Arvind Narayanan and Sayash Kapoor, two Princeton computer scientists, argue that AI will spread the way electricity did: over decades, through stages. This page times each stage separately, from new methods to reorganised work. A stage carries a status only when a published indicator, which this site calls a gauge, measures it; otherwise it reads unmeasured.
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.
| Stage | Status | Fast | Normal | Slow | Emerging | Reads who keeps the money | Votes cast |
|---|---|---|---|---|---|---|---|
| Methods | mixed | 2 | 2 | 0 | 7 | 2 | 2 |
| Products | consistent with normal | 0 | 3 | 0 | 2 | 1 | 3 |
| Early adoption | consistent with normal | 0 | 3 | 0 | 6 | 1 | 3 |
| Adaptation | consistent with normal | 0 | 14 | 0 | 7 | 1 | 11 |
| Return arrow | emerging | 0 | 0 | 0 | 7 | 1 | 0 |
Each box is a stage the technology has to pass through, each arrow the flow from one stage to the next. On the right, two things that push back: public backlash and safety brakes, and tests that keep moving the goalposts. Down the left, what use feeds back into better methods.
The statuses rest on readings from 42 sources; each stage's page names them, indicator by indicator, and the sources page lists them all.
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.
Why it matters: a technology changes the economy only as fast as its slowest stage, so a fast first stage says little on its own.
What limits its speed: None inherent; capability can move fast. The paper's claim is that this stock is not the bottleneck.
What limits its speed: Reliability, integration, evaluation and liability — the 50%→80% gap and the "last mile" of productisation.
What limits its speed: Human and organisational speed — training, procurement, trust, ROI measurement (electrification took decades).
What limits its speed: Complementary innovation, regulation, safety brakes and the pace of institutional change.
What limits its speed: Whether the deployment→training channel is open and how far it substitutes for external bottlenecks.
A status is a word. Behind it is a number and a line it has crossed, has not crossed, or is held at, so these figures show the distance as well as the verdict.
Each gauge has a normal range and a fast range, fixed before the data came in: the first is what a slow, electricity-like spread would show, the second what a much quicker one would. A mark's place is the range tonight's number falls in, before the rule's checks on its evidence; its shape is the status the site has published. Between the ranges the place is to scale. Inside a range, which has no far edge, the scale is squeezed: a mark deeper in its range is further from the line, counted in widths of that gauge's own gap between the ranges, but equal steps on the page are not equal steps in the number. Height within a strip means nothing; marks are stacked so they do not overlap.
A part-filled mark in a shaded range is a gauge whose number falls in that range but which the site has not scored. It is not a reading of fast or of normal. The rule holds a number back when it sits on or very near the line between ranges, when its margin of error reaches across both, when it rests on a single source that is not a primary one, or when its only evidence is a company describing itself. The table beneath says which applies to each gauge. Gauges read by direction instead of by a range are left out and listed in the table.
| Stage | Gauge | Number the rule reads | As of | Where it sits | Published status | Why it is not scored |
|---|---|---|---|---|---|---|
| Methods | METR 50% time horizon | 108 days | 2026-03-05 | inside the fast range | faster than normal | — |
| Methods | METR 80% time horizon | 107 days | 2026-04-07 | inside the fast range | faster than normal | — |
| Methods | 50%/80% horizon ratio | 6.34× | 2026-03-05 | inside the normal range | consistent with normal | — |
| Methods | Frontier training compute growth (doubling time) | 170 days | 2026-09-03 | inside the fast range | emerging | its margin of error spans more than one range; every reading comes from one source that is not an official one, and the site scores a status only on two independent sources or on one official one |
| Methods | Training power draw growth (doubling time) | 419 days | 2026-10-15 | inside the normal range | emerging | its margin of error spans more than one range; every reading comes from one source that is not an official one, and the site scores a status only on two independent sources or on one official one |
| Methods | Hardware price-performance growth (doubling time) | 1331 days | 2026-08-02 | inside the normal range | consistent with normal | — |
| Methods | Inference price at fixed capability (halving time) | 121 days | 2026-04-03 | between the two | emerging | its number falls between the ranges |
| Methods | Open-weights lag behind the closed frontier (months) | 7.7 months | 2026-10-26 | between the two | emerging | its number falls between the ranges |
| Methods | Public text data exhaustion year (Epoch projection) | 2028 | 2024-06-06 | on the line of the normal range | emerging | it sits on the line between ranges; every reading comes from one source that is not an official one, and the site scores a status only on two independent sources or on one official one |
| Methods | ARC-AGI-2 best score | 95.0% | 2026-09-03 | inside the fast range | emerging | every reading comes from one source that is not an official one, and the site scores a status only on two independent sources or on one official one |
| Methods | Progress against the AI 2027 quantitative predictions | 70.0% | 2026-07-31 | between the two | emerging | its number falls between the ranges |
| Products | Developer productivity uplift (METR RCTs) | -4.0% | 2026-02-24 | inside the normal range | consistent with normal | — |
| Products | Enterprise pilots with measurable P&L impact | 5.0% | 2025-06-30 | inside the normal range | consistent with normal | — |
| Products | Frontier models completing expert legal tasks end to end (Harvey LAB) | 7.1% | 2026-05-26 | inside the normal range | emerging | its best evidence is a company describing itself; every reading comes from one source that is not an official one, and the site scores a status only on two independent sources or on one official one |
| Products | Agent reliability across repeated runs (tau-bench pass^k) | 56.2% | 2026-05-18 | inside the normal range | emerging | its best evidence is a company describing itself |
| Products | Internal-to-public deployment gap at the frontier | 4 months | 2026-06-30 | inside the normal range | consistent with normal | — |
| Early adoption | Share of work hours assisted by generative AI | 5.3% | 2026-09-30 | inside the normal range | emerging | every reading comes from one source that is not an official one, and the site scores a status only on two independent sources or on one official one |
| Early adoption | US firms using AI (Census BTOS) | 24.6% | 2026-10-04 | inside the normal range | consistent with normal | — |
| Early adoption | US businesses paying for AI (Ramp AI Index) | 55.9% | 2026-09-30 | between the two | emerging | its number falls between the ranges |
| Early adoption | Augmentation vs automation share (Anthropic Economic Index) | 51.4% | 2026-05-31 | on the line of the normal range | emerging | it sits on the line between ranges |
| Early adoption | Work-related share of ChatGPT consumer messages | 27.0% | 2025-06-30 | inside the normal range | emerging | its best evidence is a company describing itself |
| Early adoption | Usage as a share of theoretical exposure | 0.44× | 2026-09-30 | inside the normal range | consistent with normal | — |
| Early adoption | US adults who have used generative AI | 60.8% | 2026-09-30 | inside the normal range | emerging | every reading comes from one source that is not an official one, and the site scores a status only on two independent sources or on one official one |
| Early adoption | Occupations with a quarter of their tasks using Claude | 49.0% | 2026-02-12 | inside the normal range | consistent with normal | — |
| Early adoption | US workers using generative AI for work every day | 12.0% | 2025-11-30 | inside the normal range | emerging | every reading comes from one source that is not an official one, and the site scores a status only on two independent sources or on one official one |
| Adaptation | US labour productivity, year on year | 2.2% | 2026-06-30 | inside the normal range | consistent with normal | — |
| Adaptation | US total factor productivity, private nonfarm business | 0.8% | 2025-12-31 | inside the normal range | consistent with normal | — |
| Adaptation | Safety brakes inside the labs (events, trailing year) | 3 | 2026-11-04 | inside the normal range | emerging | its best evidence is a company describing itself; every reading comes from one source that is not an official one, and the site scores a status only on two independent sources or on one official one |
| Adaptation | FDA-authorised AI-enabled devices, growth | 26.7% | 2026-06-30 | inside the normal range | consistent with normal | — |
| Adaptation | US state AI bills introduced (legislative year) | 1035 | 2025-12-31 | inside the normal range | emerging | every reading comes from one source that is not an official one, and the site scores a status only on two independent sources or on one official one |
| Adaptation | FDA-authorised devices built on a large language model | 1 | 2025-12-23 | inside the normal range | consistent with normal | — |
| Adaptation | Employment in AI-exposed occupations, year on year | -0.5% | 2026-09-30 | inside the normal range | consistent with normal | — |
| Adaptation | Entry-level employment shortfall in AI-exposed occupations | 19.0% | 2026-06-30 | between the two | emerging | its number falls between the ranges |
| Adaptation | Exposed vs unexposed occupation employment gap (Revelio) | -6.0% | 2026-07-31 | between the two | emerging | its number falls between the ranges |
| Adaptation | California unemployment claims from AI-exposed occupations | -4.4% | 2026-09-30 | inside the normal range | consistent with normal | — |
| Adaptation | Recent-graduate unemployment rate (NY Fed) | 5.6% | 2026-06-30 | inside the normal range | consistent with normal | — |
| Adaptation | Cross-tracker concordance (labour) | 0 | 2026-09-30 | inside the normal range | consistent with normal | — |
| Adaptation | US labour share of income, year on year | -3.4% | 2026-06-30 | between the two | emerging | its number falls between the ranges |
| Adaptation | Customer-support productivity uplift from a generative-AI assistant | 14.0% | 2023-04-30 | inside the normal range | consistent with normal | — |
| Adaptation | Executives reporting no AI impact on their own firm | 90.0% | 2026-02-28 | inside the normal range | consistent with normal | — |
| Adaptation | Workers highly exposed to AI with low adaptive capacity | 6,100,000 workers | 2026-01-31 | between the two | emerging | its number falls between the ranges |
| Adaptation | Occupational-mix dissimilarity since ChatGPT (Yale Budget Lab) | 5.56 pp | 2026-09-01 | inside the normal range | consistent with normal | — |
| Adaptation | Chatbot effect on earnings and hours in exposed occupations (Denmark) | 0.0% | 2024-12-31 | inside the normal range | consistent with normal | — |
| Adaptation | Freelance writing earnings after ChatGPT (Upwork) | -5.2% | 2023-12-01 | inside the normal range | consistent with normal | — |
| Adaptation | Wage premium for AI skills (PwC AI Jobs Barometer) | 62.0% | 2025-12-31 | inside the normal range | emerging | its best evidence is a company describing itself |
| Adaptation | Waymo paid rides per week | 500,000 rides per week | 2026-03-27 | inside the normal range | consistent with normal | — |
| Return arrow | Agent-workdays per human workday in frontier research (self-reported) | 3.10× | 2026-08-15 | on the line of the fast range | emerging | it sits on the line between ranges; its best evidence is a company describing itself; every reading comes from one source that is not an official one, and the site scores a status only on two independent sources or on one official one |
| Return arrow | Human interventions on 4–8 hour agent tasks (self-reported) | 50.0% | 2026-07-31 | on the line of the normal range | emerging | it sits on the line between ranges; its best evidence is a company describing itself; every reading comes from one source that is not an official one, and the site scores a status only on two independent sources or on one official one |
| Return arrow | Continual-learning ladder (highest production rung) | 3.00 rung | 2026-05-27 | on the line of the normal range | emerging | it sits on the line between ranges; its best evidence is a company describing itself |
| Return arrow | Active RL-environment vendors (directory count) | 23 | 2026-10-05 | inside the normal range | emerging | every reading comes from one source that is not an official one, and the site scores a status only on two independent sources or on one official one |
| Return arrow | Expert-data market run-rate (Mercor) | $2.0B | 2026-06-30 | on the line of the fast range | emerging | it sits on the line between ranges |
| Return arrow | Vendors claiming shipped continual learning | 2 | 2026-11-04 | inside the normal range | emerging | its best evidence is a company describing itself |
| Return arrow | Fastest production refresh of a continually updated model (hours) | 1.00 hours | 2026-05-27 | on the line of the normal range | emerging | it sits on the line between ranges; its best evidence is a company describing itself; every reading comes from one source that is not an official one, and the site scores a status only on two independent sources or on one official one |
| Methods | Frontier-lab revenue run-rates | — | — | not drawn: read by direction, not by a range | — | — |
| Methods | Open-weights gap to the closed frontier (ECI points) | — | — | not drawn: read by direction, not by a range | — | — |
| Products | Application layer's share of enterprise generative-AI spend (Menlo) | — | — | not drawn: read by direction, not by a range | — | — |
| Early adoption | Enterprise multi-model usage share (Menlo) | — | — | not drawn: read by direction, not by a range | — | — |
| Adaptation | Consumer surplus from generative AI (willingness to accept) | — | — | not drawn: read by direction, not by a range | — | — |
| Return arrow | Lab recoupment ratio | — | — | not drawn: read by direction, not by a range | — | — |
For each stage, a single gauge, named above its panel: the one this site is most sure of among those with more than a single reading to draw, chosen from the gauges that count toward the stage's status where any of those can be drawn. A tie goes to the gauge with more records behind it. Where no gauge counts, the panel says so and the gauge drawn is one the site has not scored. Each panel has its own scale and its own span of dates, so compare shapes within a panel and not heights across panels.
The shaded ranges appear only where the rule reads the very series plotted. Where it reads a fitted trend, or a single named study among several, the panel shows the readings and no range, and the gauge's own page shows what the rule reads. A hollow dot is a reading that is disputed; its page says by whom. One gauge does not speak for its stage: each panel's heading links to the stage's full list.
| Stage | Gauge | Latest reading | As of | Status | Readings drawn |
|---|---|---|---|---|---|
| Methods | METR 50% time horizon | 17.4 h | 2026-04-07 | faster than normal | 26 |
| Products | Developer productivity uplift (METR RCTs) | -4.0% | 2026-02-24 | consistent with normal | 2 |
| Early adoption | US firms using AI (Census BTOS) | 24.6% | 2026-10-04 | consistent with normal | 25 |
| Adaptation | Employment in AI-exposed occupations, year on year | -0.5% | 2026-09-30 | consistent with normal | 50 |
| Return arrow | Continual-learning ladder (highest production rung) | 3.00 rung | 2026-05-27 | emerging | 2 |
A stage can read normal on fresh, well-evidenced gauges or on old and thin ones. These figures show which.
Each mark is a gauge, placed by the age of its newest reading on 2026-11-04: fresh on the left, old on the right. The scale stretches recent weeks and squeezes old years, so that both can be read. A stage whose marks sit to the right is being read on old data, however its status reads.
A gauge is flagged stale, and framed here, once its newest reading is older than two of the gaps its source usually leaves between readings, or one gap and a month where that is longer. A yearly source can therefore be well over a year old, still count toward its stage, and carry no frame: its place on the scale is the only sign. A dashed frame marks a gauge that is past that limit but excused, because it is a one-off study, a count of rare events or a projection that changes only when its source revises it; the reason is in its tip and on its page. Age is counted from the date a reading describes, not the date it was fetched. Height within a strip means nothing; marks are stacked so they do not overlap.
| Stage | Gauges dated | Newest | Middle | Oldest | Flagged stale | Past its refresh, excused |
|---|---|---|---|---|---|---|
| Methods | 13 | 9 days | 94 days | 881 days | 1 | 1 |
| Products | 6 | 127 days | 212 days | 492 days | 0 | 0 |
| Early adoption | 10 | 31 days | 96 days | 1090 days | 0 | 1 |
| Adaptation | 22 | 0 days | 173 days | 1284 days | 2 | 3 |
| Return arrow | 8 | 0 days | 89 days | 161 days | 0 | 0 |
Each gauge carries a confidence score set with its status, on the rubric in the key: it grades the evidence behind the reading, not the size of the reading. Marks to the left rest on thin or vague evidence; marks to the right on good evidence. The score is this site's own grading and is written into the record with a reason each time a status changes.
A stage whose filled marks sit to the right has a status that rests on firm readings; a stage with only part-filled marks to the left has no firm reading yet. The scale stops short of certainty by rule. A mark on a line between bands belongs to the band on its right. Height within a strip means nothing; marks are stacked so they do not overlap.
(ratio ≤ 2 AND 50% horizon below the ceiling AND 80% horizon > 8 h) AND (≥ 3 trackers break) AND (hours > 20% OR TFP > trend + 1pp for 4 years)
ratio non-decreasing (within 10%) AND ladder ≤ L4 AND four releases with at most one break each
both conditions, on independently verified series (OpenAI's self-reported 3.1 and >50% do not qualify)
supported when labs are up ≥ 5pp and chips down, on gross profit; contradicted when chips' share is flat or rising (the brief's app-margin branch needs margins net of inference, which are not public)
surplus above twice the approximate revenue ceiling holds, below the enterprise-spend floor fails, in between is untestable; the doubling covers the willingness-to-accept wedge and the consumer subscriptions the ceiling omits
real pay below its end-2022 level in four consecutive quarters AND real output up on a year before
Vertical AI's share of application funding fell sharply at the end of 2025 and has risen in each of the two quarters since, so the last four quarterly moves split evenly between down and up. The direction rule needs most moves to agree, so the reading is unclear until the recovery holds or reverses.
Weighed against it: Measured end to end, the trailing ratio is still well below its mid-2025 level, which alone would read as dispersing; one large round in either group can move it by more than the dead band.
The ledger was swept under a written rule on 28 Sep 2026 (sellers of chips, cloud, data-centre capacity and power, and their AI customers, from 1 Jan 2024), adding 41 deals and correcting Amazon's Nov 2024 row to its new $4B. Over the last three quarter ends the trailing-year flow rises from $781B (Dec 2025) to $979B (Mar 2026) to $1,179B (Jun 2026), every step well beyond the $20B dead band, so the rule reads concentrating. The earlier unclear reading compared an unsearched year with a searched one.
Weighed against it: Face values count 'up to' amounts and ten-year floors; a handful of very large deals (Oracle-OpenAI $300B, Microsoft-OpenAI $250B, Amazon-OpenAI $150B across three rows) set most of the level, so one cancelled deal would move it sharply. Commitments are not cash, and the sweep may still miss deals whose dollar values were never confirmed by a party.
AMD's fiscal fourth quarters are now filled from its 10-Ks (they file full-year segment figures under a differently spelled member), so Q4 2025 exists. The four-quarter window now runs from 0.865 (Q2 2025) to 0.870 (Q2 2026), a move of 0.005 inside the 0.02 dead band. The concentrating reading set on 27 Sep compared across the missing quarter and no longer holds.
Weighed against it: Over two years the index has risen from 0.82 (Q4 2024) to 0.87, so a longer window would read concentrating; the rule's four quarters say stable. Two filers only; Broadcom, TSMC and the clouds' own chips are not in it.