026 | $17.4B | filed | S-1/A No. 2, lines 10372-10374 |
| Anthropic short-term commitments signed after Sept 30, 2026 | $6.6B | filed | S-1/A No. 2, lines 10381-10383 |
| Anthropic compute under short-term commitments, July to September 2026 (the filing: it 'currently costs significantly more than comparable compute procured under our multi-year commitments'; no price is given) | $5B | filed | S-1/A No. 2, lines 7497-7500, 7527-7531, 9369-9371 |
| Anthropic compute under short-term commitments, January to September 2026 | $6.5B | filed | S-1/A No. 2, lines 8316-8319 |
| SpaceX compute for Anthropic: the most the amended deal can come to, into 2029 | $120.5B | filed | S-1/A No. 2, lines 10419-10426: 'we agreed to spend an aggregate of up to $120.5 billion over terms extending into 2029', with SpaceX.AI LLC (the 'SpaceXAI Arrangements', entered into and amended May to October 2026). 'Following initial non-cancelable periods ending in January 2027 and March 2027 under the respective agreements, each agreement may be cancelled by either party upon 90 days' advance written notice.' The filing does not say that its short-term commitments are this deal; that is this page's reading, backed by Bloomberg (Oct 29, 2026, via Brett Kerr, https://bretkerr.substack.com/p/the-bridge-toll-what-anthropics-amended), which names Anthropic as the SpaceX customer paying $5.0B in July to September 2026, Anthropic's filed short-term compute for that quarter. |
| SpaceX compute for Anthropic: payments a month at full delivery (from December 2026) | $3.1B a month | filed | S-1/A No. 2, lines 10422-10425: 'approximately $3.1 billion in the aggregate per month once all capacity has been delivered, which we expect in December 2026, with different fees during capacity ramping periods.' |
| Price of SpaceX compute, against the usual price (the price of Anthropic's multi-year deals) | 3 × the usual price | assumed | My choice, 2026-11-17. SemiAnalysis ('SpaceX 10GW in 2027', Aug 7, 2026) says SpaceX prices its compute 'at a huge premium – up to $50B/GW/year', against about $12B a gigawatt a year at a 'conservative rental pricing rate of $3/GPU-hr': ×2.5 to ×4. A summary of a Nov 3, 2026 podcast with SemiAnalysis's Dylan Patel quotes '$40B+ per gigawatt' for SpaceX's short-term compute against '$12-15B of typical five-year contracts', about ×3 (the interviewer's comparison). Anthropic's filing says only that short-term compute 'currently costs significantly more'. The model takes all of Anthropic's filed short-term compute to be SpaceX's. |
| Price of compute bought at short notice (beyond the committed bill), against the usual price | 2 × the usual price | assumed | My choice, 2026-11-17. No source gives what a lab pays. Cloud list prices put compute at short notice at about ×1.4 to ×2.5 the price of committed compute: AWS on-demand against its 1-year plan (×1.4) and 3-year plan (×2.2), and CoreWeave's 'up to 60%' discount for committed use (at most ×2.5), read through vendor blogs. Silicon Data's B200 index showed no premium for spot over a 12-month term in September 2026. |
| Anthropic total commitments | $900.9B | filed | S-1/A No. 2, lines 10288-10289 |
| Anthropic commitments falling before Dec 31, 2030 | $451.6B | filed | S-1/A No. 2, risk factor, lines 2335-2336 |
| Anthropic commitments due in the 12 months after Sept 30, 2026 | $70.7B | filed | S-1/A No. 2, line 10386 |
| Anthropic commitments still terminable or conditional | $186.2B | filed | S-1/A No. 2, lines 10296-10299 |
| Conditional part of the next-12-month payments | $4.2B | filed | S-1/A No. 2, line 10387 |
| Anthropic revenue, nine months to Sept 30, 2026 | $33.242B | filed | S-1/A No. 2, statements of operations data, lines 1571-1600 |
| Anthropic compute and infrastructure expense, same nine months | $18.808B | filed | S-1/A No. 2, lines 1571-1600 |
| Anthropic total operating expenses, same nine months | $34.502B | filed | S-1/A No. 2, lines 1571-1600 |
| Anthropic cost of revenue, same nine months. Used as the compute that serves paying customers; slightly too high, because it also holds some staff and overhead | $4.812B | filed | S-1/A No. 2, Note 8, lines 21942-21991 (p. F-41): "Substantially all of cost of revenue consists of compute and infrastructure costs associated with customer-serving workloads". No dollar split of compute by use is given, so research compute is total compute minus this |
| Yearly change in serving compute per dollar of revenue (price cuts push it up, efficiency pushes it down) | 0 % a year | assumed | Assumed zero. No data: the filing gives no price or efficiency trend. Adjustable |
| Anthropic cost of revenue in 2025 | $2.773B | reference | S-1/A No. 2, Note 8, lines 21942-21991 (p. F-41). Filed. Not used in the arithmetic |
| Anthropic revenue in 2025 | $4.585B | reference | S-1/A No. 2, lines 1571-1600. Filed. Not used in the arithmetic |
| Anthropic run-rate revenue at Oct 31, 2026 | $81B a year | reference | S-1/A No. 2, lines 11458-11459: '$81 billion as of October 31, 2026, up $10 billion from September 30, 2026'. Checks the model's first month; it starts on Oct 1, 2026 from the Sept 30 figure |
| Anthropic cash and short-term investments at Sept 30, 2026 | $131.845B | filed | S-1/A No. 2, balance sheet data, line 1772 |
| Anthropic cash after the IPO and private placement, at the $74 midpoint | $212.684B | filed | S-1/A No. 2, balance sheet data, pro forma as adjusted column, lines 1772-1774. The offering has not priced |
| Date of Anthropic's IPO and private placement (pricing planned for the evening of 2026-11-18) | 2026.880 | reported | CNBC 2026-11-10, from the updated filing: pricing 2026-11-18, trading 2026-11-19 on Nasdaq. The money it raises is anth.cash_after_ipo less anth.cash (the filing's pro forma at the $74 midpoint), counted in the model's first quarter |
| Anthropic leases signed but not yet started (inside the total) | $451.2B | reference | S-1/A No. 2, lines 10375-10378. Start 2026-2028, average term about 10 years |
| Anthropic commitments signed after Sept 30, 2026 (inside the total) | $105.7B | reference | S-1/A No. 2, lines 10379-10381 |
| Anthropic deals with no fixed dollar amount, the filing's low estimate (outside the $900.9B total) | $150B | filed | S-1/A No. 2, lines 10290-10292 |
| Anthropic deals with no fixed dollar amount, the filing's high estimate (outside the $900.9B total) | $175B | filed | S-1/A No. 2, lines 10290-10292 |
| Amount counted for the deals with no fixed dollar amount | $162.5B | assumed | Assumed: the middle of the filing's $150B to $175B estimate (S-1/A No. 2, lines 10290-10292). Adjustable from $0 to $175B |
| Years over which the deals with no fixed dollar amount are paid, starting in 2028 | 4 years from 2028 | assumed | Assumed; the filing gives no payment dates. The largest part (Google chips, $125B to $150B) is delivered 2028 to 2031 (S-1/A No. 2, lines 10406-10413) and may be bought with five-year financing, so payments could last about nine years. The AMD part (over $20B) starts in 2027; the model starts everything in 2028. Adjustable |
| Reinvest strategy: operating profit kept, as a share of revenue | 0 % of revenue | assumed | Assumed. Starts at break-even. Used only when the reinvest strategy is chosen, and for OpenAI too. Adjustable |
| Compute-first strategy (OpenAI's starting strategy): least yearly growth of the compute bill | 31 % a year | assumed | Assumed. Starts at the yearly growth in OpenAI's leaked training-compute plan for 2027 to 2030, derived from the three OpenAI training-compute figures below (about +31%). Adjustable |
| Plan-ahead strategy (Anthropic's starting strategy): how many years ahead compute is committed | 1 years | assumed | Assumed; no lab publishes this. At 1, a year's compute is committed at the start of the year before (12 to 24 months ahead). The filing's one hint: of the leases not yet started, payments begin within about two years. At 2 (24 to 36 months ahead), 2028's compute is only what is signed today. Adjustable, 0 to 3 |
| Signed compute paid in each year after 2030, against what is paid in 2030 | 100 % of the 2030 rate | assumed | Assumed. The filing gives one total for 2031 to 2045 ($449.3 billion), no years, and says leases not yet started average about 10 years. The model pays at the 2030 rate until the total is used up (into 2034 on the starting settings). An even spread over 15 years would be about $30 billion a year |
| Anthropic purchase commitments at Dec 31, 2025 (ten months before the totals above) | $54.559B | reference | S-1/A No. 2, lines 23627-23644 |
| Anthropic leases signed but not yet started at Dec 31, 2025 | $50.2B | reference | S-1/A No. 2, lines 21612-21626 |
| Anthropic compute and infrastructure expense, 2024 | $2.522B | reference | S-1/A No. 2, lines 8507-8521 |
| Anthropic compute and infrastructure expense, 2025 | $7.326B | reference | S-1/A No. 2, lines 8507-8521 |
| Anthropic compute and infrastructure expense, nine months to Sept 30, 2025 | $4.381B | reference | S-1/A No. 2, lines 8301-8308 |
| Share of 2027 scheduled payments that fall in Jan-Sep | 0.7 share | assumed | Assumed. An even year would be 0.75; payments are rising, so slightly less. Only affects how the next-12-month figure is matched |
| Months until unscheduled payments reach their plateau | 27 months from Oct 2026 | assumed | Assumed. Default is the end of 2028 because the filing says the unstarted leases begin between 2026 and 2028. Adjustable |
| Anthropic revenue, nine months to Sept 30, 2025 (used only to state the last twelve months' revenue) | $2.545B | reference | S-1/A No. 2, lines 1571-1600. Filed |
| Midpoint of the IPO price range of $70 to $78 | $74 a share | filed | S-1/A No. 2, cover page and lines 1526-1528 |
| Bottom of the IPO price range | $70 a share | reference | S-1/A No. 2, cover page. Filed. Used only to shade the range on the share-price chart |
| Top of the IPO price range | $78 a share | reference | S-1/A No. 2, cover page. Filed. Used only to shade the range on the share-price chart |
| Shares outstanding after the IPO and the private placement (13.21B if the underwriters take their extra shares) | 13.087 B shares | filed | S-1/A No. 2, lines 1232-1239 |
| Those shares plus shares already promised: Amazon's convertible notes, donation pledges and acquisitions | 15.235 B shares | filed | S-1/A No. 2, capitalization table, lines 1442-1524: Class A total 7,430,143,944 plus Class N total 7,804,580,273. The sum is mine; the filing states no single total |
| Employee stock units outstanding, including grants after Sept 30 | 1.074 B shares | filed | S-1/A No. 2, capitalization table, lines 1442-1524 |
| Employee stock options outstanding (all of them; the model takes off the shares their exercise money would buy back at the price entered) | 1.774 B shares | filed | S-1/A No. 2, capitalization table, lines 1442-1528 |
| Average price employees pay to exercise those options | $2.91 a share | filed | S-1/A No. 2, lines 1526-1528 |
| Years from buying to the valuation date (IPO to the end of 2031, rounded) | 5 years | assumed | Assumed. The filing computes one figure as of November 19, 2026, suggesting the IPO falls about then. The company is valued on its 2031 profit; the model runs to 2032 so compute bought in the valuation year still shows its payoff |
| Return the buyer wants each year | 10 % a year | assumed | Placeholder, not an estimate. Roughly the long-run return of US shares. Adjustable |
| Company value per dollar of operating profit in the year the price is worked from (2031 unless the forecast stops sooner) | 23 × operating profit | assumed | Placeholder, not an estimate. The middle of Microsoft, Alphabet and Nvidia on 2026-11-13 (20× to 24×; see the reference table). Adjustable |
| Ceiling on the 2030 operating margin (the long-run margin) | 30 % of revenue | assumed | Placeholder, not an estimate. No filing sets it; Alphabet, Microsoft and Nvidia run at 33%, 47% and 65%. Adjustable |
| Stock-unit pay the filing says it will expense in the IPO quarter (none was expensed before) | $14.2B | reference | S-1/A No. 2, lines 10573-10580. Filed. Not used in the arithmetic |
| Stock-unit pay not yet expensed at Sept 30, 2026 (includes the IPO-quarter amount) | $32.3B | reference | S-1/A No. 2, lines 10566-10569. Filed. Not used in the arithmetic |
| Plan-ahead strategy: share of what expected revenue leaves, after serving and other costs, that buys research compute | 100 % of what is left | assumed | Set 2026-11-16 after my review: the lab buys compute for its expected revenue at today's serving cost per dollar and puts the rest into research. 100% plans to break even. Anthropic's own figure for the nine months to Sept 30, 2026 is about 110% (research and other compute 42% of revenue; revenue left 38% after serving and other costs), its operating loss of 4%. Adjustable |
| The most a lab's compute spending can multiply by in one year | 3 × a year | assumed | A starting figure; nothing fixes it. Anthropic's compute expense grew 2.9 times in 2025 and 4.3 times in the nine months to Sept 30, 2026 against the same months of 2025 (S-1/A No. 2, quarterly table), the latter with $5.0B of short-term compute the filing says 'costs significantly more'. In physical compute: AI chips sold 2.5 times a year (Epoch AI), the AI 2027 forecast 3.4 times for the leading company. Signed compute is always delivered. Adjustable; 0 = no limit |
| Anthropic cost of revenue, nine months to Sept 30, 2025 | $1.796B | filed | S-1/A No. 2, Note 8, lines 21942-21991 (p. F-41). Filed. Not used in the arithmetic: it shows the serving cost per dollar of revenue was 71 cents then, against 14.5 cents a year later |
| Revenue lost to price cuts, second to third quarter of 2026 | $0.6B | filed | S-1/A No. 2, quarterly results discussion, lines 9351-9357: 'a decrease of approximately $600 million from price reductions introduced as part of new model versions and contractual discounts'. Not used in the arithmetic |
| Revenue lost to price cuts, third to fourth quarter of 2025 | $0.07B | filed | S-1/A No. 2, quarterly results discussion, lines 9351-9357: 'a decrease of approximately $70 million'; and 'Price changes did not have a material impact on the other quarterly revenue trends presented'. Not used in the arithmetic |
| Anthropic run-rate revenue at Sept 30, 2026 | $71B a year | filed | S-1/A No. 2, lines 11458-11459: '$81 billion as of October 31, 2026, up $10 billion from September 30, 2026', so $71 billion. Run-rate is the last 28 days of paid-by-use revenue times 13 plus monthly subscription fees times 12. The pace at the model's start, Oct 1, 2026 |
| Anthropic compute and infrastructure expense, third quarter of 2026 | $9.863B | filed | S-1/A No. 2, quarterly statements of operations. Includes $5.0 billion of short-term compute. The starting point for the limit on compute growth |
| Anthropic: subscriptions as a share of revenue, third quarter of 2026 | 22 % of revenue | filed | Derived from filed figures: subscription-based revenue of $6.936B in the nine months to Sept 30, 2026 (S-1/A No. 2) less $3.2B in the first half (first S-1 filing, rounded there), over the quarter's revenue of $16.956B. Subscriptions serve 'both enterprises and prosumers', so this overstates consumer revenue, which the filing does not give. Used as the starting consumer share. Adjustable |
| Secure-a-share strategy: the share of the world's AI compute Anthropic secures | 14 % of the world's AI compute | assumed | Used only by the secure-a-share strategy, which is off to start. AI Futures Project compute forecast (April 2025): Anthropic's share goes 'from 4% to 14%' from the end of 2024 to the end of 2027. Epoch AI (2026-05-20) counted Anthropic at about 1.19 million H100-equivalents at the end of 2025 against OpenAI's 1.74 million, which with OpenAI at 10% to 15% puts Anthropic near 7% to 10%. Announced capacity (up to 5 GW Amazon, about 3.5 GW Google chips from 2027, 1 GW Azure) is 9.5 GW. Adjustable |
| OpenAI | |||
| OpenAI planned compute spend through 2030, low | $600B | leak | CNBC via Reuters/CGTN, 2026-02-21: 'roughly $600 billion in total compute spend through 2030'. https://news.cgtn.com/news/2026-02-21/OpenAI-sees-600-billion-in-compute-spending-through-2030-Reuters-1KWAar1MJLW/p.html |
| OpenAI planned compute spend through 2030, middle | $665B | leak | The Information via The Decoder, 2026-02-21: 'expects to spend $665 billion training and operating AI models'. https://the-decoder.com/openai-adds-111-billion-to-its-cash-burn-forecast-as-ai-costs-spiral-beyond-projections/ |
| OpenAI planned compute spend through 2030, high | $856B | leak | Attributed to an FT report of a company presentation. Seen only in a search summary; the original was not read, so this is the weakest of the three |
| OpenAI planned training compute, 2026 | $32B | leak | The Information via The Decoder, 2026-02-21 (same URL as the middle total) |
| OpenAI planned training compute, 2027 | $65B | leak | The Information via The Decoder, 2026-02-21 |
| OpenAI planned training compute, total through 2030 | $440B | leak | The Information via The Decoder, 2026-02-21: 'nearly $440 billion through 2030' |
| OpenAI annualized revenue at the end of Sept 2026 | $50B a year | leak | FT via Bloomberg Law, 2026-10-08. An earlier Axios figure of about $70B used a different basis. https://news.bgov.com/artificial-intelligence/openais-annualized-revenue-nears-50-billion-ft-says |
| OpenAI's planned compute spend in each year after 2030, against its 2030 level | 100 % of the plan's 2030 level | assumed | Assumed. The leaked plan stops at 2030 and nothing is known after it. Planned spending is held at its 2030 level in 2031 and 2032; the strategy can add to it |
| OpenAI cash and securities, about the end of Q1 2026 | $73B | leak | The Information via The Decoder, 2026-06-20: 'more than $73 billion'. Stale: later burn and later fundraising are not reflected. https://the-decoder.com/openai-tripled-revenue-to-5-7-billion-in-q1-but-burned-through-3-7-billion-to-get-there/ |
| OpenAI round now closing: about $15B at $1.25T before the money | $15B | reported | CNBC 2026-11-12 (via TradingView https://www.tradingview.com/news/seekingalpha:b0ba9e182094b:0-openai-raises-15b-at-1-25t-valuation-in-latest-fundraising-round-cnbc/): about $15B raised; the target was cut to about $20B and investors were described as tepid. Bloomberg 2026-11-06: about $11B wired at the first close. The last $5B to the target is not counted (not confirmed) |
| When that round is counted: Nov to Dec 2026 | 2026.900 | reported | First close 2026-11-06 (Bloomberg); the rest reported as coming in the following weeks. Counted in the model's first quarter |
| What OpenAI raises at its IPO | $60B | assumed | Assumed: no outlet has reported a size, and Polymarket's market on it is too thin to use (about $1k a bucket). My choice, 2026-11-18, from the agent's suggestion. Yardstick: Anthropic's IPO, up to $64B plus a $26B placement |
| When OpenAI's IPO is counted: July 2027 | 2027.500 | market | Where prediction markets put even odds, read 2026-11-18: Polymarket 'OpenAI IPO by...?' 22.5% by 2027-03-31, 45.5% by 2027-06-30, about 75% by 2027-12-31 (https://polymarket.com/event/openai-ipo-by); Kalshi similar. The company: 'will be a public company in 2027' (CNBC 2026-08-19) |
| What Microsoft booked from OpenAI in the quarter to Sept 30, 2026 (cloud spend plus revenue share) | $10B | reference | Microsoft Form 10-Q filed 2026-10-28, related-party note. Filed, read directly. Not used in the arithmetic |
| OpenAI's own 2027 revenue projection (Feb 2026 version) | $62B | reference | The Information via The Decoder, 2026-02-21. Leak; not used in the arithmetic |
| OpenAI's own 2030 revenue projection (Feb 2026 version) | $280B | reference | CNBC via Reuters/CGTN, 2026-02-21: 'more than $280 billion'. Leak; not used in the arithmetic |
| OpenAI's reported gross margin, first quarter of 2026 (33% in 2025), which would put serving costs near 61% of revenue | 0.39 share | reference | The Information via The Decoder, 2026-06-20 and 2026-02-21. Leak; how the margin is defined is not given. Not used in the arithmetic |
| OpenAI: consumer subscriptions as a share of revenue, third quarter of 2026 | 50 % of revenue | leak | Dataconomy, 2026-11-11: about $10B in the quarter, 'split evenly between enterprise and consumer'. Second-hand, no primary outlet named. In August OpenAI's finance chief told investors 'the majority of our revenue is now enterprise' (CNBC, 2026-08-14). Adjustable |
| OpenAI revenue, third quarter of 2026 | $10B | leak | Dataconomy, 2026-11-11: about $10 billion, up about 50% on the second quarter. Second-hand. Context only |
| Secure-a-share strategy: the share of the world's AI compute OpenAI secures | 20 % of the world's AI compute | assumed | Used only by the secure-a-share strategy, which is off to start. AI Futures Project compute forecast (April 2025): 'OpenAI's usage share of the world's AI compute jumps from 5% at the end of 2024 to 20% and Anthropic from 4% to 14%' by the end of 2027. Epoch AI (2026-05-20) put OpenAI at 10% to 15% of the world's working AI compute at the end of 2025. Adjustable |
| Capability (from the AI Futures model's public source) | |||
| Date the AI Futures model treats as the present (its settings are from then) | 2026.578 | their model | AI Futures Model source, tests/golden_data/default_trajectory.json at commit 1c40ecdb (https://github.com/AI-Futures-Project/aifm-public), parameter present_day. 2026.578 is about Aug 1, 2026 |
| ECI score of the best model at the anchor date | 162 ECI points | their model | AI Futures Model source, tests/golden_data/default_trajectory.json at commit 1c40ecdb (https://github.com/AI-Futures-Project/aifm-public), parameter eci_at_present_day. ECI is Epoch AI's Capabilities Index |
| Date of that ECI score | 2026.525 | their model | AI Futures Model source, tests/golden_data/default_trajectory.json at commit 1c40ecdb (https://github.com/AI-Futures-Project/aifm-public), parameter eci_reference_date. 2026.525 is about Jul 11, 2026 |
| How fast the best model's ECI score was rising at the anchor date | 15 ECI points a year | their model | AI Futures Model source, tests/golden_data/default_trajectory.json at commit 1c40ecdb (https://github.com/AI-Futures-Project/aifm-public), parameter eci_growth_rate |
| How fast effective compute was growing at the anchor date (1.6 = a factor of about 40 a year) | 1.608 log10 a year | their model | AI Futures Model source, tests/golden_data/default_trajectory.json at commit 1c40ecdb (https://github.com/AI-Futures-Project/aifm-public), series progress_rates, interpolated between two points of its 0.1-year grid. The model converts effective compute to ECI at 'ECI points a year' divided by a growth rate at this date. The exact rate it uses comes from an internal run that is not in the shipped file; this is the rate on the finished default trajectory. The other candidate in the file (the run with no AI help in research, 1.447) would make each factor of 10 worth about 10% more ECI points |
| The same growth rate on the model's run with no AI help in research (the other candidate for the ECI conversion; not used) | 1.447 log10 a year | reference | AI Futures Model source, tests/golden_data/default_trajectory.json at commit 1c40ecdb (https://github.com/AI-Futures-Project/aifm-public), series human_only_progress_rates, interpolated between two points of its 0.1-year grid |
| Effective compute at the anchor date | 28.332 log10 | their model | AI Futures Model source, tests/golden_data/default_trajectory.json at commit 1c40ecdb (https://github.com/AI-Futures-Project/aifm-public), series effective_compute, interpolated between two points of its 0.1-year grid |
| Effective compute at the model's present day | 28.418 log10 | their model | AI Futures Model source, tests/golden_data/default_trajectory.json at commit 1c40ecdb (https://github.com/AI-Futures-Project/aifm-public), series effective_compute, interpolated between two points of its 0.1-year grid |
| Training compute at the model's present day | 27.413 log10 | their model | AI Futures Model source, tests/golden_data/default_trajectory.json at commit 1c40ecdb (https://github.com/AI-Futures-Project/aifm-public), series training_compute, interpolated between two points of its 0.1-year grid |
| How fast methods were improving at the model's present day (1.08 = a factor of about 12 a year) | 1.084 log10 a year | their model | AI Futures Model source, tests/golden_data/default_trajectory.json at commit 1c40ecdb (https://github.com/AI-Futures-Project/aifm-public), series software_progress_rates, interpolated between two points of its 0.1-year grid. Used for the 'today's pace continues' line, which is this project's own construction, not an AI Futures trajectory |
| AI Futures setting: ten times less compute for experiments makes methods research this many times slower | 2.8 × per 10× compute | their model | AI Futures Model source, tests/golden_data/default_trajectory.json at commit 1c40ecdb (https://github.com/AI-Futures-Project/aifm-public), parameter inv_compute_anchor_exp_cap. Described in components/PlanAExplorerClient.tsx as 'The factor by which research progress slows if experiment compute is cut by 10×'. Their range is 1 to 10. In their model this holds at mid-2026 levels with the number of researchers unchanged; this page applies it as a constant rule in both directions and to a lab's whole research compute, which is this project's simplification |
| AI Futures law for methods: how good methods are = (all research done so far) to this power | 3.186 power | their model | AI Futures Model source, tests/golden_data/default_trajectory.json at commit 1c40ecdb (https://github.com/AI-Futures-Project/aifm-public). Not a recorded setting: the model calibrates it inside a run (the recorded starting value is r_software = 2.4). Derived here from three series of the finished default run, as software_progress_rates x ln 10 x research_stock / research_efforts, which gives the same number at every point of the trajectory (spread 2.2e-15) |
| AI Futures revenue rule: the leading lab's revenue multiplies by this for each year's worth of progress at the mid-2026 pace | 5 × a year | their model | AI Futures Model source, tests/golden_data/default_trajectory.json at commit 1c40ecdb (https://github.com/AI-Futures-Project/aifm-public), parameter present_revenue_growth_rate. Their note on it: 'Based on recent revenue growth, adjusted downward due to expectations of slowing soon.' Their range in constants/parameters.ts is 2 / 5 / 12.5 (low / middle / high). The rule itself (progress_model/README.md): revenue = present_revenue x growth^((progress - progress at the reference date) / progress rate at the reference date) |
| Date the AI Futures revenue rule starts from | 2026.374 | their model | AI Futures Model source, tests/golden_data/default_trajectory.json at commit 1c40ecdb (https://github.com/AI-Futures-Project/aifm-public), parameter revenue_reference_date. 2026.374 is about May 17, 2026 |
| How fast effective compute was growing at that date: one 'year's worth of progress' in the revenue rule | 1.582 log10 a year | their model | AI Futures Model source, tests/golden_data/default_trajectory.json at commit 1c40ecdb (https://github.com/AI-Futures-Project/aifm-public), series progress_rates, interpolated between two points of its 0.1-year grid. As with the ECI conversion, the exact rate the model uses comes from an internal run that is not in the shipped file; this is the rate on the finished default trajectory. The other candidate in the file (1.455) would make revenue grow about 9% faster per tenfold of effective compute |
| The same growth rate on the model's run with no AI help in research (the other candidate for the revenue rule; not used) | 1.455 log10 a year | reference | AI Futures Model source, tests/golden_data/default_trajectory.json at commit 1c40ecdb (https://github.com/AI-Futures-Project/aifm-public), series human_only_progress_rates, interpolated between two points of its 0.1-year grid |
| Revenue the AI Futures rule starts from: Anthropic's annualized revenue in May 2026 | $47B a year | reference | AI Futures Model source, tests/golden_data/default_trajectory.json at commit 1c40ecdb (https://github.com/AI-Futures-Project/aifm-public), parameter present_revenue. components/ParameterHoverContext.tsx describes it as 'Anthropic's most recent reported annualized revenue'. Not used by this page: the rule is applied as a growth multiplier to each lab's own starting revenue |
| AI Futures guess of the leading lab's revenue when AI can do all of a lab's coding work | $1,000B a year | reference | AI Futures Model source, tests/golden_data/default_trajectory.json at commit 1c40ecdb (https://github.com/AI-Futures-Project/aifm-public), parameter revenue_required_for_ac. Described there as 'in between current revenue and the revenue that would be obtained by automating a huge fraction of white-collar labor'. Other presets in constants/parameters.ts use $7 trillion and $3 trillion. Shown for scale next to the market ceiling; not used in the arithmetic |
| ECI score the AI Futures model uses as its mark for AI that can do all of a lab's coding work | 250 ECI points | their model | AI Futures Model source, tests/golden_data/default_trajectory.json at commit 1c40ecdb (https://github.com/AI-Futures-Project/aifm-public), parameter eci_required_for_ac. Used by the model only in one of its optional modes; shown on the chart as a reference line |
| How often better methods halve the compute needed for the same result (Epoch AI) | 8 months | reported | Ho et al., 'Algorithmic progress in language models' (Epoch AI; arXiv 2403.05812, 2024-03-09; NeurIPS 2024): 'the compute required to reach a set performance threshold has halved approximately every 8 months, with a 95% confidence interval of around 5 to 14 months', over 2012-2023 (language-model scores on Wikitext and Penn Treebank). They add that 'the increase in compute made an even larger contribution to overall performance improvements over this time period'. Read on arXiv by the agent's research helper, 2026-11-17. Used as the pace of methods on the starting line from today on (the Owner's choice of 2026-11-17: Epoch AI for the start) |
| How fast the largest training runs grow, through 2030 (Epoch AI) | 4 × a year | reported | Epoch AI, 'Can AI scaling continue through 2030?' (2024-08-20): the trend of 'approximately 4x per year' can go on, and '2e29 FLOP training runs will likely be feasible by 2030'. Epoch AI, 'Training compute of frontier AI models grows by 4-5x per year' (2024-05-28): the ten largest models 4.2x a year since about 2018 (90% CI 3.6x to 4.9x), notable models 4.1x a year since 2010. Read on Epoch's pages by the agent's research helper, 2026-11-17 |
| The end of Epoch AI's compute trend on the starting line (the end of 2030); after it the AI Futures Model's training compute growth | 2031.000 | assumed | The agent's choice (the Owner, 2026-11-17: 'lean towards epoch unless there's a reason AI Futures fits better'): Epoch's report runs 'through 2030' and says little after it; the AI Futures Model is the only forecast past 2030 on this page, and slows (about x2 a year in 2031, x1.7 in 2032) |
| The AI market, shares and keeping pace | |||
| Placeholder: the most all AI labs together could earn in a year | $10,000B a year | assumed | No data sets this. A round figure, set on 2026-11-16. For scale: about 8% of world output in 2026 (about $128 trillion), and 1.6 times world IT spending ($6.39 trillion). The AI Futures model guesses $1, $3 and $7 trillion for ONE lab's revenue when AI does all coding. A stand-in until there is a demand model |
| Revenue pace today of every AI lab other than Anthropic and OpenAI | $87B a year | assumed | Nobody publishes this. Derived: Counterpoint Research put Anthropic at 31.4% and OpenAI at 29% of global LLM revenue in the first quarter of 2026, about 60% together, so everyone else earns two-thirds of what the two labs earn together ($81B + $50B) |
| How strongly a lab's share follows its capability | 1 power | assumed | No study measures this. At 1, a lab's pull on customers is proportional to its effective compute; at 0 shares never move. Only qualitative evidence: Menlo Ventures (Nov 2026) has Anthropic's share of enterprise API spend going from 40% to 55% in a year |
| How much more compute a dollar buys each year | 49 % a year | assumed | Epoch AI, 'AI and compute' (updated 2026-02-05, https://epoch.ai/research/ai-and-compute): 'AI chip performance per dollar has grown by 49% per year since 2023'. Assumed to continue through 2030. Epoch's older GPU studies found 32% to 39% a year |
| How a lab's research compute moves its own effective compute | 1 power | assumed | Assumed 1 to 1: twice the research compute gives twice the effective compute, and better methods reach every lab equally. The AI Futures model has no per-lab version; its settings imply more experiment compute also speeds up methods (experiment_compute_exponent 0.655), left out here, so a lab's own compute probably matters somewhat more |
| World output in 2026 | $128,000B | reported | IMF World Economic Outlook, October 2026, read through a second-hand page (ceoworld.biz, 2026-11-05: 'more than $128 trillion of output in 2026, measured in current U.S. dollars'). The IMF's own table was not reached |
| Share of world income paid as wages | 52.6 % of world income | reported | ILO, Employment and Social Trends 2026 (January 2026), read directly: 'The global labour income share, at 52.6 per cent in 2025' |
| What the world spends on all IT in 2026 | $6,391B | reported | Gartner press release of 2026-10-21, read through second-hand pages that reproduce its table (techintelpro.com, saastr.com). Gartner's own page was not reachable |
| Anthropic's share of global LLM revenue, first quarter of 2026 | 31.4 % of global LLM revenue | reported | Counterpoint Research, as reported by The Register on 2026-04-30 (https://www.theregister.com/2026/04/30/openai_anthropic_top_lines_research_counterpoint/). Taken from a search summary; the article itself was not opened |
| OpenAI's share of global LLM revenue, first quarter of 2026 | 29 % of global LLM revenue | reported | Counterpoint Research, as reported by The Register on 2026-04-30. Taken from a search summary; the article itself was not opened |
| Anthropic and OpenAI together, share of enterprise LLM API spend | 86 % of enterprise API spend | reported | Menlo Ventures, '2026: The State of Enterprise AI' (2026-11-03), a survey of about 400 US decision-makers, read through second-hand pages. Anthropic 55%, OpenAI 31%, Google 9%. Menlo is an investor in Anthropic |
| How fast customers move: the part of the distance to its new share that a lab covers in a year | 70 % of the distance, each year | assumed | A fit on six periods (2024 to the third quarter of 2026) of Anthropic's revenue over OpenAI's against the gap between their best models on METR's task-length measure: the best form closes half the distance in about 7 months, which is 70% in a year; a looser form gives 'at once'. Adding the third quarter of 2026 moved the fits from 75% and 40%, so it is not stable; METR has no measurements after April 2026. Suggestive only. tools/fit_competition.py. Adjustable |
| Price sensitivity of demand: a 10% price cut raises the work bought by this many times 10% | 1 power | assumed | Could not be measured: Anthropic's prices per unit barely moved in its filed history (cuts took about $600M from one quarter and $70M from another). OpenRouter's 'State of AI' (December 2025) finds a 10% price cut goes with only 0.5 to 0.7% more usage (0.05 to 0.07); Anthropic's filing says the opposite in words. At 1 a price change leaves revenue unchanged. Adjustable |
| Change in the price of a unit of work each year, with the cost of serving it unchanged | 0 % a year | assumed | Assumed zero: the filing says cost-per-task cuts were made 'while preserving our margins', so price and serving cost moved together. A cut lowers the markup over serving cost (6.9 times today). Adjustable |
| History: Anthropic's and OpenAI's revenue together against the best model's ECI, 2024 to mid-2026 | 3.8 × per year of progress | reference | This project's fit (tools/fit_competition.py, run 2026-11-17) on Epoch AI's ECI, METR's task-length data and inputs/competition_history.json. Six periods. The best model's score and the calendar moved together (correlation 0.999), so this gives better models all the growth: an upper limit. Not used in the arithmetic |
| History: Anthropic's own quarterly revenue against the best model's ECI, late 2024 to the third quarter of 2026 | 9.6 × per year of progress | reference | This project's fit (tools/fit_competition.py, run 2026-11-17) on Epoch AI's ECI, METR's task-length data and inputs/competition_history.json. Eight filed quarters; 7.8 against Anthropic's own best model. An upper limit too. Not used in the arithmetic |
| History: lowest share sensitivity the fits give | 0.3 power | reference | This project's fit (tools/fit_competition.py, run 2026-11-17) on Epoch AI's ECI, METR's task-length data and inputs/competition_history.json. Anthropic's revenue over OpenAI's against the gap in best-model METR task length, with a time trend (0.29); a Menlo Ventures survey of enterprise API spend (2025 to 2026) gives 0.3 for Anthropic against OpenAI. Six periods; suggestive only. Not used in the arithmetic |
| History: highest share sensitivity the task-length fits give | 1.1 power | reference | This project's fit (tools/fit_competition.py, run 2026-11-17) on Epoch AI's ECI, METR's task-length data and inputs/competition_history.json. Gradual-switching form (1.06); plain form 0.93. The same survey gives 2.9 for Google against OpenAI. On ECI no fit works: Anthropic gained on OpenAI in almost every period while level or behind on ECI. Not used in the arithmetic |
| ECI of Anthropic's best released model, 2026-11-16 | 173.4 ECI points | reported | Epoch AI, eci_scores.csv, downloaded 2026-11-16: Claude Fable 5.2, released 2026-11-09, 173.44 (their range 169.9 to 177.5). Anthropic's starting position (my decision, 2026-11-16). On Oct 1, 2026 Anthropic's best was 168.5, below OpenAI's |
| ECI of OpenAI's best released model, 2026-11-16 | 169 ECI points | reported | Epoch AI, eci_scores.csv, downloaded 2026-11-16: GPT-6 Astra, released 2026-09-03, 169.0 (their range 166.1 to 172.3). OpenAI's starting position (my decision, 2026-11-16) |
| Every other lab: consumer subscriptions as a share of revenue | 15 % of revenue | assumed | No other lab publishes this. Menlo Ventures puts all consumer AI spending at about $40B in 2026; OpenAI's consumer revenue and Anthropic's subscriptions already come to about $30B of it, leaving about $10B for everyone else, roughly 15% of the $87B pace assumed for the other labs. Placeholder |
| Business revenue (or all revenue, with one kind of customer) multiplies by this for each year's worth of progress of the best model | 2 × per year of progress | assumed | My decision, 2026-11-17: start at the low end of the AI Futures Project's range for its revenue rule, 2 / 5 / 12.5 (low / middle / high; constants/parameters.ts in their source, see cap.revenue_multiplier). This project's fit gives 3.8 for the two labs together and 5.1 for business revenue alone, both upper limits. Adjustable |
| Consumer revenue multiplies by this for each year's worth of progress of the best model | 1.5 × per year of progress | assumed | The agent's choice, 2026-11-17, when I set the business multiplier to 2: keep the relation the fit found, not the fitted number. The fit (tools/fit_competition.py, four periods) gives consumer revenue x2.76 and business revenue x5.14 per year of progress, so consumer revenue follows capability about 0.62 times as strongly (log 2.76 / log 5.14). With a business multiplier of 2 that is 2^0.62 = 1.54, rounded to 1.5. Adjustable |
| Placeholder: the most all labs together could earn from consumers in a year (part of the $10 trillion ceiling) | $1,000B a year | assumed | No data. A round figure: about 4 billion adults paying $20 a month would be $1 trillion a year. The rest of the ceiling is business. Adjustable |
| Adoption delay: the time added to how long the market needs to double, when all labs together earn $1 trillion a year | 0.4 years | assumed | No source gives this number. My decisions of 2026-11-17: adoption gets harder smoothly as revenue grows, with no point where a limit switches on ("as revenue increases, further adoption becomes harder"; "I like the delay idea"), and the delay is set so that Anthropic's 2028 revenue on the starting settings equals the middle of Manifold's market, $419 billion as read on 2026-11-17 (play money, 8 traders; "I trust manifold more than anthropic's projection"; "2028 pinned"). Found with tools/fit_adoption_delay.py, rounded to two decimals; not moved when the market moves. Adjustable; 0 = no delay |
| How much longer the adoption delay gets each time the revenue of all labs together doubles | 2 times per doubling of the market | assumed | No source gives this number. My choice of 2026-11-17 ("2x is good") from four shown: 1.5, 2, 2.8 and 4. Other large markets' growth (data item 55) fell by about a third to a half per doubling above about $300 billion a year: the agent's reading of a handful of cases, not a fit. Nothing checks it; the 2028 pin fixes the delay at one size only, so this sets the years after 2028. Adjustable; 1 = the delay is the same at every size |
| Raising when cash gets low: the most a lab can raise in any 12 months, as a multiple of its yearly revenue (0 = off) | 1 x a year's revenue | assumed | My choice, 2026-11-18 ('1x'). For scale: OpenAI raised about 11 times its revenue in 2025 and about 9 times in the March 2026 round; Anthropic about 1.4 times in May 2026; OpenAI's round now closing is about 0.4 times (docs/research/fundraising.md). Money comes only while revenue grows year on year |
| Raising when cash gets low: a lab raises when its cash would fall below this many months of its spending | 6 months | assumed | Assumed: the agent's choice, not confirmed. No lab publishes a cash floor |
| The figure the adoption delay was set to: the middle of Manifold's market on Anthropic's revenue for 2028, as read on 2026-11-17 | $418.7B | reference | Manifold, market anthropic-2028-revenue-56quO59zEA, read 2026-11-17: play money, 8 traders; the middle is worked out by tools/fetch_revenue_markets.py from the chances of the listed ranges (10% to 90%: $130 billion to $942 billion). Not read by the model; it records what link.assume.adoption_delay was pinned to |
| ECI of the best released model of any other lab, 2026-11-16 | 166.5 ECI points | reported | Epoch AI, eci_scores.csv, downloaded 2026-11-16: Gemini 4 Argon (Google DeepMind), limited access, released 2026-09-30, 166.48. An unreleased Google model is listed at 169.76. Used as the other labs' starting position |
| How far the best open-weight model is behind the best closed model, 2026 (about four months) | 8 ECI points | reported | Epoch AI, 'Open models lag state-of-the-art closed models by 4 months' (Data Insight, Jack Edwards and Luke Emberson, 2026-05-29), epoch.ai/data-insights/open-closed-eci-gap: 'Since January 2026, the most capable open-weight models have lagged frontier closed models by an average of four months'; the average ECI gap was 8 points; six months under a stricter test (the open model's point estimate strictly higher). Read on Epoch's page by the agent's research helper, 2026-11-17 |
| Gross margin of a company that hosts open models: the margin a lab falls to when its best alternative catches up | 50 % gross margin | reported | Contrary Research on Fireworks AI: 'As of July 2025, gross margins were ~50% with management targeting ~60% through improved GPU utilization' (it cites DatacenterDynamics; read on Contrary's page by the agent's research helper, 2026-11-17). Sacra gives the same ~50% (seen only in search snippets). Third-hand, and no host of open models reports its margin: Together AI about 45% (a Substack with no method, snippet only), nothing found for Baseten or DeepInfra. Used as the floor of the margin rule (the agent's construction; the Owner, 2026-11-17: 'I like the price competition rule but apply it to openai too ofc. Let's do margin only') |
| Open-source models' share of enterprise model spend, September 2026 | 4 % of enterprise model spend | reported | Menlo Ventures, '2026: The State of Enterprise AI' (2026-11-03): 'Open-source remains 4% of spend despite 19% of usage' (6% a year earlier; 19% of tokens, from OpenRouter). A survey of about 400 US enterprise decision-makers, spend inferred from workload shares; consumer use and self-hosting are not in it, and 'open-source' is not defined. Read on Menlo's page (as a summary) by the agent's research helper, 2026-11-17. Menlo is an investor in Anthropic. Used for consumers too, for want of a figure (the agent's choice, not confirmed) |
| Date of the starting ECI scores (2026-11-16) | 2026.874 | reported | The day Epoch AI's file was downloaded. The forecast's ECI line is moved up or down so that it passes through the best measured model on this date |
| A lab's own edge: ECI points a year its best model gains beyond what its research compute gives | 0 ECI points a year | assumed | No data on what one lab's methods add over another's. History gives no edge either way: from 2024 to the third quarter of 2026 the best model of Anthropic gained 17.3 ECI points a year and OpenAI's 17.1 (tools/fit_competition.py). Zero for both labs. Adjustable, per lab |
| The forecast stops when the best model passes this ECI score | 250 ECI score of the best model | assumed | My rule (2026-11-17): stop the forecast where it becomes too hard to forecast (past superintelligence), with a visible label. The level is the agent's choice, not confirmed: 250, the AI Futures Project's score for AI that can do all of a lab's coding work (cap.eci_for_automated_coder). On their default run superintelligence is reached at about ECI 245 on this page's conversion (link.aifm_asi_eci), so 250 sits just past it. Adjustable; 0 never stops |
| ECI score at which the AI Futures default run reaches its superintelligence milestone, by this page's conversion | 245 ECI points | reference | Worked out from the AI Futures Model source at commit 1c40ecdb (https://github.com/AI-Futures-Project/aifm-public): tests/golden_data/default_trajectory.json and progress_model/_impl.py, compute_milestones. Their 'ASI' milestone is research skill, not an ECI score: ai_research_taste_sds reaching 12.36 standard deviations, four times the best human researchers. On the default run it is reached at 2028.91, effective compute 10^37.19; this page's straight-line conversion to ECI gives 244.6. Read on a 0.1-year grid, good to about 5 points. Earlier marks on the same run: about 198 (SAR), 214 (TED-AI), 229 (SIAR). Context for the stop setting; not used in the arithmetic |
| The forecast stops when all AI labs together earn more than this a year | $5,000 billions a year, all AI labs together | assumed | My rule (2026-11-17): stop where the forecast becomes too hard to forecast. The level is the agent's choice, not confirmed: half the $10 trillion ceiling, a placeholder that at half has already cut the rule's revenue in half, so from there the placeholder decides the result. Adjustable; 0 never stops |
| How many times faster Anthropic's models are improving than before, measured by METR in October 2026 | 1.9 times the earlier pace | reported | METR, 'Independent Assessment of AI R&D Acceleration at Anthropic in October 2026' (Tom Cunningham, Nate Rush, Parker Whitfill, Hjalmar Wijk; November 5, 2026; https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/): Anthropic's own Epoch Capabilities Index has grown since October 2025 at '1.9x [95% confidence interval: 1.5–2.8x]' its rate from March 2024 to October 2025 (about 29 points a year against 15 on this page's ECI scale: the points are this page's arithmetic, METR gives the multiple). METR: 'we would not be surprised if the true acceleration was closer to 1.3x or 3x'. They guess about 80% of it is due to AI, and say 'New internal models make a permanent slope change look more plausible than a one-time jump at Mythos Preview'. On a pace that keeps rising: 'While we do not see clear evidence of this in the AECI trends, there is other evidence that makes this seem more plausible which Anthropic redacted'; their forecast for a future model is redacted. Their index counts Anthropic's unreleased models; Epoch AI's public scores (eci_scores.csv, downloaded 2026-11-16) show less: Anthropic's best released model 13.9 points a year from March 2024 to October 2025 and 22.9 since (1.6 times); the best of any lab 15.4 and 20.8 (1.3 times) |
| The low end of METR's 95% range for that speed-up | 1.5 times the earlier pace | reference | METR, 'Independent Assessment of AI R&D Acceleration at Anthropic in October 2026' (Tom Cunningham, Nate Rush, Parker Whitfill, Hjalmar Wijk; November 5, 2026; https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/): '1.9x [95% confidence interval: 1.5–2.8x]' |
| The high end of METR's 95% range for that speed-up | 2.8 times the earlier pace | reference | METR, 'Independent Assessment of AI R&D Acceleration at Anthropic in October 2026' (Tom Cunningham, Nate Rush, Parker Whitfill, Hjalmar Wijk; November 5, 2026; https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/): '1.9x [95% confidence interval: 1.5–2.8x]' |
| The same speed-up for Anthropic's best released model, from public scores | 1.6 times the earlier pace | reference | Epoch AI, eci_scores.csv (https://epoch.ai/eci, downloaded 2026-11-16), tools/eci_pace_change.py: the best released model's score on March 1, 2024, October 1, 2025 and November 16, 2026. Anthropic: 13.9 points a year before October 2025, 22.9 since, 1.6 times. Released models only, not METR's measure |
| The same speed-up for the best released model of any lab, from public scores | 1.3 times the earlier pace | reference | Epoch AI, eci_scores.csv (https://epoch.ai/eci, downloaded 2026-11-16), worked out by tools/eci_pace_change.py: the best released model's score on March 1, 2024, on October 1, 2025 and on November 16, 2026. Any lab: 15.4 points a year before October 2025 and 20.8 a year since, 1.3 times |
| AI improves faster than before, by (the starting value) | 1.9 times the old pace | assumed | My decision, 2026-11-17: the page opens on the pace METR measured, held steady. 1.9 is METR's middle estimate (link.metr_speedup); its range is 1.5 to 2.8 and Epoch AI's public scores show 1.3 to 1.6 (link.public_speedup_anthropic, link.public_speedup_any_lab) |
| If the speed-up keeps growing: how long it takes to double | 1 years | assumed | The agent's choice, not confirmed by me. METR measured one change (the pace about doubled over about a year) and says the faster pace looks more likely to last than to fall back; it sees no clear evidence of a pace that keeps rising and gives no rate (forecast redacted). One year is the reading 'what happened over the last year happens again each year'. Adjustable. Counted in years of compute growth as the forecast has it |
| The ECI score of the Kalshi market used as a point the line must reach | 185 ECI points | reported | Kalshi, series KXEPOCHSCORE ('Epoch AI capability score'), https://kalshi.com/markets/kxepochscore: the only threshold still trading on 2026-11-17. Used only by settings that switch on 'pass through the prediction market's point'; the starting settings do not (my call, 2026-11-17: 'don't anchor, just include an indicator'). I then asked (2026-11-17) that the AI Futures and Amodei scenarios be at least at what the market predicts |
| The date by which that market gives even odds (January 23, 2027) | 2027.060 | reported | Kalshi public API, read 2026-11-17 06:49 UTC by tools/fetch_kalshi_eci.py: a model above 185 by 2027-01-01 at 34%, by 2027-02-01 at 57% (middle of best bid and ask); even odds fall on 2027-01-23 on a straight line between the two. One chance in four by 2026-12-18, three in four by 2027-04-09. Thin: 4,824 contracts in all. A snapshot: the chart marker follows the live market, this number does not |
| The world's AI compute | |||
| The world's AI data-centre power at the end of 2025 | 15 gigawatts, end of year | their model | AI Futures Project, 'Compute Forecast' (Romeo Dean, April 2025, https://ai-2027.com/research/compute-forecast), section 1 table, row 'Total AI datacenter power requirement | 5GW | 9GW | 15GW | 29GW | 62GW' for 2023 to 2027 (read 2026-11-17). Taken as end of year (an assumption; the table gives no day). The summary says 60GW for 2027, the table 62GW; the table is used. A forecast, not a measurement: Epoch AI's count of tracked sites was 15.1 GW in November 2026, covering only the sites it tracks |
| The world's AI data-centre power at the end of 2026 | 29 gigawatts, end of year | their model | AI Futures Project, 'Compute Forecast' (Romeo Dean, April 2025, https://ai-2027.com/research/compute-forecast), section 1 table, row 'Total AI datacenter power requirement | 5GW | 9GW | 15GW | 29GW | 62GW' for 2023 to 2027 (read 2026-11-17). Taken as end of year (an assumption; the table gives no day). The summary says 60GW for 2027, the table 62GW; the table is used. A forecast, not a measurement: Epoch AI's count of tracked sites was 15.1 GW in November 2026, covering only the sites it tracks |
| The world's AI data-centre power at the end of 2027 | 62 gigawatts, end of year | their model | AI Futures Project, 'Compute Forecast' (Romeo Dean, April 2025, https://ai-2027.com/research/compute-forecast), section 1 table, row 'Total AI datacenter power requirement | 5GW | 9GW | 15GW | 29GW | 62GW' for 2023 to 2027 (read 2026-11-17). Taken as end of year (an assumption; the table gives no day). The summary says 60GW for 2027, the table 62GW; the table is used. A forecast, not a measurement: Epoch AI's count of tracked sites was 15.1 GW in November 2026, covering only the sites it tracks |
| The world's AI data-centre power in 2030, middle case | 156 gigawatts, 2030 | reported | McKinsey, 'The cost of compute: a $7 trillion race to scale data centers' (April 2025). SECOND-HAND, NOT CHECKED (the page timed out on 2026-11-17): from other outlets' reports. AI data-centre capacity of 156 GW in 2030 in its middle case, 124 GW added from 2025 to 2030, against 78 GW added in its constrained case and 205 GW in its accelerated case. Middle case. Sets the starting yearly growth of world AI compute from 2028: (156 / 62)^(1/3) = 1.36. A different source than the 2025 to 2027 figures; they disagree (the 2027 figure is already 40% of it). No public year-by-year figure for 2028 to 2032, none at all for 2031 or 2032; those years repeat the same growth (an assumption). Adjustable |
| The world's AI data-centre power in 2030, low case | 110 gigawatts, 2030 | assumed | McKinsey, 'The cost of compute: a $7 trillion race to scale data centers' (April 2025). SECOND-HAND, NOT CHECKED (the page timed out on 2026-11-17): from other outlets' reports. AI data-centre capacity of 156 GW in 2030 in its middle case, 124 GW added from 2025 to 2030, against 78 GW added in its constrained case and 205 GW in its accelerated case. Low case, derived here: the middle total less the middle addition plus the constrained addition, 156 - 124 + 78 = 110. Growth from 2028: (110 / 62)^(1/3) = 1.21 |
| The world's AI data-centre power in 2030, high case | 237 gigawatts, 2030 | assumed | McKinsey, 'The cost of compute: a $7 trillion race to scale data centers' (April 2025). SECOND-HAND, NOT CHECKED (the page timed out on 2026-11-17): from other outlets' reports. AI data-centre capacity of 156 GW in 2030 in its middle case, 124 GW added from 2025 to 2030, against 78 GW added in its constrained case and 205 GW in its accelerated case. High case, derived here: 156 - 124 + 205 = 237. Growth from 2028: (237 / 62)^(1/3) = 1.56 |
| What a gigawatt of AI compute costs to rent for a year | $12.5 billions per gigawatt-year | assumed | The agent's choice, not confirmed by me: the middle of Dario Amodei's range, 'each gigawatt costs ... on the order of $10-15 billion a year' (Dwarkesh Patel interview, 2026-02-13). Others: Epoch AI's total cost of ownership $8.5 billion per gigawatt-year; five-year contracts with compute renters near $12 billion and long-term rents of $20 to $25 billion (both second-hand). A higher figure makes every lab's share smaller. Adjustable |
| What everyone but Anthropic and OpenAI spends on AI compute | 67 % of what the world's AI compute costs at the usual price | assumed | The agent's choice, not confirmed by me. I asked (2026-11-17) for the price of compute to rise by supply and demand; this is the number that rule needs. The rest of the world is taken to bring 67% of the build-out cost at the usual price, every year: the two labs hold 33% of the world's AI compute today on this page's numbers, and the AI Futures Project forecast (April 2025) gives them 34% at the end of 2027 (OpenAI 20%, Anthropic 14%). No source gives the rest of the world's spend. Adjustable |
| Who pays: shown on the Who pays panel, not used by the forecast | |||
| Anthropic: consumption-based revenue (pay for what is used, mostly the API), nine months of 2026 | $26.3B | filed | S-1/A No. 2, MD&A, lines 7607-7609: '$26.3 billion, or 79% of total revenue, for the nine months ended September 30, 2026' |
| Anthropic: subscription revenue, nine months of 2026 | $6.9B | filed | S-1/A No. 2, MD&A, lines 7609-7611: '$6.9 billion, or 21% of total revenue' |
| Anthropic: revenue from the United States, nine months of 2026 | $23.47B | filed | S-1/A No. 2, notes to the financial statements, revenue by geography, line 20590: United States $23,470 million |
| Anthropic: revenue from outside the United States, nine months of 2026 | $9.772B | filed | S-1/A No. 2, same table, line 20596: International $9,772 million |
| Anthropic: share of revenue from customers paying $100,000 or more a year (more than 11,900 of them), twelve months to September 2026 | 74 % of revenue | filed | S-1/A No. 2, lines 7798-7803: 'more than 11,900 as of September 30, 2026 ... $26.1 billion ... 74% of total revenue ... The remaining 26% ... was generated by approximately 40 million customers' |
| Anthropic: customers paying under $100,000 a year (26% of revenue) | 40 million customers | filed | S-1/A No. 2, lines 7803-7805: 'approximately 40 million customers, each with less than $100,000 of revenue on a TTM basis' |
| Anthropic: share of revenue from its five largest customers, July to September 2026 (36% in January to March) | 21 % of revenue | filed | S-1/A No. 2, lines 7819-7822: 'Approximately 79% of our revenue came from customers outside of our top five customers for the three months ended September 30, 2026, up from approximately 64% for the three months ended March 31, 2026'. No single customer was over 10% in the nine months |
| Anthropic: government agencies (US and others), at most this share of revenue | 2 % of revenue | filed | S-1/A No. 2, risk factors, lines 4213-4214: 'Revenue from government agency contracts represented, in aggregate, less than 2% of our revenue in each of the periods presented'. The only industry the filing gives |
| ChatGPT consumer messages: practical guidance (how-to, tutoring, advice) | 28.8 % of messages | reported | Chatterji et al., 'How People Use ChatGPT', NBER working paper w34255 (nber.org/papers/w34255, 2025-09-15), Figure 7, messages at the end of the window (June 2025); consumer plans only (Free, Plus, Pro), not business plans or the API. Read in the original PDF by the agent's research helper, 2026-11-18. Use, not revenue |
| ChatGPT consumer messages: seeking information | 24.4 % of messages | reported | Chatterji et al., 'How People Use ChatGPT', NBER working paper w34255 (nber.org/papers/w34255, 2025-09-15), Figure 7, messages at the end of the window (June 2025); consumer plans only (Free, Plus, Pro), not business plans or the API. Read in the original PDF by the agent's research helper, 2026-11-18. Use, not revenue |
| ChatGPT consumer messages: writing (editing, drafting, translating) | 23.9 % of messages | reported | Chatterji et al., 'How People Use ChatGPT', NBER working paper w34255 (nber.org/papers/w34255, 2025-09-15), Figure 7, messages at the end of the window (June 2025); consumer plans only (Free, Plus, Pro), not business plans or the API. Read in the original PDF by the agent's research helper, 2026-11-18. Use, not revenue |
| ChatGPT consumer messages: technical help (programming 4.2% of all messages) | 5.1 % of messages | reported | Chatterji et al., 'How People Use ChatGPT', NBER working paper w34255 (nber.org/papers/w34255, 2025-09-15), Figure 7, messages at the end of the window (June 2025); consumer plans only (Free, Plus, Pro), not business plans or the API. Read in the original PDF by the agent's research helper, 2026-11-18. Use, not revenue |
| ChatGPT consumer messages that are about work (73% are not), June 2025 | 27 % of messages | reported | Chatterji et al., 'How People Use ChatGPT', NBER working paper w34255 (nber.org/papers/w34255, 2025-09-15), Table 1, June 2025; consumer plans only (Free, Plus, Pro), not business plans or the API. Read in the original PDF by the agent's research helper, 2026-11-18. Use, not revenue |
| Businesses' spending on AI coding tools, all vendors, the year to September 2026 | $34B | reported | Menlo Ventures, '2026: The State of Enterprise AI' (menlovc.com/perspective/2026-the-state-of-enterprise-ai/, published 2026-11-03, data to September 2026; the web page, read by the agent's research helper 2026-11-18). Buyers' spending across all vendors, a survey and model by one venture firm that is an Anthropic investor; not any lab's revenue. Claude Code and Codex together $24.0B of it (70%); Menlo does not split the two |
| Of which Claude Code and OpenAI's Codex together | $24B | reported | Menlo Ventures, '2026: The State of Enterprise AI' (menlovc.com/perspective/2026-the-state-of-enterprise-ai/, published 2026-11-03, data to September 2026; the web page, read by the agent's research helper 2026-11-18). Buyers' spending across all vendors, a survey and model by one venture firm that is an Anthropic investor; not any lab's revenue |
| Businesses' spending on AI chat assistants, all vendors | $25B | reported | Menlo Ventures, '2026: The State of Enterprise AI' (menlovc.com/perspective/2026-the-state-of-enterprise-ai/, published 2026-11-03, data to September 2026; the web page, read by the agent's research helper 2026-11-18). Buyers' spending across all vendors, a survey and model by one venture firm that is an Anthropic investor; not any lab's revenue |
| Businesses' spending on other AI work tools: customer support $2.7B, IT $1.6B, sales $1.4B, marketing $1.2B, HR $0.9B and more | $8.8B | reported | Menlo Ventures, '2026: The State of Enterprise AI' (menlovc.com/perspective/2026-the-state-of-enterprise-ai/, published 2026-11-03, data to September 2026; the web page, read by the agent's research helper 2026-11-18). Buyers' spending across all vendors, a survey and model by one venture firm that is an Anthropic investor; not any lab's revenue. 'Horizontal' apps other than coding and chatbots |
| Businesses' spending on AI tools for one industry: healthcare $3.7B, legal $1.8B and more | $10.1B | reported | Menlo Ventures, '2026: The State of Enterprise AI' (menlovc.com/perspective/2026-the-state-of-enterprise-ai/, published 2026-11-03, data to September 2026; the web page, read by the agent's research helper 2026-11-18). Buyers' spending across all vendors, a survey and model by one venture firm that is an Anthropic investor; not any lab's revenue. 'Vertical' apps; no figure for finance or government on the page |
| Anthropic's share of businesses' spending on AI models through the API (40% a year before) | 55 % of spending | reported | Menlo Ventures, '2026: The State of Enterprise AI' (menlovc.com/perspective/2026-the-state-of-enterprise-ai/, published 2026-11-03, data to September 2026; the web page, read by the agent's research helper 2026-11-18). Buyers' spending across all vendors, a survey and model by one venture firm that is an Anthropic investor; not any lab's revenue |
| OpenAI's share of the same (27% a year before) | 31 % of spending | reported | Menlo Ventures, '2026: The State of Enterprise AI' (menlovc.com/perspective/2026-the-state-of-enterprise-ai/, published 2026-11-03, data to September 2026; the web page, read by the agent's research helper 2026-11-18). Buyers' spending across all vendors, a survey and model by one venture firm that is an Anthropic investor; not any lab's revenue |
| Google's share of the same (21% a year before) | 9 % of spending | reported | Menlo Ventures, '2026: The State of Enterprise AI' (menlovc.com/perspective/2026-the-state-of-enterprise-ai/, published 2026-11-03, data to September 2026; the web page, read by the agent's research helper 2026-11-18). Buyers' spending across all vendors, a survey and model by one venture firm that is an Anthropic investor; not any lab's revenue |
| Claude.ai conversations that are computer and maths tasks, February 2026 | 35 % of conversations | reported | Anthropic Economic Index, March 2026 report (anthropic.com/research/economic-index-march-2026-report), data 2026-02-05 to 02-12, one million conversations, grouped by occupation (O*NET), not by the customer's industry. Anthropic's own study; read by the agent's research helper 2026-11-18. Use, not revenue |
| Requests to Anthropic's own API that are computer and maths tasks, November 2025 | 46 % of API requests | reported | Anthropic Economic Index, January 2026 report, data 2025-11-13 to 11-20, one million API transcripts; the report gives 46% in chapter 1 and 52% in chapter 2 without saying why (46% used). Anthropic's own study; read by the agent's research helper 2026-11-18. Use, not revenue |
| Company | Twelve months to | Revenue | Operating profit | Market value | Value ÷ revenue | Value ÷ operating profit | Operating margin |
|---|---|---|---|---|---|---|---|
| Microsoft | 2026-09-30 | $346B | $163B | $3.90T | 11.3× | 23.9× | 47% |
| Alphabet | 2026-09-30 | $471B | $157B | $3.66T | 7.8× | 23.3× | 33% |
| Nvidia | 2026-07-26 | $303B | $198B | $3.95T | 13.1× | 20.0× | 65% |
Revenue and profit are the last twelve months filed with the SEC. Market value is the closing share price on 2026-11-13 (from a market data site, not a filing) times the diluted share count of the latest quarter; it is not adjusted for cash or debt.
This is a personal research model, not investment advice. It can be wrong, and several of its inputs are leaks or assumptions.