“The share of value-production running through the computational substrate. It is small today but rising, and it climbs in a fixed order — media first, born legible; cognitive services now, as language falls; the physical domains next.”
Denominator
Task-hours, wage-weighted — not value. A value share erases itself as prices fall to the cost of compute.
The authorship fraction — the share of end-setting on the substrate, three conditions per domain
Built
Latest
One economy, six readings
Every published “AI adoption” number sits on a different rung. The fraction the book names is the bottom rung — the share of work that actually runs through the substrate — and it is an order of magnitude below the headlines. United States, latest available.
Firms using AI in any business function
Census BTOS Q7, national, biweekly. Question rewritten Nov 2025 (“in producing goods or services” → “in any business function”); old series ran 3.7% Sep 2023 → ~10% Sep 2025. census.gov
Employed adults using genAI for work
Bick, Blandin & Deming, Real-Time Population Survey, quarterly, via FRED. Intensity in the same survey: 6.3% of work hours assisted, 2.2% saved (Q2 2026, FRED Blog, 27 Aug 2026).
Does it climb in the book’s order?
Media first, cognitive services now, the physical domains next. Each row is tagged with the tier the book would put it in; if the claim holds, the tags should stratify when sorted by any adoption measure.
media — born legiblecognitive servicesphysical domainsembodied / presence (the remainder)
The order holds at the top and fails at the bottom. Information, professional services and finance lead on every instrument. But the book’s “physical next” tier is not a tier: manufacturing and construction have moved faster than arts and hospitality in the American data, and the true floor is not the physical economy but the embodied one — care, food, accommodation — which is what Chapter Two calls the remainder. The ladder has three rungs in the text and four in the data.
Media first — but which fraction?
The book says music went first because it was born legible. The data agree, and then show why the choice of denominator decides everything.
Fully AI-generated tracks, share of daily uploads to Deezer
Deezer newsroom, 21 Jul 2026 (Jun 2026 >50%, ~90,000 tracks/day); intermediate points from Deezer’s Apr 2026 and 2025 releases. Shaded band: AI tracks’ share of streams, 1–3%, up to 85% of it fraudulent.
Three fractions for one sector
>50%of new units produced (uploads) — the substrate has taken production1–3%of consumption (streams) — listeners have not moved≈0%of value — an AI track earns nothing, so a value-share tracker reports media as barely touched
Counted by what is made, music has crossed. Counted by what is heard, it has not. Counted by what is paid, subsumption is invisible at the moment it is most complete. This is Ben’s deflation point (B76) made concrete, and it is why this tracker refuses a value denominator.
Other media and creative readings
By country
No single national subsumption number exists; five instruments triangulate it. Diffusion is who uses the substrate; enterprise adoption is whether firms run work through it; the Anthropic index is intensity relative to population; preparedness is capacity to absorb more; the data-centre share is the joule side of the meter.
Preparedness vs diffusion
x: IMF AI Preparedness Index (2023). y: Microsoft AI Diffusion Rate, share of 15–64 population using a genAI product, Q1 2026. Hover a point for the country.
The joule side: data centres as share of national electricity
IEA Energy & AI via Our World in Data; CSO Ireland; CBS Netherlands; LBNL 2024. Singapore is a secondary 2020 figure — indicative only.
Diffusion: Microsoft AI Economy Institute Q1 2026 (◌ = regionally imputed). Enterprise: Eurostat 2025 (10+ employees) where available, otherwise national statistics office (UK ONS Jun 2026, StatCan Q2 2026, ABS 2024–25, Statistics Korea 2024, US Census Aug 2026, Japan MIC 2025 — Japan’s sample skews large). AUI: Anthropic AI Usage Index, Nov 2025 (usage share ÷ working-age population share; automation = share of conversations that are directive or feedback-loop). AIPI: IMF 2023. Exposure: IMF SDN/2024/001, only six countries quoted.
The crossover, task by task
“Each task has a price at which the substrate beats the wage … that price is the only brake the system currently has.” Here is the price, the wage, and the brake the text does not name.
Human cost of an expert task
$3616.7 h
GDPval median task: 404 minutes of an expert with ~14 years’ experience, at BLS May 2024 median wages. OpenAI, Sep 2025.
Substrate cost, naïve
≈$3.6~100× cheaper
OpenAI’s own multiple: frontier models “roughly 100× faster and 100× cheaper” on pure inference time and API billing — excluding oversight.
Substrate cost, with review
1.63×cheaper, not 100×
Same paper, “try, then fix”: once an expert reviews the output (109 min, $86) and redoes failures, GPT-5 is 1.63× cheaper and 1.39× faster. Expert parity or better: 47.6% of tasks (Opus 4.1).
The inference price has crossed the wage for every legible task in the table below, by one to two orders of magnitude. What has not crossed is the cost of verification. The brake the system currently has is not the price of the substrate; it is the price of checking its work, and that price is set by a human wage. Chapter Two names the wrong brake, and Chapter Nine already knows the right one: where checking is far cheaper than producing, the substrate wins outright.
Wage vs substrate, by occupation (US, $/hour)
Bars: BLS OEWS May 2025 median hourly wage. Vertical line: substrate cost per human-hour of output at current frontier list prices, ≈$0.54 (GDPval $3.6 ÷ 6.7 h) — the bar it would have to reach. Colour is the book’s tier. Every bar lies to the right of the line; the ratio, not the crossing, is the reading.
Where the price is going
9–900×/yrfall in price to reach a fixed capability, by task (Epoch AI, Mar 2025)>280×fall in GPT-3.5-level inference cost, Nov 2022–Oct 2024 (Stanford AI Index 2025)5–10×/yron the price–performance frontier, Apr 2024–Nov 2025 (Gundlach et al.)129 daysdoubling time of the 50% task horizon since 2023; Opus 4.6 ≈12 h, Mythos Preview ≈17 h (METR, May 2026; wide intervals, and METR treats readings above 16 h as unreliable)−19%experienced developers were slower with AI tools in a randomised trial, while believing they were 20% faster (METR, Jul 2025) — verification cost, measured
Frontier list prices, $ per million tokens
OpenRouter pass-through, 7 Sep 2026 — spot-check against provider pages before quoting. Anchor: GPT-3.5-level ≈ $20/M in Nov 2022.
“For the present below it” — is the price a subsidy?
The authorship fraction
The subsumption fraction counts how much of the work has moved onto the substrate. This instrument counts the narrower quantity that sits above it: how much of the end-setting has. It is read per domain, never as one number, and only when three conditions come true together. The laboratories’ own behavioural reporting appears here as evidence of capacity — the precondition of the first condition — and not as evidence of the claim.
Where each domain sits
A domain sits inside a circle when that condition holds (crossed) and on its rim when it is crossing; the centre is the threshold, where all three hold together. Each glyph is a three-sector dial in the same orientation as the circles: top = no standing instruction, lower left = acted on unreviewed, lower right = withdrawal uncompetitive. Hover for the readings; click to jump to the row.
Three conditions, domain by domain
crossedcrossing — partial or contestedopen — not met on available evidence
Status is a judgement on the best available reading, not a computed value; the reading and its source sit in each cell. ◌ marks a figure reached only through secondary coverage. Trading and advertising are baselines: they crossed all three conditions before the chapter’s story begins.
Capacity — what systems can now do unattended
METR 50%-success time horizon, frontier model at each reading (hours)
Threshold status under the laboratories’ own frameworks
Individual-model dispositions — what the laboratories choose to measure
Delegation — the review seat
Is anyone still reading it before it takes effect?
What is handed over
The review seat, in four readings
Grey: the human seat. Blue: the machine seat, or the machine in use. Black: the earlier reading in the same series. Each pair is one source and one denominator; sources are in the list above.
Lock-in — where withdrawal is the uncompetitive option
What was fed in, day by day
Every item the weekly scan turned up, on the day it was run: what the source said, whether it was keyed into the data, and which tab it touched. Live series appear when a new cycle or wave arrives; everything else is hand-keyed or noted and left out.
keyed — entered into the datalive — an automated series refreshednoted — read, not keyed (no primary figure, or out of scope)status — a judgement changed on the authorship grid
◌ marks an item reached only through secondary coverage. The tab name after each item links to where it landed. Runs are recorded once in the update log on the Movements tab; this tab is the item-level record.
Quarter on quarter
The funnel, then and now
Quarterly series
One row per snapshot. US firms (BTOS), US workers (RPS), work hours assisted and saved (RPS), paid subscriptions (Ramp).
Biggest industry movers — US firms (BTOS)
Biggest industry movers — US workers (RPS)
Country movers — population diffusion (Microsoft)
Country movers — Anthropic AUI
Country movers — EU enterprise adoption
Data-centre share movers
Update log
Method & sources
What is being measured
Definition (book). The share of value-production running through the computational substrate, climbing media → cognitive → physical, counted task by task at the price where the substrate beats the wage.
Denominator (instrument). Task-hours weighted by the pre-subsumption wage bill, not value. As a good’s price falls toward the cost of compute its weight in measured output shrinks, so a value share understates subsumption exactly where it is most advanced. Value share is kept as a secondary reading.
Three layers, never collapsed into one index.Exposure: what could cross (IMF, ILO, Anthropic task coverage). Adoption: what has crossed, observed (Census BTOS, RPS, Eurostat, Ramp, Deezer, ELIS). Crossover: substrate price per task against the wage, with the review cost and the subsidy flag alongside.
The funnel is the headline. Any single rung quoted alone misleads: firm adoption overstates, hours saved understates the direction of travel. Publish the rungs together.
Tiers are a hypothesis under test. Every industry row carries the book’s tier tag so the “fixed order” claim can be falsified by sorting.
Known failure modes
Survey instruments changed wording (BTOS, Nov 2025). Trend lines are only drawn within a consistent series.
Adoption is not intensity. RPS is the only source measuring hours; it is a survey of 5,000 and its industry cells are noisy.
The Anthropic index measures one vendor’s users; China is absent by construction.
Microsoft’s diffusion rate is telemetry-modelled and regionally imputed for much of Africa and Central Asia.
IMF per-country exposure exists for ~125 countries as a chart only; the data file is available on request from IMF Research and should be obtained.
Inference list prices are not costs. Gross margins at the large labs turned positive in 2025; the subsidy now sits in training, capex and the rate of price decline, not in the marginal token.
Refresh cadence
BTOS biweekly (live), RPS quarterly (live via FRED), Eurostat annual each December (live), IMF AIPI on release (live). Ramp monthly, Anthropic Economic Index roughly quarterly, Microsoft Diffusion semi-annual, Deezer ad hoc, ILO/IMF exposure on publication — hand-keyed into curated.json. Pipeline: pipeline.py → build.py, no API keys.