# AI Lab Finances > An interactive model of what Anthropic and OpenAI must earn to pay for the compute they have committed to, how that changes with how fast AI improves, and what Anthropic's IPO share price needs. Built from public filings and named sources. It is a model, not a forecast: the person sets the assumptions, and the settings can be set to give almost any outcome. Not investment advice. Inputs last updated 2026-11-17; this file was built 2026-11-18. This file is for AI assistants and other programs. The page itself is drawn by a script in the browser, so its HTML holds little text. Everything you need is here and in the data file. - The page: https://lucas-klein.com/ai-lab-finances/ - The method in prose, with every source: https://lucas-klein.com/ai-lab-finances/about.html - Every number with its source, the settings, the scenarios and their results, as JSON: https://lucas-klein.com/ai-lab-finances/data.json - This guide: https://lucas-klein.com/ai-lab-finances/llms.txt This file is long (about 67,000 characters). Its order is: what you can do, how the model works, the link format, the questions to ask, what is not known, then the two long lists: every setting, and the ready-made scenarios with their results. If your tool cut the file short at 40,000 characters, you have lost only the last scenarios, in whole or in part; their ids are `start`, `lucas`, `old_pace`, `metr_grow`, `oai_race`, `both_race`, `aif`, `amodei`, `naam`, `zitron`, `cap_stall`, `rev_stall`, `adopt`, `bubble`, `open`, `moratorium`, `research_ban`, `cap_us` and `both_limits`, and all of them are in data.json under `scenarios`. Say so if you could not read them. ## What you can and cannot do - You can explain the model, ask a person the questions below, and build a link that opens the page with their answers as its settings. - You cannot run the model. It runs in the person's browser. The results of the ready-made scenarios are below and in data.json. For any other settings, give the person the link and let the page show the result. Do not work out or guess results yourself. - Many numbers in the model are guesses, and each one says so. In data.json every input has a `basis`: `filed` (an SEC filing), `leak` (a press report of leaked figures), `reported` (published by a named organisation), `forecast` (from the AI Futures Project's model), `context` (for comparison only), `assumed` (a choice; its `cite` says whether the site owner or the AI agent that built the page made it). When you explain a result, say which numbers behind it are guesses. Never present an assumed number as data. - Scenarios named after people or groups are inspired by what they have said in public. They did not make, check or endorse them. Say so if you mention one. - Use plain words. The site's owner is not a finance person and wants the page readable by people who are not. - Never present a result as a prediction. Say what it follows from: "on these settings, ...". The starting settings are one view, not the site's forecast. ## How the model works, in short - It runs by quarter from October 2026 to the end of 2032 for two labs, Anthropic and OpenAI, always together. - **The compute bill.** Anthropic's comes from its IPO filing: dated commitments, undated ones the model spreads over time, deals with no fixed price, and its short-term deal with SpaceX, kept only while needed at the start. OpenAI files nothing, so its bill is a leaked plan. On top of that each lab follows a strategy for buying more compute (the starting one: plan a year ahead for the revenue it expects, and put what is left after costs into research compute, aiming to break even). - **Capability.** How good the best AI model is, measured by Epoch AI's Capabilities Index (ECI). The forecast line is either a steady line built from Epoch AI's measured trends (the biggest training runs grow ×4 a year to 2030, better methods halve the compute needed every 8 months) or a finished run of the AI Futures Project's model, in which AI takes over AI research and progress explodes. A speed-up setting gives the pace today as a multiple of the old pace (about 15 ECI points a year, up to late 2025). The page opens on the steady line at ×1.9, the speed-up METR measured in 2026, held steady. - **Revenue.** The revenue of all AI labs together follows the best model: it multiplies by a set number for each year's worth of progress, up to a ceiling. Revenue chases that figure with an adoption delay that grows with the market: the time a market needs to double is the time the rule alone would take plus about 5 months when all labs together earn $1 trillion a year, twice that at $2 trillion. The site owner set the delay so that Anthropic's 2028 revenue on the starting settings equals the middle of a Manifold market on it. The ceiling is a placeholder and sets most of the result in the later years. - **Shares.** Each lab's share of that revenue follows how its research compute compares with the other's. On the starting settings only these two labs push the frontier, so their research compute also sets how fast the best model improves. That closes the loop: revenue pays for compute, compute moves capability, capability moves revenue. - **The price of compute.** The usual price is what Anthropic pays under its multi-year deals. SpaceX compute costs ×3 that and compute bought at short notice ×2 (the site owner's choices, from estimates; both are settings). New compute also costs more when the labs and everyone else together want more than the world builds. - **Money.** Each lab lives on its cash, its revenue and the money it raises: a list of raises for each lab (at the start Anthropic's IPO, about $81 billion, from its filing; OpenAI's round now closing, $15 billion, reported; and an OpenAI IPO in July 2027, $60 billion, the site owner's choice since no size is reported), and, when its cash would fall below 6 months of its spending, more money up to one year's revenue in any 12 months, only while its revenue grows. OpenAI's starting cash is from the end of March 2026, the latest found. - **Where the forecast stops.** The forecast ends at the first quarter's end where the best model passes ECI 250 or all labs together earn more than $5 trillion a year (both adjustable), or where either lab runs out of money: its cash and what it can raise cannot pay what it owes. If Anthropic is the one that runs out, before the end of 2031, its shares are worth nothing. Nothing after that is shown. - **The share price.** For Anthropic only: the price paid, grown at the return wanted to the end of 2031, gives the profit and so the revenue the company needs in 2031. The page compares that with the revenue the model gives, and shows the price that revenue supports. A forecast that stops before the end of 2031 is priced from its last full forecast year instead, with the shares held that many fewer years; one that stops before its first full year is over has no share price. Nothing after a stop is forecast, so tell the person to read a price worked from an early year with care. The share price is also unstable on many settings: it rests on one year's profit of a lab that plans to break even. ## The link format A link is the page's address with settings after a question mark: `https://lucas-klein.com/ai-lab-finances/?name=value&name=value` - `scenario=` starts from a ready-made scenario (ids below). Without it, the link starts from the starting settings. - `traj=` picks the capability line: `today` the steady line on Epoch AI's trends, `aifm_default` the AI Futures model at its default settings, `aifm_slow` a slower test run shipped with their code. - `lab=` picks which lab's tab is shown. - Every other name is a setting from the list below, written `group.name`: `link.` settings are shared, `anthropic.` and `openai.` settings belong to one lab, `vp.` settings are the share-price section. - Values are in the model's own units. Shares and yearly changes are fractions: `0.3` means 30%. `none` switches off a setting whose range says so. - Settings named in the link are applied on top of the scenario. A name the page does not have is ignored, and the page says so. A number outside a setting's range is moved to the nearest end. - The page keeps its own address up to date as the person changes settings, so a person can always copy the address back to you. Examples: - `https://lucas-klein.com/ai-lab-finances/?scenario=aif` : the AI Futures Project scenario as it is. - `https://lucas-klein.com/ai-lab-finances/?link.speedFactor=1&link.multiplier=3` : the starting settings, with AI improving at the old pace (no speed-up) and revenue multiplying by 3 per year of progress. - `https://lucas-klein.com/ai-lab-finances/?traj=aifm_default&link.speedFactor=1.9&link.trajStretch=2&link.marketAnchor=true&anthropic.strategy=capacity&anthropic.capacityShare=0.2&vp.price=90` : the AI Futures run started from the measured pace and unfolding twice as slowly, lifted to the prediction market's point, Anthropic securing 20% of the world's AI compute, at $90 a share. ## Questions to ask For a tutorial, ask these one at a time and explain each. For a quick set-up, ask a few at a time. Give the starting value, and let the person keep it by saying "skip". Translate their words into settings yourself; do not make them read setting names. 1. **How fast is AI improving, and will it take off?** The choices: steady, at the faster pace METR measured (the start: `traj=today` with `link.speedFactor=1.9`; METR's range is 1.5 to 2.8, and public scores show 1.3 to 1.6); steady at the old pace of before late 2025 (`link.speedFactor=1`); faster and still speeding up, or slowing down (`link.speedGrowth`: how much the pace changes each year, from -0.5 to 2; 0 holds it, 1 doubles it every year, -0.4 makes it fall 40% a year so capability levels off); or a takeoff as in the AI Futures model (`traj=aifm_default`). For a takeoff, `link.speedFactor` is the pace today: above 1 the run is started from a point where it goes that fast, which brings everything more than a year forward, and 1 leaves the run as published. Ask too whether it unfolds more slowly than their run (`link.trajStretch`, 1 to 3). 2. **Should the forecast be at least where the prediction market is?** Kalshi's market: even odds that a model scores above ECI 185 by Jan 23, 2027 (read Nov 17, 2026, 10:00 pm Central). `link.marketAnchor=true` lifts the line to that point as it stood when the page was published. Prediction markets on revenue, for comparison only (Manifold: play money, a handful of traders each; none was found on Polymarket or Kalshi): Anthropic, revenue in January to March 2027: middle $35B (7 traders); Anthropic, revenue in April to June 2027: middle $50B (11 traders); Anthropic, revenue in 2027: middle $240B (9 traders); Anthropic, revenue in October to December 2027: middle $75B (6 traders); Anthropic, revenue in 2028: middle $419B (8 traders); Anthropic, highest yearly pace reported by the end of 2026: middle $98B (56 traders); OpenAI, revenue in 2026: middle $38B (35 traders); OpenAI, highest yearly pace reported by the end of 2026: middle $68B (33 traders). The page draws them on the Revenue and bill chart. The model does not use them. 3. **How strongly does revenue follow capability?** `link.multiplier`: business revenue multiplies by this for each year's worth of progress (about 15 ECI points). 2 is the start and the low end of the AI Futures range; 5 is their middle; 12.5 their top. `link.consumerMultiplier` is the same for consumer revenue (1.5 at the start). If the person thinks benchmark scores overstate real usefulness, lower `link.capabilityShare` (0.9 is a mild version). 4. **How big can the market for AI get?** `link.ceiling`, in $ billions a year for all labs together. 10000 ($10 trillion) is a placeholder with no data behind it, about 8% of world output. `link.consumerCeiling` is the consumer part. Ask too how fast a large market takes up AI: `link.adoptDelay` is the adoption delay in years when all labs together earn $1 trillion a year (0.4 at the start; 0 = none), and `link.adoptGrow` how many times longer it gets each time the market doubles (2 at the start; 1 = the same at any size). 5. **What do the labs do about compute?** How long Anthropic keeps its short-term SpaceX compute (`anthropic.spacexNeeded`: true at the start, keeping from April 2027 only the SpaceX compute Anthropic needs after its long-term compute; false keeps it to its end in 2029; `anthropic.spacexStop`: none at the start, a date cancels it, e.g. 2027.25 for April 2027), and what that and short-notice compute cost (`link.spacexMultiple`, 3; `link.shortNotice`, 2). Keeping SpaceX moves the share price far more than its price does. For each lab, `strategy`: `plan` (the start: commit a year ahead for the revenue it expects; `planLead` years ahead, `planReinvest` the share of what is left that goes to research compute, 1 = break even), `capacity` (secure a share of all the AI compute the world builds, whatever its revenue: `capacityShare`, `capacityYear`), `margin`, `grow`, `pace` or `none`. 6. **How much money can the labs raise?** For each lab, `raises` lists its raises (e.g. `openai.raises=2026.9:15,2027.5:60:ipo`; `none` for no raises), and `raiseMultiple` how much it can raise when its cash runs low, as a multiple of a year's revenue in any 12 months (1 at the start, 0 = never; `raiseFloorMonths`, 6, is when it raises). A lab that cannot pay what it owes runs out of money and the model stops; if that is Anthropic, before the end of 2031, its shares are worth nothing. 7. **How fast does the world build AI compute?** `link.worldGrowth`, times a year from 2028: 1.36 at the start, 2.14 if the AI Futures pace goes on. A slower build-out makes new compute scarce and dearer. 8. **What could go wrong?** Nine switches, all off at the start: capability slows (`link.slowFrom`, `link.slowPace`); revenue stops following capability (`link.stallEci`, `link.multAfter`); slow adoption (`link.maxMarketGrowth`, `link.adoptFrom`); a fall in demand (`link.shockFrom`, `link.shockDrop`, `link.shockRecovery`); open models take the market (`link.openShare`, `link.openBy`); a limit on building AI compute, such as a data-centre moratorium, which stops new commitments while what the two labs have signed is still built (`link.buildFrom`, `link.buildGrowth`); a limit on AI research, such as a ban (`link.researchFrom`, `link.researchLimit`); a cap on training runs, everywhere or in the US only (`link.capFrom`, `link.capScope`); no new money, where neither lab raises anything beyond Anthropic's IPO (`link.noNewMoney`). Open models also compete by the share rule on every setting (`link.openNow`, `link.openGap`, under Competition). Each lab's margin over serving cost falls toward a commodity margin as its lead over the best alternative shrinks (`link.marginCompete`, `link.commodityMargin`, under Competition; revenue is not changed). The scenario list has a ready-made case of each. 9. **Where should the forecast stop?** `link.stopEci` (250 at the start) and `link.stopRevenue` (5000 at the start); 0 means never. On the starting settings it stops during 2029. On faster settings it stops within a year or two, and then the share price is worked from the last full forecast year, or there is none. The person may prefer a higher level, or no stop. 10. **The share price.** `vp.price` (the price paid per share), `vp.requiredReturn` (the yearly return wanted, 0.1 at the start), `vp.multiple` (company value per $1 of operating profit in 2031, 23 at the start), and `vp.researchBasis`: `committed` counts everything the forecast pays for compute in that year, so the profit is the forecast's own (the start), `standstill` only what it takes to stand still. The last one moves the result a great deal. 11. **Anything else they hold a view on.** Offer the rest of the settings list by section name, in plain words. Then build the link, list each setting you changed with its value, and say which of their answers the model is most sensitive to: the capability line and the speed-up, the revenue multiplier, the ceiling, the stop levels and `vp.researchBasis`. ## What is not known Say so if asked; do not fill these in. - Any lab's revenue other than Anthropic's (filed) and OpenAI's (press reports). Every other lab's is an assumption. - OpenAI's costs: it is measured with Anthropic's cost shares. - OpenAI's cash after March 2026, and how much of its March 2026 round was still to be paid. The size of its IPO, and any raise either lab makes after the ones listed. - The world's AI compute for each year from 2028 to 2032, and what buyers other than the two labs spend on it. - How demand responds to price. - How much more Anthropic pays for short-term (SpaceX) compute, or a lab for compute at short notice, than for multi-year compute. Anthropic's filing says "significantly more" and gives the amounts ($5.0 billion in July to September 2026, $6.5 billion in the first nine months of 2026), not the price. The ×3 for SpaceX rests on SemiAnalysis's estimates; the ×2 for short notice on cloud price lists. The filing describes the SpaceX deal; that its short-term compute is that deal is this page's reading, backed by a press report. - A real-money prediction market on either lab's revenue. None was found on Polymarket or Kalshi; the ones the page shows are on Manifold, where the stakes are play money and each market has a handful of traders. - How fast a large market takes up AI. The adoption delay was set to match one play-money market with 8 traders on Anthropic's 2028 revenue, and how it grows with size is the site owner's choice; no source gives either. - A capability score that marks superintelligence. The stop level of ECI 250 is this page's choice. - What buyers other than the two labs have already signed for. Under a limit on building, nothing is added for them. - How fast the speed-up METR measured keeps growing, if it does. METR says the faster pace looks more likely to last than to fall back. On a pace that keeps rising it sees no clear evidence in the scores and says evidence removed from its public report makes it more plausible; its forecast for a future model is removed too. - The AI Futures Project's forecast after METR's report. Their runs here are from their public code as of August 2026. - Who has proposed a data-centre moratorium or a ban on AI research, and in what form. The dates and sizes in those scenarios are this page's choice. - ECI scores far above today's are a straight line extended, not something a benchmark has measured. - Either lab's revenue by industry (banks, health, law and so on). Neither publishes it. The page's Who pays chart shows only what is published: Anthropic's revenue by how it is sold, where the customer is and how big; OpenAI's split between people and businesses (second-hand); what ChatGPT's users ask about (messages, not money); and businesses' spending on AI tools by use across all vendors (Menlo Ventures, an Anthropic investor). The forecast uses none of it. ## Settings Each line: the name to use in a link, what it is, its starting value, what it takes. ### Capability - `link.speedFactor`: AI improves faster than before, by. Starts at `1.9`. Takes 1 to 3; times the pace up to late 2025 (about 15 ECI points a year). 1 = the pace up to late 2025. METR measured about 1.9 for Anthropic (range 1.5 to 2.8), and the page opens on that. Public scores show 1.3 to 1.6. - `link.speedGrowth`: Each year that pace changes by. Starts at `0`. Takes -0.5 to 2; a fraction a year: 0 = the pace holds, 1 = it doubles every year, -0.5 = it halves every year. Only for the steady line (traj=today). The pace today (link.speedFactor) is multiplied by 1 + this for each year, so it speeds up or slows down exponentially; below 0 capability levels off. No source gives it: 0 is the start, and the values in the scenarios are the agent's choices. - `link.trajStretch`: From today, the run unfolds this many times slower. Starts at `1`. Takes 1 to 3; times. Only for an AI Futures run. At 1.5, what the run does in two years takes three. - `link.marketAnchor`: The line must reach the prediction market's point. Starts at `false`. Takes `true` or `false`. true: a line below ECI 185 on January 23, 2027 (Kalshi's even-odds date at publication) is lifted to pass through it. ### Revenue - `link.mode`: Where revenue comes from. Starts at `loop`. Takes one of `loop` (Capability, and labs compete for share), `fixed` (Capability, shares stay fixed), `path` (A path set by hand). In loop and fixed, all AI labs' revenue follows the capability line. In path, each lab's revenue is the hand-set yearly growth (anthropic.growths, openai.growths: six numbers for 2027 to 2032, separated by spaces, 1 = +100%). - `link.segmented`: Businesses and consumers as separate markets. Starts at `true`. Takes `true` or `false`. false: every customer in one market that follows link.multiplier. - `link.multiplier`: Business revenue multiplies by, per year of progress. Starts at `2`. Takes 1.5 to 12.5; times, per about 15 ECI points. The AI Futures Project's revenue rule. Their middle is 5, their range 2 to 12.5. The page starts at 2 (my choice). - `link.consumerMultiplier`: Consumer revenue multiplies by, per year of progress. Starts at `1.5`. Takes 1.1 to 12.5; times, per about 15 ECI points. The same rule for consumer revenue. 1.5 at the start, the agent's choice. This page's fit gave 2.8 when business revenue fitted 5. - `link.capabilityShare`: Of a gain in ECI, the part customers pay for. Starts at `1`. Takes 0.5 to 1; a fraction: 0.3 means 30%. ECI is a benchmark score. Below 1, the revenue rule counts less progress than the score. No source gives it. - `link.restConsumerShare`: Share of other labs' revenue from consumers today. Starts at `0.15`. Takes 0 to 1; a fraction: 0.3 means 30%. A guess: no other lab publishes it. - `link.ceiling`: Most all AI labs can earn in a year. Starts at `1e+04`. Takes 1000 to 30000; $ billions a year. A placeholder with no data behind it. It sets most of the later years' result. - `link.consumerCeiling`: Of which the most consumers can pay. Starts at `1000`. Takes 200 to 5000; $ billions a year. A placeholder. Must be at least 200 below link.ceiling. - `link.adoptDelay`: Adoption delay when all labs together earn $1 trillion a year. Starts at `0.4`. Takes 0 to 2; years; 0 = no delay. A market's doubling time is the rule's time plus this delay. 0.4 at the start: I set it so Anthropic's 2028 revenue equals the middle of a Manifold market ($419 billion, read November 17, 2026). - `link.adoptGrow`: The adoption delay gets longer, each time the market doubles, by. Starts at `2`. Takes 1 to 4; times. 2 at the start, my choice: the delay is 0.4 years at $1 trillion a year, 0.8 at $2 trillion, 0.2 at $500 billion. 1 = the same delay at every size. - `link.priceChange`: Price of a unit of work changes, each year. Starts at `0`. Takes -0.3 to 0.1; a fraction: 0.3 means 30%. With the cost of serving unchanged. 0 at the start. - `link.priceSensitivity`: Price sensitivity of demand. Starts at `1`. Takes 0 to 3; elasticity. 1: a price cut leaves revenue unchanged. Matters only when link.priceChange is not 0. Could not be measured. - `anthropic.runRate`: Revenue pace at the start (October 1, 2026). Starts at `71`. Takes 40 to 200; $ billions a year. Anthropic: $71B from its filing. OpenAI: $50B, a press report. - `openai.runRate`: Revenue pace at the start (October 1, 2026). Starts at `50`. Takes 20 to 150; $ billions a year. Anthropic: $71B from its filing. OpenAI: $50B, a press report. - `anthropic.consumerShare`: Share of the lab's revenue from consumers today. Starts at `0.22`. Takes 0 to 1; a fraction: 0.3 means 30%. Anthropic: subscriptions in its filing (too high as a consumer share). OpenAI: a second-hand press report. - `openai.consumerShare`: Share of the lab's revenue from consumers today. Starts at `0.5`. Takes 0 to 1; a fraction: 0.3 means 30%. Anthropic: subscriptions in its filing (too high as a consumer share). OpenAI: a second-hand press report. ### Where the forecast stops - `link.stopEci`: Stop when the best model passes ECI. Starts at `250`. Takes 0 to 400; ECI points; 0 = never. The forecast ends at the first quarter's end where the best model is above this. 250 at the start: the agent's choice, not confirmed. - `link.stopRevenue`: Stop when all labs' revenue passes. Starts at `5000`. Takes 0 to 20000; $ billions a year; 0 = never. 5000 at the start (half the starting ceiling): the agent's choice, not confirmed. ### Competition - `link.restRunRate`: Every other lab's revenue pace today. Starts at `87`. Takes 0 to 300; $ billions a year. Not published. From one research firm's estimate that the two labs held about 60% of the market. - `link.restEci`: Every other lab's best model today. Starts at `166.5`. Takes 150 to 190; ECI points. Measured by Epoch AI (Google's best model). - `link.shareSensitivity`: How strongly share follows capability. Starts at `1`. Takes 0 to 3; power. 0: shares never move. Only in link.mode=loop. - `link.switchSpeed`: How fast customers move. Starts at `0.7`. Takes 0.1 to 1; the part of the distance covered each year. 1 = at once. - `link.rest`: Who pushes the frontier. Starts at `leader`. Takes one of `leader` (Only Anthropic and OpenAI (the closed loop)), `pace` (Others too: the forecast holds whatever these two do). Only in link.mode=loop. - `link.openNow`: Open models' share of what customers pay today. Starts at `0.04`. Takes 0 to 0.3; a fraction: 0.04 means 4%; 0 = no competition from open models. Open models win customers by the same rule as a lab, from this share, and what they take pays the labs nothing. Menlo Ventures: 4% of US business spending on models in September 2026 (19% of the words processed). - `link.openGap`: Open models behind the best model today. Starts at `8`. Takes 0 to 30; ECI points. Epoch AI: 8 points through 2026, four to six months. They keep that lag in time: the gap in points grows when the line speeds up and shrinks when the frontier slows. - `link.marginCompete`: Margins fall as rivals catch up. Starts at `true`. Takes `true` or `false`. true: each lab's margin over serving cost holds at today's while its lead over the best alternative (every other lab, and open models) is at least today's, and falls in a straight line to link.commodityMargin as that lead goes to nothing. Revenue does not change. - `link.commodityMargin`: Margin once a lab has no lead. Starts at `0.5`. Takes 0 to 0.8; a fraction: 0.5 means 50%. Only with link.marginCompete. Fireworks AI, a host of open models: about 50% in July 2025 (third-hand). - `link.methodsGain`: Ten times the research compute gives research output of. Starts at `2.8`. Takes 1 to 10; times. The AI Futures model's figure is 2.8. Only in the closed loop. - `anthropic.startEci`: The lab's best model today. Starts at `173.4`. Takes 155 to 190; ECI points. Measured by Epoch AI on November 16, 2026. - `openai.startEci`: The lab's best model today. Starts at `169`. Takes 155 to 190; ECI points. Measured by Epoch AI on November 16, 2026. - `anthropic.ownEdge`: Extra progress of the lab's own. Starts at `0`. Takes -10 to 10; ECI points a year. On top of what its research compute gives. 0 at the start: history shows no edge either way. - `openai.ownEdge`: Extra progress of the lab's own. Starts at `0`. Takes -10 to 10; ECI points a year. On top of what its research compute gives. 0 at the start: history shows no edge either way. ### What could go wrong - `link.slowFrom`: 1. Capability slows down from the start of. Starts at `none`. Takes 2027 to 2031, whole numbers, or `none`; a year, 2027 to 2031; none = off. Above the level reached on that date, the same effort gives only link.slowPace of the progress. - `link.slowPace`: After that, the same effort gives. Starts at `1`. Takes 0 to 1; a fraction: 0.3 means 30%. Only with link.slowFrom. 0 = no progress. - `link.stallEci`: 2. Revenue stops following capability once the best model passes ECI. Starts at `none`. Takes 180 to 260, or `none`; ECI points; none = off. Past it, a year of progress multiplies revenue by link.multAfter. - `link.multAfter`: After that, revenue multiplies per year of progress by. Starts at `1`. Takes 1 to 2; times. Only with link.stallEci. 1 = flat. - `link.maxMarketGrowth`: 3. Adoption is slow: a market grows in a year by at most. Starts at `0`. Takes 0 to 5; a fraction: 0.3 means 30%; 0 = no limit. A hard limit on each market's growth, on top of the adoption delay under Revenue. - `link.adoptFrom`: The adoption limit applies from the start of. Starts at `none`. Takes 2028 to 2031, whole numbers, or `none`; a year; none = from the start. Only with link.maxMarketGrowth above 0. - `link.shockFrom`: 4. Demand falls during. Starts at `none`. Takes 2027 to 2031, whole numbers, or `none`; a year, 2027 to 2031; none = off. Over that year demand falls to link.shockDrop below the rule's path. - `link.shockDrop`: Demand ends the year below the rule's path by. Starts at `0`. Takes 0 to 0.9; a fraction: 0.3 means 30%. Only with link.shockFrom. - `link.shockRecovery`: Of what was lost, comes back each year. Starts at `0`. Takes 0 to 1; a fraction: 0.3 means 30%. 0 = none of it comes back. - `link.openShare`: 5. Open models take even more, of what the labs would have earned. Starts at `0`. Takes 0 to 0.9; a fraction: 0.3 means 30%. On top of what open models win as competitors (link.openNow). Rises in a straight line from nothing today to this share at the end of link.openBy. - `link.openBy`: They reach that share by the end of. Starts at `2030`. Takes 2027 to 2032, whole numbers; a year. Only with link.openShare above 0. - `link.buildFrom`: 6. A limit on building AI compute from the start of. Starts at `none`. Takes 2027 to 2031, whole numbers, or `none`; a year, 2027 to 2031; none = off. From that date the world's AI compute grows by at most link.buildGrowth a year. Only new commitments stop; what the two labs already signed is still built. - `link.buildGrowth`: After that, the world's AI compute grows in a year by at most. Starts at `0`. Takes 0 to 1.5; a fraction: 0.3 means 30%; 0 = a moratorium. Only with link.buildFrom. - `link.researchFrom`: 7. A limit on AI research from the start of. Starts at `none`. Takes 2027 to 2031, whole numbers, or `none`; a year, 2027 to 2031; none = off. From that date each lab's research compute is held at link.researchLimit of the last quarter's. Signed compute it cannot use is still paid for. - `link.researchLimit`: After that, research compute is held at. Starts at `1`. Takes 0 to 1; a fraction: 0.3 means 30%; 0 = a ban, 1 = a freeze. Only with link.researchFrom. - `link.capFrom`: 8. A cap on training runs from the start of. Starts at `none`. Takes 2027 to 2031, whole numbers, or `none`; a year, 2027 to 2031; none = off. No training run may be bigger than the largest before that date. Labs still find better methods, still serve customers, and hold research compute where it was (unless link.researchFrom says otherwise). Other labs and open models keep their lag, so the gaps shrink. - `link.capScope`: The cap holds. Starts at `world`. Takes one of `world` (Everywhere (a treaty)), `us` (In the US only). Only with link.capFrom. In the US only, open models keep the pace the frontier would have had and stay their lag behind it. - `link.noNewMoney`: 9. No new money. Starts at `false`. Takes `true` or `false`. true: neither lab raises any more money. Each lives on its cash and revenue, signs new compute only as far as its cash would last a year with revenue flat (research gives way first), and runs out of money when its cash cannot pay the compute it already owes. The forecast then stops at the end of that quarter, for both labs, like the ECI and revenue stops. If the lab that runs out is Anthropic, before the end of 2031, its shares are worth nothing. ### Strategy: new compute - `link.hardwareGain`: Compute a dollar buys grows, each year. Starts at `1.49`. Takes 1 to 2; times a year: 1.49 means +49%. Turns the capability forecast's compute path into dollars. - `link.worldGrowth`: The world's AI compute grows, each year from 2028. Starts at `1.36`. Takes 1.1 to 3; times a year. 1.36 at the start (to McKinsey's middle case for 2030, second-hand). The AI Futures pace is 2.14. - `link.gwCost`: A gigawatt of AI compute costs, a year. Starts at `12.5`. Takes 8 to 25; $ billions. Turns gigawatts into dollars. The agent's choice (the middle of Dario Amodei's $10 to $15 billion). - `link.restBudget`: Everyone else's spending on AI compute, as a share of the world's. Starts at `0.67`. Takes 0.3 to 1; a fraction: 0.3 means 30%. The agent's choice, not confirmed. Sets when new compute gets scarce. - `link.shortNotice`: Compute bought at short notice costs. Starts at `2`. Takes 1 to 3; times the usual price. Compute bought beyond the committed bill, when demand outruns it or research would fall below last quarter's. 2 at the start: my choice. Cloud prices put short notice at about ×1.5 to ×2.5 the price of committed compute. - `anthropic.strategy`: What the lab does beyond what is signed. Starts at `plan`. Takes one of `none` (Sign nothing more), `margin` (Reinvest profit above a set share), `grow` (Compute first: grow the bill yearly), `pace` (Follow the forecast's compute path), `plan` (Plan ahead for expected revenue), `capacity` (Secure a share of the world's AI compute). plan at the start, for both labs. - `openai.strategy`: What the lab does beyond what is signed. Starts at `plan`. Takes one of `none` (Sign nothing more), `margin` (Reinvest profit above a set share), `grow` (Compute first: grow the bill yearly), `pace` (Follow the forecast's compute path), `plan` (Plan ahead for expected revenue), `capacity` (Secure a share of the world's AI compute). plan at the start, for both labs. - `anthropic.planLead`: Compute is committed this many years ahead. Starts at `1`. Takes 0 to 3, whole numbers; years. Only with strategy=plan. A guess: no lab publishes it. - `openai.planLead`: Compute is committed this many years ahead. Starts at `1`. Takes 0 to 3, whole numbers; years. Only with strategy=plan. A guess: no lab publishes it. - `anthropic.planReinvest`: Research compute, share of what is left. Starts at `1`. Takes 0 to 1.5; a fraction: 0.3 means 30%. Only with strategy=plan. 1 = plans to break even; above 1 a planned loss. - `openai.planReinvest`: Research compute, share of what is left. Starts at `1`. Takes 0 to 1.5; a fraction: 0.3 means 30%. Only with strategy=plan. 1 = plans to break even; above 1 a planned loss. - `anthropic.capacityShare`: Share of the world's AI compute it secures. Starts at `0.14`. Takes 0.05 to 0.6; a fraction: 0.3 means 30%. Only with strategy=capacity. - `openai.capacityShare`: Share of the world's AI compute it secures. Starts at `0.2`. Takes 0.05 to 0.6; a fraction: 0.3 means 30%. Only with strategy=capacity. - `anthropic.capacityYear`: Reaches that share by the end of. Starts at `2030`. Takes 2027 to 2032, whole numbers; a year. Only with strategy=capacity. - `openai.capacityYear`: Reaches that share by the end of. Starts at `2030`. Takes 2027 to 2032, whole numbers; a year. Only with strategy=capacity. - `anthropic.targetMargin`: Operating profit kept, share of revenue. Starts at `0`. Takes -0.5 to 0.4; a fraction: 0.3 means 30%. Only with strategy=margin. - `openai.targetMargin`: Operating profit kept, share of revenue. Starts at `0`. Takes -0.5 to 0.4; a fraction: 0.3 means 30%. Only with strategy=margin. - `anthropic.billGrowth`: The bill grows at least. Starts at `0.31`. Takes 0 to 1.5; a fraction: 0.3 means 30% a year. Only with strategy=grow. - `openai.billGrowth`: The bill grows at least. Starts at `0.31`. Takes 0 to 1.5; a fraction: 0.3 means 30% a year. Only with strategy=grow. - `anthropic.strategyUntil`: Keeps doing this through. Starts at `2032`. Takes 2028 to 2032, whole numbers; a year. After it the lab signs nothing more. - `openai.strategyUntil`: Keeps doing this through. Starts at `2032`. Takes 2028 to 2032, whole numbers; a year. After it the lab signs nothing more. - `anthropic.maxComputeGrowth`: Most compute spending can grow in a year. Starts at `3`. Takes 0 to 6; times a year; 0 = no limit. A starting figure: nothing fixes it. - `openai.maxComputeGrowth`: Most compute spending can grow in a year. Starts at `3`. Takes 0 to 6; times a year; 0 = no limit. A starting figure: nothing fixes it. - `anthropic.unfixedAmount`: Deals agreed with no fixed price: amount counted. Starts at `162.5`. Takes 0 to 175; $ billions. The filing estimates $150B to $175B. 0 = left out. - `anthropic.unfixedYears`: Those deals are paid over. Starts at `4`. Takes 4 to 9, whole numbers; years, from 2028. The filing gives no payment dates. - `anthropic.raiseMultiple`: Raise money when cash runs low: at most this many times a year's revenue in any 12 months. Starts at `1`. Takes 0 to 5; × a year's revenue (0 = never). 1 at the start, for both labs (my choice, November 18, 2026). The listed raises count toward it. Money comes only while revenue is above the same quarter's a year before. For scale: OpenAI raised about 11 times its revenue in 2025, Anthropic about 1.4 times in May 2026. - `openai.raiseMultiple`: Raise money when cash runs low: at most this many times a year's revenue in any 12 months. Starts at `1`. Takes 0 to 5; × a year's revenue (0 = never). 1 at the start, for both labs (my choice, November 18, 2026). The listed raises count toward it. Money comes only while revenue is above the same quarter's a year before. For scale: OpenAI raised about 11 times its revenue in 2025, Anthropic about 1.4 times in May 2026. - `anthropic.raiseFloorMonths`: It raises when cash would fall below this many months of spending. Starts at `6`. Takes 0 to 24; months. 6 at the start: the agent's choice, not confirmed. No lab publishes a floor. - `openai.raiseFloorMonths`: It raises when cash would fall below this many months of spending. Starts at `6`. Takes 0 to 24; months. 6 at the start: the agent's choice, not confirmed. No lab publishes a floor. ### Margins - `anthropic.servingShare`: Compute to serve customers, per $1 of revenue. Starts at `0.1448`. Takes 0.05 to 0.8; a fraction: 0.3 means 30%. Anthropic's filing: 14.5 cents. Used for OpenAI too, as a yardstick. - `openai.servingShare`: Compute to serve customers, per $1 of revenue. Starts at `0.1448`. Takes 0.05 to 0.8; a fraction: 0.3 means 30%. Anthropic's filing: 14.5 cents. Used for OpenAI too, as a yardstick. - `anthropic.servingDrift`: Change in that serving cost, each year. Starts at `0`. Takes -0.3 to 0.3; a fraction: 0.3 means 30%. 0 at the start: no data picks a number. - `openai.servingDrift`: Change in that serving cost, each year. Starts at `0`. Takes -0.3 to 0.3; a fraction: 0.3 means 30%. 0 at the start: no data picks a number. - `anthropic.researchShare`: Research compute, share of revenue today. Starts at `0.421`. Takes 0 to 1; a fraction: 0.3 means 30%. From Anthropic's filing. Used only for the 'revenue needed at today's margins' line. - `openai.researchShare`: Research compute, share of revenue today. Starts at `0.421`. Takes 0 to 1; a fraction: 0.3 means 30%. From Anthropic's filing. Used only for the 'revenue needed at today's margins' line. - `anthropic.otherShare`: All other costs, share of revenue. Starts at `0.4721`. Takes 0.1 to 0.8; a fraction: 0.3 means 30%. From Anthropic's filing, held in every year. - `openai.otherShare`: All other costs, share of revenue. Starts at `0.4721`. Takes 0.1 to 0.8; a fraction: 0.3 means 30%. From Anthropic's filing, held in every year. ### Money raised - `anthropic.raises`: The raises. Starts at `2026.88:80.839:ipo:done`. Takes a list of raises as in the help, or `none`. Each raise as when:amount ($ billions), with :ipo for an IPO and :done for money already raised (No new money keeps only those), separated by commas; none for no raises. A raise counts in the quarter its date falls in. Anthropic at the start: its IPO and private placement, 2026.88:80.84:ipo:done (from its filing, at $74 a share). OpenAI: the round now closing, 2026.9:15 (reported), and an IPO in July 2027, 2027.5:60:ipo (the date where prediction markets put even odds; the amount is my guess). - `openai.raises`: The raises. Starts at `2026.9:15,2027.5:60:ipo`. Takes a list of raises as in the help, or `none`. Each raise as when:amount ($ billions), with :ipo for an IPO and :done for money already raised (No new money keeps only those), separated by commas; none for no raises. A raise counts in the quarter its date falls in. Anthropic at the start: its IPO and private placement, 2026.88:80.84:ipo:done (from its filing, at $74 a share). OpenAI: the round now closing, 2026.9:15 (reported), and an IPO in July 2027, 2027.5:60:ipo (the date where prediction markets put even odds; the amount is my guess). ### Anthropic's commitments - `anthropic.rampMonths`: Undated payments stop rising after this many months from October 2026. Starts at `27`. Takes 12 to 51; months. How the model spreads the filing's undated commitments. - `anthropic.sched27ShareJanSep`: Share of 2027's dated payments due January to September. Starts at `0.7`. Takes 0.5 to 0.9; a fraction: 0.3 means 30%. An assumption, needed to use the filing's 12-month total. - `anthropic.conditionalDrop`: Conditional commitments that fall away. Starts at `0`. Takes 0 to 1; a fraction: 0.3 means 30%. Of the $186.2B the filing says can still be cancelled. - `link.spacexMultiple`: SpaceX compute costs. Starts at `3`. Takes 1 to 5; times the usual price. The usual price is what Anthropic pays under its multi-year deals. 3 at the start: my choice. SemiAnalysis puts SpaceX's price at up to $50B a gigawatt a year, against about $12B for multi-year rental. - `anthropic.spacexStop`: SpaceX deal cancelled: nothing paid from. Starts at `none`. Takes 2027.25 to 2029.5, or `none`; a date in years (2027.25 = April 1, 2027); none = kept to its end in 2029. Either side can cancel on 90 days' notice after January and March 2027. none at the start: kept only while needed (anthropic.spacexNeeded), or with anthropic.spacexNeeded false kept to its end in 2029. - `anthropic.spacexNeeded`: SpaceX deal kept only while needed. Starts at `true`. Takes `true` or `false`. true: from April 2027 Anthropic keeps only the SpaceX compute it needs to serve customers and hold research at last quarter's level, after its long-term compute, never more than the quarter before; its strategy signs long-term compute in its place. true at the start: my choice. false: kept to its end in 2029. anthropic.spacexStop still cancels it from a date. ### OpenAI's compute plan - `openai.openaiTotal`: Planned compute spend through 2030. Starts at `600`. Takes 600 to 856; $ billions. Press reports of leaked plans give 600, 665 and 856. ### Share price - `vp.price`: Price paid per share. Starts at `74`. Takes 30 to 250; $. The filing's IPO range is in the data file (anth.ipo_price_low, anth.ipo_price_high). - `vp.basis`: Shares counted. Starts at `promised`. Takes one of `promised` (All shares already promised), `ipo` (Only shares outstanding at the IPO). - `vp.requiredReturn`: Return you want. Starts at `0.1`. Takes 0 to 0.3; a fraction: 0.3 means 30% a year. To the end of 2031, or of the last full forecast year when the forecast stops sooner. - `vp.multiple`: Company value per $1 of operating profit, in the year the price is worked from (2031, or the last full forecast year). Starts at `23`. Takes 8 to 50; times. - `vp.researchBasis`: Compute counted against that year's profit. Starts at `committed`. Takes one of `committed` (Everything committed for that year), `standstill` (What it takes to stand still). Moves the share price a great deal. No data settles it. - `vp.longRunMargin`: Highest operating margin in that year. Starts at `0.3`. Takes 0.05 to 0.6; a fraction: 0.3 means 30%. ## Scenarios, and what the model gives for each Dollar figures are a year's revenue. Where the forecast stops, only the years before the stop are listed. ### Starting settings (`scenario=start`) The settings the page opens with. AI improves at the faster pace METR measured (×1.9 the pace up to late 2025, one step up and then steady). Revenue follows capability by the simple rule, with an adoption delay that grows with the market (about 5 months when all labs together earn $1 trillion a year, twice that at $2 trillion). Both labs plan a year ahead and aim to break even. Nothing under What could go wrong is on. Where it comes from: My choices of November 17, 2026: open on METR's pace, and set the adoption delay so that Anthropic's revenue for 2028 equals the middle of Manifold's market on it ($419 billion that day; play money, 8 traders). No source gives the delay. Every other scenario starts here and changes only what it names. Note: METR measured the speed-up. It did not make these settings, check them or endorse them. Settings it changes: none. Result: the forecast stops at the end of September 2029 (the best AI model passes ECI 250); nothing after that is forecast. Anthropic's revenue 2027: $155B, 2028: $405B. Best model's ECI at the end of each of those years: 205, 233. Share price Anthropic's revenue supports: about $42, worked from 2028, the last full forecast year. Money Anthropic raises: $284B. ### Lucas's settings (`scenario=lucas`) My own view. Capability follows Epoch AI's trends, going ×1.5 the old pace today (the low end of the speed-up METR measured), and lifted to the prediction market's point. Customers pay for 90% of a gain in ECI. Adoption follows the starting adoption delay, and each market grows at most 250% a year (×3.5). The world's AI compute grows ×1.8 a year from 2028, a little slower than the AI Futures pace of ×2.14. Revenue multiplies ×2 for each year of progress, as on the starting settings. Where it comes from: My words: close to the AI Futures Project on capability but "slightly slower, like the slower end"; ECI "isn't a perfect measurement for capability", a mild effect; "a little bit of a slowdown in compute build out". On November 17, 2026 I set the speed-up to METR's low end, capped each market at 250% a year on top of the adoption delay, and moved capability from the AI Futures run to Epoch AI's trends. Two numbers were picked by this page's AI agent to fit my words and are not confirmed: the 90% and the ×1.8. Settings it changes: `link.speedFactor=1.5`, `link.marketAnchor=true`, `link.capabilityShare=0.9`, `link.maxMarketGrowth=2.5`, `link.worldGrowth=1.8`. Result: the forecast stops at the end of 2029 (the best AI model passes ECI 250); nothing after that is forecast. Anthropic's revenue 2027: $181B, 2028: $365B, 2029: $706B. Best model's ECI at the end of each of those years: 206, 228, 252. Share price Anthropic's revenue supports: about $8, worked from 2029, the last full forecast year. Money Anthropic raises: $312B. ### No speed-up: the old pace (`scenario=old_pace`) AI improves at its pace up to late 2025, about 15 ECI points a year, as if the speed-up METR measured had not happened or does not last. Where it comes from: Epoch AI's public scores show a smaller speed-up than METR's ×1.9: ×1.6 for Anthropic's best released model, ×1.3 for the best released model of any lab, and almost none for OpenAI. Note: The agent's scenario. Epoch AI is named only as the source of the scores: it did not make, check or endorse it. Settings it changes: `link.speedFactor=1`. Result: the forecast stops at the end of September 2031 (OpenAI runs out of money); nothing after that is forecast. Anthropic's revenue 2027: $113B, 2028: $204B, 2029: $361B, 2030: $505B. Best model's ECI at the end of each of those years: 190, 205, 219, 227. Share price Anthropic's revenue supports: about $4, worked from 2030, the last full forecast year. Money Anthropic raises: $446B. OpenAI runs out of money by September 2031. ### The measured speed-up keeps growing (`scenario=metr_grow`) AI improves ×1.9 as fast as before, as on the starting settings, and the pace then doubles again every year. Where it comes from: METR sees no clear evidence in the scores of a pace that keeps rising, and says other evidence, removed from its public report, makes it more plausible. The one-year doubling time is the agent's choice, not confirmed. Note: METR measured the speed-up. It did not make this scenario, check it or endorse it, and its own forecast is blacked out in the public report. Settings it changes: `link.speedFactor=1.9`, `link.speedGrowth=1`. Result: the forecast stops at the end of June 2028 (the best AI model passes ECI 250); nothing after that is forecast. Anthropic's revenue 2027: $183B. Best model's ECI at the end of each of those years: 222. Share price Anthropic's revenue supports: about $32, worked from 2027, the last full forecast year. Money Anthropic raises: $81B. ### OpenAI races for compute (`scenario=oai_race`) OpenAI secures 20% of all the AI compute the world builds by the end of 2027 and holds that share, whatever its revenue. Anthropic keeps planning a year ahead. Where it comes from: The 20% is the AI Futures Project's figure for OpenAI at the end of 2027 (compute forecast, April 2025). How much compute the world builds is a rough path. Note: The 20% is the AI Futures Project's figure; the scenario is the agent's. They did not make it, check it or endorse it. Settings it changes: `openai.strategy=capacity`, `openai.capacityShare=0.2`, `openai.capacityYear=2027`. Result: the forecast stops at the end of September 2029 (the best AI model passes ECI 250); nothing after that is forecast. Anthropic's revenue 2027: $155B, 2028: $405B. Best model's ECI at the end of each of those years: 205, 233. Share price Anthropic's revenue supports: about $40, worked from 2028, the last full forecast year. Money Anthropic raises: $287B. ### Both labs race for compute (`scenario=both_race`) Each lab secures 20% of all the AI compute the world builds by the end of 2027 and holds that share, whatever its revenue. Where it comes from: 20% is the AI Futures Project's figure for OpenAI. For Anthropic it is the agent's choice: their figure for Anthropic is 14%, less than planning ahead already buys it here. Note: The 20% is the AI Futures Project's figure; the scenario is the agent's. They did not make it, check it or endorse it. Settings it changes: `anthropic.strategy=capacity`, `anthropic.capacityShare=0.2`, `anthropic.capacityYear=2027`, `openai.strategy=capacity`, `openai.capacityShare=0.2`, `openai.capacityYear=2027`. Result: the forecast stops at the end of September 2029 (the best AI model passes ECI 250); nothing after that is forecast. Anthropic's revenue 2027: $155B, 2028: $396B. Best model's ECI at the end of each of those years: 205, 235. Share price Anthropic's revenue supports: about $35, worked from 2028, the last full forecast year. Money Anthropic raises: $97B. ### AI Futures Project (`scenario=aif`) Capability follows their model at its default settings, started today where the run first goes ×1.9 the old pace (the speed-up METR measured) and lifted to the prediction market's point. On their own clock that point comes in early 2028, so the takeoff comes more than a year sooner here. Business revenue multiplies ×5 for each year of progress (their middle figure) and consumer revenue ×2.8. The world's AI compute keeps growing ×2.1 a year after 2027. No adoption delay: their rule has none. Where it comes from: Theirs: the run and the ×5. Not theirs: starting the run from METR's pace and lifting it to the market (my requests of November 17, 2026: the takeoff "should have already started just now", and a scenario named after a forecaster should not predict less progress than the prediction market; how it is done is the agent's construction); the consumer figure (the agent's fit); the compute growth after 2027 (their last year's pace carried on, 62 gigawatts against 29); and the $10 trillion ceiling on all labs' revenue. Their own forecast after METR's report is not in their public code and may differ. Note: Inspired by what the AI Futures Project has published. The name is only a label: they did not make this scenario, check it or endorse it, and this page has no link to them. Settings it changes: `traj=aifm_default`, `link.multiplier=5`, `link.consumerMultiplier=2.8`, `link.worldGrowth=2.138`, `link.speedFactor=1.9`, `link.marketAnchor=true`, `link.adoptDelay=0`. Result: the forecast stops at the end of June 2027 (all AI labs together earn more than $5.00T a year); nothing after that is forecast. No full year is forecast, so there is no yearly revenue to give. No share price: the forecast stops before its first full year is over. Money Anthropic raises: $81B. ### Dario Amodei (`scenario=amodei`) Capability follows the AI Futures default run, the nearest here to Dario Amodei's "country of geniuses in a data center" within one to three years, started today where the run first goes ×1.9 the old pace (the speed-up METR measured) and lifted to the prediction market's point. Revenue multiplies ×1.5 for each year of progress for businesses and ×1.3 for consumers. Anthropic plans a year ahead and puts 80% of what is left into research compute, not 100%. No adoption delay: none is stated. Where it comes from: Dario Amodei's words in 2026: "hard for me to see that there won't be trillions of dollars in revenue before 2030"; "I plan for the lower side". The numbers that turn these into settings are the agent's reading and are not confirmed. The early start and the lift to the market are my requests, as for the AI Futures Project. The world's build-out is left at its starting value, which this path outruns: new compute becomes scarce and its price rises. Note: Inspired by what Dario Amodei has said in public. The name is only a label: neither they nor Anthropic made this scenario, checked it or endorsed it, and this page has no link to either. Settings it changes: `traj=aifm_default`, `link.multiplier=1.5`, `link.consumerMultiplier=1.3`, `link.speedFactor=1.9`, `link.marketAnchor=true`, `link.adoptDelay=0`, `anthropic.planReinvest=0.8`. Result: the forecast stops at the end of September 2027 (the best AI model passes ECI 250); nothing after that is forecast. No full year is forecast, so there is no yearly revenue to give. No share price: the forecast stops before its first full year is over. Money Anthropic raises: $81B. ### Ramez Naam (`scenario=naam`) Capability keeps the steady pace it had up to late 2025, with no takeoff and no speed-up: Ramez Naam puts the frontier at about 16 ECI points a year, and this page's line gains about 15. All AI labs together can earn at most $3 trillion a year. No adoption delay: none is stated, and the $3 trillion ceiling does that work. Where it comes from: Naam is not a bear on revenue: "I can see our way to a trillion, a couple trillion, three trillion in revenue going to AI companies". What they argue against is an intelligence explosion. Using the $3 trillion as the ceiling is the agent's reading, not confirmed. Note: Inspired by what Ramez Naam has said in public. The name is only a label: they did not make this scenario, check it or endorse it, and this page has no link to them. Settings it changes: `link.speedFactor=1`, `link.ceiling=3000`, `link.consumerCeiling=300`, `link.adoptDelay=0`. Result: the forecast stops at the end of March 2031 (OpenAI runs out of money); nothing after that is forecast. Anthropic's revenue 2027: $111B, 2028: $192B, 2029: $315B, 2030: $391B. Best model's ECI at the end of each of those years: 190, 205, 219, 226. Share price Anthropic's revenue supports: about $3, worked from 2030, the last full forecast year. Money Anthropic raises: $225B. OpenAI runs out of money by March 2031. ### Ed Zitron (`scenario=zitron`) No new money: OpenAI raises nothing more and Anthropic keeps only its IPO money; each lab lives on its cash and revenue, and fails when it cannot pay the compute it owes. Revenue stops growing with capability once the best model passes ECI 175, about today's best. AI's pace slows 50% a year from today (it halves every year), so progress levels off. The forecast stops at the end of March 2028, when OpenAI runs out of money. Prices rise 10% a year as the labs stop selling below cost, and customers are price-sensitive (1.5), so some leave. Where it comes from: Ed Zitron in 2026: the bubble breaks "when the money runs out"; Anthropic and OpenAI "cannot afford to pay their bills without a near-infinite flow of venture capital investments"; "every single AI service you use subsidized compute", and at their actual rates "many would've bounced off the product"; "AI is not working." He gives no numbers: ECI 175, the slowdown (from his 2025 writing on diminishing returns), the 10% and the 1.5 are the agent's choices, not confirmed. Note: Inspired by what Ed Zitron has said in public. The name is only a label: they did not make this scenario, check it or endorse it, and this page has no link to them. Settings it changes: `link.noNewMoney=true`, `link.stallEci=175`, `link.multAfter=1`, `link.priceChange=0.1`, `link.priceSensitivity=1.5`, `link.speedGrowth=-0.5`, `openai.raises=[]`. Result: the forecast stops at the end of March 2028 (OpenAI runs out of money); nothing after that is forecast. Anthropic's revenue 2027: $79B. Best model's ECI at the end of each of those years: 196. Share price Anthropic's revenue supports: about $9, worked from 2027, the last full forecast year. Money Anthropic raises: $81B. OpenAI runs out of money by March 2028. ### Capability stalls (`scenario=cap_stall`) AI's pace slows 40% a year from today (it halves about every 16 months), so the best model levels off near ECI 220 instead of passing 250. Every lab ends near the same level, so customers move away from the lab that was ahead. Where it comes from: Argued by Gary Marcus and Yann LeCun. None of them gives numbers for it: the numbers are the agent's choice, not confirmed. Note: Inspired by what the people named have argued in public. None of them made, checked or endorsed this scenario or its numbers. Settings it changes: `link.speedGrowth=-0.4`. Result: the forecast stops at the end of 2031 (OpenAI runs out of money); nothing after that is forecast. Anthropic's revenue 2027: $139B, 2028: $267B, 2029: $399B, 2030: $503B, 2031: $525B. Best model's ECI at the end of each of those years: 198, 210, 218, 221, 221. Share price Anthropic's revenue supports: $0. Money Anthropic raises: $493B. OpenAI runs out of money by December 2031. ### Revenue stops following capability (`scenario=rev_stall`) Models keep improving, but once the best one passes ECI 190 customers stop paying much more for it: a further year of progress multiplies revenue by ×1.1, not ×2. Where it comes from: The "good enough" argument. Nobody who argues it gives numbers: the numbers are the agent's choice, not confirmed. Settings it changes: `link.stallEci=190`, `link.multAfter=1.1`. Result: the forecast stops at the end of September 2029 (the best AI model passes ECI 250); nothing after that is forecast. Anthropic's revenue 2027: $135B, 2028: $181B. Best model's ECI at the end of each of those years: 205, 233. Share price Anthropic's revenue supports: about $6, worked from 2028, the last full forecast year. Money Anthropic raises: $179B. ### Slow adoption (`scenario=adopt`) Each market grows at most 30% a year however good the models get, a hard limit on top of the adoption delay: firms take years to change how they work. Where it comes from: Argued by Arvind Narayanan and Sayash Kapoor ("Diffusion occurs over decades, not years") and by Daron Acemoglu. None of them gives numbers for it: the numbers are the agent's choice, not confirmed. Note: Inspired by what the people named have argued in public. None of them made, checked or endorsed this scenario or its numbers. Settings it changes: `link.maxMarketGrowth=0.3`. Result: the forecast stops at the end of September 2029 (the best AI model passes ECI 250); nothing after that is forecast. Anthropic's revenue 2027: $87B, 2028: $113B. Best model's ECI at the end of each of those years: 205, 233. Share price Anthropic's revenue supports: about $5, worked from 2028, the last full forecast year. Money Anthropic raises: $183B. ### The bubble pops (`scenario=bubble`) During 2028 demand falls to 55% below the rule's path, a fall of about 15% in actual revenue, and none of it comes back. From 2029 each market grows at most 10% a year. Compute committed before the fall is still owed. Where it comes from: The size of the fall is Cisco's after 2000: revenue down 15% in a year, then flat. Argued by Ed Zitron, Michael Burry and Jeremy Grantham. The date and the slow growth afterwards are the agent's choice, not confirmed. Note: Inspired by what the people named have argued in public. None of them made, checked or endorsed this scenario or its numbers. Settings it changes: `link.shockFrom=2028`, `link.shockDrop=0.55`, `link.shockRecovery=0`, `link.maxMarketGrowth=0.1`, `link.adoptFrom=2029`. Result: the forecast stops at the end of September 2029 (the best AI model passes ECI 250); nothing after that is forecast. Anthropic's revenue 2027: $155B, 2028: $342B. Best model's ECI at the end of each of those years: 205, 233. Share price Anthropic's revenue supports: about $10, worked from 2028, the last full forecast year. Money Anthropic raises: $251B. ### Open models take the market (`scenario=open`) Open models compete by the same rule as a lab on every setting. Here they also take, by the end of 2030, 60% of what customers would have paid the labs, and the labs cut the price of the rest by 30% a year. Customers are not very price-sensitive (0.5), so a cheaper price brings only a little more use and the labs earn less. Where it comes from: For scale: open models were about 4 months behind the best in 2026 (Epoch AI) and 4% of US business spending on models in September 2026, though 19% of the words processed (Menlo Ventures). Nobody who argues this gives numbers: the 60% and the 30% are the agent's choice, not confirmed; the price sensitivity of 0.5 is mine (2026-11-18: at 1 the price cut left revenue unchanged). Note: Epoch AI and Menlo Ventures are named only as the sources of figures for scale. Neither made, checked or endorsed this scenario. Settings it changes: `link.openShare=0.6`, `link.openBy=2030`, `link.priceChange=-0.3`, `link.priceSensitivity=0.5`. Result: the forecast stops at the end of September 2029 (the best AI model passes ECI 250); nothing after that is forecast. Anthropic's revenue 2027: $125B, 2028: $267B. Best model's ECI at the end of each of those years: 205, 232. Share price Anthropic's revenue supports: about $8, worked from 2028, the last full forecast year. Money Anthropic raises: $185B. ### Data-centre moratorium (`scenario=moratorium`) From the start of 2028 nobody may commit to new AI compute. What the two labs had already signed for is still built, on top of what the world had then. Nothing else is added, so new compute becomes scarce and its price rises. Where it comes from: A pause on new data centres has been called for in public; who proposed what was not looked up. The date and a full stop everywhere are the agent's choice, not confirmed. My rule (November 17, 2026): a moratorium stops new commitments only, because what is already signed has its approvals. What other buyers had signed is not known, so nothing is added for them (a later date allows for it). Compute a lab planned to buy but had not signed counts as new. Settings it changes: `link.buildFrom=2028`, `link.buildGrowth=0`. Result: the forecast stops at the end of September 2029 (the best AI model passes ECI 250); nothing after that is forecast. Anthropic's revenue 2027: $155B, 2028: $405B. Best model's ECI at the end of each of those years: 205, 233. Share price Anthropic's revenue supports: about $39, worked from 2028, the last full forecast year. Money Anthropic raises: $288B. ### Cap on training runs (`scenario=research_ban`) From the start of 2028, everywhere, no training run may be bigger than the largest before it. Labs keep finding better methods (with the speed-up), keep serving customers and hold research compute where it was. Other labs and open models stay their lag behind, so the gaps shrink. Where it comes from: Caps like this have been proposed in public: a US bill (S. 5493, September 2026) would pause training runs above 10^25 operations, and MIRI's draft treaty would ban them above 10^24 worldwide. My choices of November 17, 2026: better methods go on at Epoch AI's pace, research compute is held flat. The date and "everywhere" are the agent's choice, not confirmed. Not modelled: runs moved abroad in secret, a lab that breaks the cap, prices falling as the gaps close. Note: The bill and MIRI's draft are named only as examples of what has been proposed. Neither made, checked or endorsed this scenario. Settings it changes: `link.capFrom=2028`. Result: the forecast runs to the end of 2032. Anthropic's revenue 2027: $155B, 2028: $275B, 2029: $297B, 2030: $315B, 2031: $327B, 2032: $341B. Best model's ECI at the end of each of those years: 205, 218, 227, 235, 240, 245. Share price Anthropic's revenue supports: about $13. Money Anthropic raises: $81B. ### Cap on training runs, US only (`scenario=cap_us`) As the cap above, but only in the US. Open models, mostly from Chinese labs, keep the pace the frontier would have had, stay their lag behind it, and soon pass the capped labs. Where it comes from: Closer to the US bill (S. 5493), which binds only US labs. Assumed by the agent, not confirmed: open models keep their lag to an uncapped frontier, though today they partly learn from the US labs' models. Labs abroad that do not publish their models are not modelled. Note: The bill is named only as an example of what has been proposed. Its authors did not make, check or endorse this scenario. Settings it changes: `link.capFrom=2028`, `link.capScope=us`. Result: the forecast stops at the end of March 2030 (the best AI model passes ECI 250); nothing after that is forecast. Anthropic's revenue 2027: $155B, 2028: $274B, 2029: $311B. Best model's ECI at the end of each of those years: 205, 219, 246. Share price Anthropic's revenue supports: about $9, worked from 2029, the last full forecast year. Money Anthropic raises: $81B. ### Moratorium and training cap (`scenario=both_limits`) Both at once from the start of 2028: no new commitments to AI compute anywhere, and no training run bigger than the largest before it. Where it comes from: The two scenarios above, combined. The date and sizes are the agent's choice, not confirmed. Settings it changes: `link.buildFrom=2028`, `link.buildGrowth=0`, `link.capFrom=2028`. Result: the forecast runs to the end of 2032. Anthropic's revenue 2027: $155B, 2028: $275B, 2029: $297B, 2030: $315B, 2031: $327B, 2032: $334B. Best model's ECI at the end of each of those years: 205, 218, 227, 235, 240, 245. Share price Anthropic's revenue supports: about $13. Money Anthropic raises: $81B.