The story in pictures

The steep line is what models can do; the flatter ones are firms and working hours catching up. The methods line follows the best model so far, leaving out readings METR flags as beyond what its tests can measure. Too few readings to draw: Developer productivity uplift (METR RCTs).

Source: METR time horizons (v1.1); Census Business Trends and Outlook Survey; FRED (Federal Reserve Economic Data). METR, Measuring AI Ability to Complete Long Tasks (Horizon v1.1). U.S. Census Bureau, Business Trends and Outlook Survey. Bick, Blandin & Deming, Real-Time Population Survey, retrieved from FRED.

A drawing of the argument made by Arvind Narayanan and Sayash Kapoor, not of any data. The order of the stages and what holds each back are theirs, as this site's records word them. The even spacing, the gentler slope of each later curve and the equal heights are drawn for the eye and measure nothing: the drawing shows that each stage must wait for the last, not how long the wait is. The figures above and below show what is measured.

One mark is one published gauge, and each links to its page. A gauge is what this site calls an indicator: a published series read against a rule written down in advance. A filled mark casts a vote toward its stage's status; gauges that read the same survey or tracker share a single vote, so the rest of them are outlined.

Emerging is this site's word for a gauge that has readings but no verdict yet: its number sits between the ranges or on the line between them, its margin of error reaches across both, it rests on a single source that is not a primary one, or its only evidence is a company describing itself. A mark here does not show how large any reading is; the figure of ranges further down the page does.

Each strip has its own scale, starting at nothing, so compare a mark with its own ranges and not with the row above. The ranges are published in advance, set from how earlier technologies spread and from the long-run trend in output per hour; each variable's own page gives the reasoning. A number sits inside a range only when the site scores it there. A number hanging beneath its strip is one the site has not scored: the essay's sentence above reads where the number falls, and the site's own status does not yet confirm it. The lower rows have no range: the site grades them by whether the number rose or fell over a set window, and a rise can mean the profit gathering in one place or spreading out, never fast or slow.

It does not show how sure each reading is, how each got here over time, or the fast variables the essay sets these against. The chip makers' share is a share of the gross profit this site can measure, which includes an estimate for OpenAI and Anthropic; the value users keep is one survey's estimate.

This figure was drafted by Claude, a model made by Anthropic, and Anthropic is among the labs whose estimated profit sits inside the total the chip makers' share is measured against. Check it against the records it links to.

The strip is the whole $7.0T of knowledge payroll scored. From the census (v3.2-2026-09-24), a screen scored by three AI models, not a claim about what AI can do and not a forecast. All three models pass $158B, which is 2.3% of the whole. The last part is everything else: mostly work with no existing check that is also slow to judge or costly when wrong; work whose most-given reason is a missing check but on which the models split over whether anything else holds it; and a small part that is a sign-off a person answers for.

Every bar is a share of the $7.0T of knowledge payroll scored. From the census (v3.2-2026-09-24), a screen scored by three AI models, not a claim about what AI can do and not a forecast. Each task that fails carries the reason the models that voted no gave most often, set by the census's own code; where reasons tie, the task counts under each, so the middle bars overlap and do not add up. A question a task also fails, but less often, is not counted, so each bar is a floor for that question.

The marker is placed by a published rule, using spending on buildings and equipment at 4.38× the AI revenue we can measure, and $998B of commitments signed in four quarters in deals where the seller also funds the buyer. By that rule this is frenzy. The rule: Installation while Microsoft, Alphabet, Amazon, Meta and Oracle spend at least a dollar on buildings and equipment for each dollar of AI revenue we can measure. Inside it, frenzy while deals in which a seller also funds its buyer are growing on a year before, irruption before any frenzy, and a turning point once a confirmed frenzy cools. Deployment once that spending falls below half a dollar per dollar and such deals are no more than a year before; synergy and maturity inside it wait for a revenue-growth reading the site does not yet hold. A turning point in between. A year-ago reading counts only when the ledger was searched for the whole year behind it, and the half is shown when the finer phase cannot be told apart. A change counts only after two quarters in a row agree, and the phase is untestable when either series has no reading.

Source: SEC EDGAR XBRL company facts; Epoch AI data hub — AI companies; Menlo Ventures, State of Generative AI in the Enterprise; SEC EDGAR filings (8-K, 6-K, 10-Q exhibits); NVIDIA Newsroom; Official Microsoft Blog; About Amazon company news; CNBC; SoftBank Group press releases. SEC EDGAR, XBRL company facts API. Epoch AI, 'Data on AI Companies' (CC BY 4.0), epoch.ai/data. Menlo Ventures, '2025: The State of Generative AI in the Enterprise'. SEC EDGAR archives. NVIDIA Newsroom. Official Microsoft Blog. About Amazon. CNBC. SoftBank Group.

Each column is one quarter's gross profit across the parts we can measure: NVIDIA and AMD at the bottom, Microsoft's cloud business above them, and an estimate for OpenAI and Anthropic, hatched, at the top.

Derived (gross_profit_share_by_layer) from: SEC EDGAR XBRL company facts; Epoch AI data hub — AI companies; SEC EDGAR 10-Q/10-K XBRL instances (segment data). SEC EDGAR, XBRL company facts API. Epoch AI, 'Data on AI Companies' (CC BY 4.0), epoch.ai/data. SEC EDGAR, inline XBRL filings (segment disclosures).

What each bar rests on. Sales a year is a run-rate: a recent month or quarter of sales multiplied up to a year, or the yearly value of current subscriptions, as told to the press or stated by the lab and compiled by Epoch AI. It is not booked revenue, no auditor has checked it, labs differ on whether they count sales before or after a partner's cut, and this site marks every such figure as disputed. Each lab is drawn at the date of its own latest report, and those dates differ. Money raised is the sum of the funding rounds on record up to that date; it leaves out debt and any round this site does not hold, so it is a floor. Payments signed up to is the full written amount of the contracts and commitments on the financing ledger in which the lab is the buyer: they run over many years, amounts described as up to a ceiling are counted in full, some are press reports that the supplier has not confirmed, and none of it is money paid. It is a floor too: deals with no dollar figure are left out, and a lab's own statement of its total commitments is not used. A lab with no such bar has no priced deal on the ledger, which is not the same as having promised nothing.

Labs with only a round or two of funding on record, or whose latest figure is much older than the others, are listed beneath without a multiple. Meta and Google DeepMind have no run-rate on record, and Cohere and Moonshot AI have a run-rate but no funding round, so none of them has a row. This page is drafted by a Claude model, made by Anthropic, which is drawn here. Anthropic's sales figure is a press report, drawn hatched by the same rule as any other lab's press-reported figure, and the model has no figure of its own.

Each band joins a supplier or investor, on the left, to a lab, on the right, and its thickness is the full amount written in the signed deals between them on the financing ledger; the newest deal on it is dated September 2026. That is every signed deal the ledger holds since its first row, so it is larger than the gauge of circular financing, which counts only deals signed in the trailing year. Circular financing means a supplier funding the customer that then spends the money with that supplier. The ledger does not record which way money moves, so a band is a tie, not a flow: where a pair has a band of each colour, the same company has both a sale to the lab and a stake in it or a guarantee behind it.

A few very large deals make up most of the total, so a single cancelled or renegotiated deal would redraw the picture. These are promises, most of them running for many years: amounts described as up to a ceiling are counted in full, and nothing here says how much has been paid. A guarantee is a promise to pay only if the company it backs does not, so it may never be drawn on. Letters of intent and talks are on the ledger but not drawn. Suppliers beyond the largest are gathered into a single band, and deals with no frontier lab in them, such as a data-centre owner leasing to a cloud company, run along the bottom; labs outside the frontier group, such as Mistral, fall in the bottom band. Meta is on the right because this site files it with the frontier labs, though these deals serve its whole business; Alphabet, which owns a rival lab, is on the left because here it is the investor. The ledger holds what this site has found and read, so it is a floor.

The layers run in the order work flows, from power and chips to the people who use the products. The bars count each company on the market map below once, by the part of the chain the tracker files it under. The map sorts the same companies into its own finer layers and categories and lets a company sit in more than one, so its counts are larger and do not line up with these bars. The folded table lists the parts behind each bar. A bar is the tracker's reach, not the size of any market: a tall bar means many companies followed. The last layers hold none because the site reads them through surveys and statistics, not through companies.

Each direction is a judgement, not a measurement, drafted by Claude, a model made by Anthropic, whose own part of the chain (frontier labs) is among those judged: follow it to read the argument, the view against it and what would prove it wrong. 169 of the 849 companies sit in parts with no judgement yet, and 3 are on the map without a part of the chain. The parts with no judgement are ones the site has not yet written up, not ones it judged unimportant.

Short names are the tracker's own; each links to a plain statement of the judgement. RL is reinforcement learning, training by practice and grading; HBM is the fast memory stacked beside an AI chip; FDE is an engineer the seller places inside a customer; evals are tests of a model.

Each solid mark is an input's tightness score from the scorecard below, tightest first. A hatched mark is a score this site holds with low confidence because its gauge is a stand-in for the thing itself; it is drawn apart and is not ranked with the rest. A hollow mark is an input with no score: it has no place on the scale, so it is drawn beside it, never in the middle. Inputs that sit close together are not in a meaningful order. Each mark links to its row in the scorecard, where the gauges and their sources are.

This site's picture of where the kinds of firm on this page are heading. It is a model of an idea, joined from the arguments on the page before; no reading tests it whole.

One mark is one dispute from the first figure: a pair of rival positions, or a position that argues against one drawn elsewhere; it links to the position. Its fill is how much this site could read on it tonight, taking the claim that got furthest on either side, in this order: a reading that met a claim's line, then a reading that missed one, then a reading both sides expect, then a test written and waiting, then no test at all. In 41 of the 59 disputes no claim on either side has a test, so nothing this site reads bears on them yet. A dark mark does not mean either side has the better of the argument. A single claim that met its line is enough to fill it, and in only 2 of the 10 dark disputes, the ringed marks, does that claim carry a test of what the rival expects; in the rest this site has not written down what the rival would expect of the same reading. An orange mark is a dispute where a claim missed its line and none met one, which does not hand the argument to the other side either. Across the page, 9 of the 11 claims that are holding have no test for the rival. These figures were drafted by Claude, an AI model made by Anthropic. Several of these positions are about that lab's business, and among the writers drawn here are advisers to it, a writer at its institute and the founder of a firm that holds a stake in it; each writer's ties are given with their source at the foot of the page. Every count and mark comes from the ledger and its nightly tests, not from the model.

Each bar is every forecast in that section, split by tonight's word; the count beside a name is how many forecasts the bar holds. A bar shows shares, so a short section and a long one look alike: read the counts in the table. The words are this site's reading on 2026-11-04, not a final verdict.

Every bar here is hatched because it draws a judgement, not a reading. The judge is claude-opus-5-5, a model made by Anthropic; some of these forecasts are about Anthropic and its rivals, and no person reviewed each lean, so treat the leans on those forecasts with extra care. It gave each lean on 2026-10-05, and a lean changes no word above. A bar is the shares of that section's leaned-on forecasts; the numbers at its ends count the forecasts leaning true and leaning false, and the count beside a name is how many the bar holds. Bars are shares, so a section with few leans can draw the longest bar: read the counts. Forecasts with no lean are in the table, not drawn.

The scale is not even: the half century from 1950 and the half century from 2050 are squeezed so the years around now get most of the width (dashed lines mark where it changes). A call that gives odds, or bets against a date, without naming a year is listed under the chart. A forecast is placed at the start of its year, so one due by the end of this year sits left of today while its window is still open. Shaded: the years still to come.

One mark is one dated forecast of the milestones in the forecaster table: 37 in all, each linked to its record. The dashed steps are the year a forecast was made, so a mark higher above them looked further ahead. Because the scale up the side is not even, equal heights are not equal numbers of years: the table gives each forecast's years. A mark under the line marked this year names a year that has already passed; a mark on it names this year, which is not over. Neither scale is even: the years before 2020 are squeezed along the bottom, and the years before 2020 and after 2040 up the side (dashed lines mark where each scale changes).

Where one mark would cover another, the later is set beside it, to the right or the left, and tied back by a short line to the date it was made: 7 marks are set aside this way, some by several months, and no mark is set right of today. Read such a mark's date from the far end of its line, or from the table. No mark is named in the drawing: the table lists every forecast in the order drawn, left to right, with its writer. Not drawn: 3 steps toward a milestone and 14 forecasts that give odds or doubt a date without naming a year.

The marks fall toward the lower right because this site's list holds different people making different claims in different years, not because anyone here changed a date: no forecaster's earlier and later date for the same claim is on record, so the figure cannot show a forecast being brought forward and is not evidence that forecasts are converging. Where one writer has more than one mark on a milestone they are different claims, at different odds or on different conditions; each links to its record. A mark is the year the timeline places a forecast at, which may be a deadline, a most likely year or a year given only some odds, and a range bar may be a spread of odds or a plain "give or take". No trend line is fitted. This figure was drafted by a Claude model, made by Anthropic, and the heads of Anthropic and of rival labs are among the forecasters it draws; each is placed by the same rule, from its own record in this site's ledger of forecasts, and no placing was adjusted for anyone.

The grid is the outlook page's: how far capability goes, down, and how the rules settle, across. Each world is drawn over the cells this site gives it; the placing is this site's, a drawing of an idea, and no world is crowned. A glyph is one signpost: a claim the outlook tests each night, listed in the table.

A world reads consistent until a signpost in one of its cells fails. On 2026-11-04, of the signposts on the grid, counting one that sits in more than one cell each time, 8 are expected by both sides of the argument, 3 cannot be tested yet, 0 hold and 0 fail, and 2 cells have no signpost at all. So consistent here means not ruled out: it is no evidence that a world is on its way.