What is holding AI up, and where

Gallium, germanium, rare earths, copper and the other raw materials chips and power equipment are made from.

Why it is not scored: This site reads no series on how much of these minerals is mined or sold.

• Capability / invention: Chips, power plants and the grid are built from them; a shortage slows how fast computing capacity can grow.

The power stations, turbines, reactors and solar farms the electricity comes from.

What its score reads: Demand only: how fast the electricity that all AI chips sold so far would draw is growing. It says nothing about how much power can be supplied.

• Capability / invention: Training a frontier model draws power for months; without the electricity, the computers cannot run.
• Innovation: products and applications: Serving a product to everyone who uses it draws power every day, on top of training.

Read by: Growth in the power drawn by AI chips sold

Transformers, switchgear, transmission lines and the queue to be connected. The wire between a power station and a rack of chips.

Why it is not scored: The wait to connect a large new load, such as a data centre, is what scarcity here would show up in. Only one utility has stated its wait, and only through local press, so there is no series to read.

• Capability / invention: A new site waits in line to be connected before it can draw power from the grid.
• Innovation: products and applications: The same wait holds back the sites built to serve products.

Expected: This site: “Data-centre buildings and their power hookups are the bottleneck.” holding so far. Byrne Hobart: “Data centres arrive faster and with less warning than the grid's planners assume, so getting connected holds back new American sites.” can't be tested yet.

The shells, cooling and power hookups that racks of chips sit in.

What its score reads: The power Epoch expects at the large sites it tracks, against the power it counts as built there today.

• Capability / invention: The buildings, cooling and hookups that house the chips take a long time to put up.
• Innovation: products and applications: Serving capacity needs the same buildings.

Read by: Planned power against built power

Expected: This site: “Data-centre buildings and their power hookups are the bottleneck.” holding so far. David Cahn: “The cloud companies start storing new chips instead of installing them, the sign Cahn watches for that data centres are running late.” can't be tested yet. Byrne Hobart: “Too few electricians, welders and other tradespeople hold back the building of data centres, so the largest technology firms keep paying to train them.” can't be tested yet.

The machines that print, deposit, etch and inspect chips, made by a handful of firms such as ASML.

Why it is not scored: This site reads the equipment makers' filings, but their backlogs do not measure demand from AI: one of the largest files no backlog, and the others' backlogs fell while the AI build-out grew, following the wider chip cycle.

• Capability / invention: A handful of firms make the machines that print advanced chips, and each new factory needs them.

The factories, above all TSMC's, that make the most advanced logic chips.

What its score reads: How fast the supply of leading-edge logic wafers is growing, year on year.

• Capability / invention: Almost all advanced AI chips come out of a few factories, so their capacity caps how many chips exist.
• Innovation: products and applications: Chips for serving models come out of the same factories.

Read by: Growth in leading-edge logic supply; Growth TSMC states for its newest node

The step that joins processor dies to stacks of fast memory on one base. Nearly all of it is TSMC's process, called CoWoS.

Why it is not scored: None of its gauges has a usable reading today: each is missing or older than its age limit.

What its score reads: How fast packaging supply is growing, and how much of it a few buyers take.

• Capability / invention: The step that joins a chip to its memory has capped how many AI chips could ship.
• Innovation: products and applications: Chips for serving pass through the same step.

Read by: Growth in packaging supply; How concentrated the buyers are

Expected: This site: “Memory and chip packaging are short.” can't be tested yet.

The fast stacked memory, called high-bandwidth memory, that sits beside an AI chip, made by SK Hynix, Samsung and Micron.

Why it is not scored: None of its gauges has a usable reading today: each is missing or older than its age limit.

What its score reads: How fast the supply of high-bandwidth memory is growing. Epoch counts it in dollars, so a price rise looks the same as more memory.

• Capability / invention: The fast memory stacked beside each AI chip comes from a few makers.
• Innovation: products and applications: Serving large models is limited by memory as much as by arithmetic.

Read by: Growth in memory supply, in dollars

Expected: This site: “Memory and chip packaging are short.” can't be tested yet.

The processors that train and run models, from NVIDIA, AMD, Google, Amazon and others.

What its score reads: How much of the chip supply one designer holds, and whether the cost of a unit of computing power is still falling.

• Capability / invention: The chips that train models are the input labs compete for most directly.
• Innovation: products and applications: The same chips serve the products.

Read by: How concentrated chip supply is by designer; Yearly change in the cost of computing power; Yearly change in the rent for NVIDIA's most-rented AI chip

Expected: This site: “NVIDIA and AMD take more than 50% of the gross profit they share with Microsoft's cloud arm and the estimates for OpenAI and Anthropic.” tonight's reading fits both sides.

The switches, cables and optical links that join thousands of chips into one machine.

Why it is not scored: This site reads no series on networking gear for AI clusters.

• Capability / invention: Training one model means wiring a building's worth of chips together.

The money behind all of it, above all the spending of the five largest builders of data centres.

What its score reads: How fast Microsoft, Alphabet, Amazon, Meta and Oracle are raising their spending on buildings and equipment, over the past four quarters against the four before.

• Capability / invention: Frontier training runs are paid for long before they earn anything.
• Innovation: products and applications: Serving capacity is built ahead of the demand it is meant for.

Read by: Growth in the largest builders' spending

Expected: This site: “Money is not what is scarce.” holding so far. David Cahn: “The biggest cloud and platform companies spend more each year on buildings and equipment than AI earns in revenue.” tonight's reading fits both sides.

The low-level programs, such as NVIDIA's CUDA, that let models run fast on a given chip.

Why it is not scored: The activity of the open-source projects is public, but this site does not read it, and more activity could mean either a shortage being fixed or none at all.

• Capability / invention: The software that makes chips usable decides which chips a lab can switch to.

The software and know-how that keep tens of thousands of chips working on one training run without wasting them.

Why it is not scored: Epoch lists how fully some training runs used their chips, but almost none are the largest runs this input is about, since the labs running them publish nothing, and the figures are measured against different chip speeds, so the same skill can read as twice as tight.

• Capability / invention: Running one training job across a whole building of chips without failing is its own skill.

The labs that train the largest models, such as OpenAI, Anthropic, Google DeepMind, xAI and Meta.

• Capability / invention: A handful of teams know how to build a frontier model from scratch.

Read by: How fast the computing power behind frontier models doubles

Expected: Dan Sundheim: “Only a handful of labs stay at the frontier, because it costs too much capital for more.” holding so far.

Everything done to a model after its first training to make it useful, including practice tasks with automatic graders and feedback from experts.

Why it is not scored: Nobody publishes what this work costs or how long buyers wait for it. This site counts the sellers and keeps a tally of disclosed deals, which says the business exists, not whether it is short.

• Capability / invention: Turning a raw model into a useful one needs data, graders and practice tasks built by people.
• Innovation: products and applications: A model is shaped for each kind of work in the same way.

Expected: This site: “Practice tasks with automatic graders, and the expert-written data they consume, are where the shortage goes next.” failing its test. Sean Cai: “Labs spend more each year with the firms that sell practice tasks, graders and expert data.” holding so far.

Both claims here turn on one hand-kept tally of the labs' disclosed deals with these sellers. This site holds its own prediction to that tally; a named writer's claim is not judged on it until every seller counted a year before has a deal on file for the latest year, because most such buying is never announced.

Getting a model to keep learning from its use after release without forgetting what it knew.

Why it is not scored: No public test measures whether a model keeps learning after release. This site tracks what vendors claim, which is not a measure of scarcity.

• Capability / invention: Models do not yet learn on the job; each improvement waits for the next training run.
• Innovation: products and applications: A product cannot get better from its own use without it.
• Diffusion: early adoption: Firms adopt faster what adapts to them, and today's systems mostly do not.

The benchmarks and testing firms that say what a model can do.

Why it is not scored: This site holds many benchmark scores, but none of them measures whether good testing is itself in short supply.

• Capability / invention: Labs can only improve what they can test.
• Innovation: products and applications: A product is shipped only once its builders can tell it works.
• Diffusion: early adoption: Checking an AI system's work is a cost every buyer pays before trusting it.

Text written or checked by paid experts, generated data, and licensed archives.

Why it is not scored: Sellers' revenues, which this site tracks from press reports, show activity. Nobody publishes prices or waiting times, which are what scarcity would show up in.

• Capability / invention: The stock of public text is running out, so new data increasingly has to be licensed, made or bought.

Expected: This site: “Practice tasks with automatic graders, and the expert-written data they consume, are where the shortage goes next.” failing its test. Dwarkesh Patel and Jerry Han: “Better data, more than better model designs, drives how much a model learns from its training, so progress comes to rest on finding new data.” can't be tested yet. Sean Cai: “Training for long work whose results are hard to check, in fields such as biology, security and matters of taste, comes to rest on records of real work inside firms, because tasks staged by contractors stop teaching models.” can't be tested yet. Sean Cai: “Labs buy more of their training data straight from leading app companies and specialists, squeezing the firms that sell general practice environments.” can't be tested yet. Sean Cai: “The market for training data stays split among many specialist sellers rather than gathering around one or two.” can't be tested yet. Sean Cai: “Firms outside the labs become large buyers of training data and practice tasks, so the labs stop being almost the only customers.” can't be tested yet.

Serving a trained model to users, which the industry calls inference.

What its score reads: The price, in dollars for a million tokens of output, of the cheapest model among a named basket of flagship models on one marketplace. A token is a word or a piece of one. One vendor sets it.

• Innovation: products and applications: Every answer a product gives costs computing time.
• Diffusion: early adoption: What a firm pays per answer decides where using AI is worth it.

Read by: Cheapest model in the basket; Yearly change in the rent for NVIDIA's most-rented AI chip

Expected: Sean Cai: “With many agents working on a task at once, most valuable office work comes to turn on the cost of computing and the systems around it, not on whether models are good enough.” can't be tested yet.

Services that send each request to whichever model is cheapest or best for it.

Why it is not scored: The spread of list prices is set almost entirely by one very expensive model, so it says nothing dependable about how much a router can save.

• Innovation: products and applications: Picking the right model for each request is part of making a product cheap enough.
• Diffusion: early adoption: A firm that can switch models easily pays less and risks less.

Models that carry out tasks by using software tools, for minutes or hours at a time.

• Innovation: products and applications: A system that acts for people needs tools, permissions and a record of what it did.
• Diffusion: early adoption: Firms hand work to agents only as far as they can see and undo what the agents do.

Read by: How fast the length of task an agent can finish doubles

Expected: This site: “Proof of what an AI agent (a model that carries out tasks with software tools on a person's behalf) did, and who answers for it, becomes a bottleneck once agents handle money.” can't be tested yet. Byrne Hobart and Nikhil Davar; Soren Larson: “Large platforms move to keep outside agents off them more often than they let them in.” can't be tested yet.

Companies that host and run models for other businesses.

Why it is not scored: No public series separates these platforms' capacity from the cost of running models, which the row for running models reads.

• Innovation: products and applications: The platforms that host models for firms decide how quickly a product can reach them.
• Diffusion: early adoption: Security reviews and contracts for those platforms set how fast firms can start.

The products people and companies use, from coding assistants to legal and medical tools.

Why it is not scored: Most of these companies are private, and too few publish their finances to say how hard money is for them to raise.

• Diffusion: early adoption: Many firms first use AI inside products they already buy, for tasks they already have.
• Diffusion: adaptation: A firm changes how it works only once products fit its own processes.

Expected: Soren Larson: “Apps that resell the labs' models see their margins shrink as their use grows.” can't be tested yet. This site: “Companies selling AI products keep lower gross margins than classic software, because every use costs them the model's price.” holding so far. Byrne Hobart; Sean Cai: “Engineers who fit AI into each customer's own work stay scarce, so their pay rises or employers settle for weaker hires.” can't be tested yet.

• Innovation: products and applications: Their own first stage, from a method that works in the lab to a product people can use.

Read by: 50%/80% horizon ratio ●consistent with normal; Developer productivity uplift (METR RCTs) ●consistent with normal; Enterprise pilots with measurable P&L impact ●consistent with normal; Human interventions on 4–8 hour agent tasks (self-reported) ◐emerging; METR 80% time horizon ▲faster than normal; Price per unit of capability (cheapest METR-hour on OpenRouter) ◐emerging

Expected: Rabanser, Kapoor and colleagues: “The agents Princeton's tracker tests grow more dependable far more slowly than they grow more accurate.” tonight's reading fits both sides.

Their barriers by domain: agents 4 14; general 1 2 3 5 6 9 13 17; law 10 11 16; medicine 12; software engineering 7 8 15

• Diffusion: early adoption: Their own second stage, people choosing to use the products.

Read by: Augmentation vs automation share (Anthropic Economic Index) ◐emerging; Human interventions on 4–8 hour agent tasks (self-reported) ◐emerging; Share of work hours assisted by generative AI ◐emerging; US businesses paying for AI (Ramp AI Index) ◐emerging; US firms using AI (Census BTOS) ●consistent with normal

Expected: Arvind Narayanan and Sayash Kapoor: “The share of American firms using AI stays within the pace this site calls normal, below 35%.” tonight's reading fits both sides.

Their barriers by domain: agents 26; general 18 19 20 21 22 24 27; software engineering 23 25

• Diffusion: adaptation: Their own third stage, organisations changing how they work.

Read by: Developer productivity uplift (METR RCTs) ●consistent with normal; Enterprise pilots with measurable P&L impact ●consistent with normal; Entry-level employment shortfall in AI-exposed occupations ◐emerging; Human interventions on 4–8 hour agent tasks (self-reported) ◐emerging; Recent-graduate unemployment rate (NY Fed) ●consistent with normal; Share of work hours assisted by generative AI ◐emerging; US firms using AI (Census BTOS) ●consistent with normal; US labour productivity, year on year ●consistent with normal; US total factor productivity, private nonfarm business ●consistent with normal

Their barriers by domain: agents 40; general 28 29 30 31 32 36 37 38 39 42; law 41; safety 34 35; software engineering 33

• Diffusion: early adoption: One of their own items: not knowing who is liable when AI errs deters firms from adopting it.
• Diffusion: adaptation: They say these limits act across every stage; this site draws them where firms reorganise. Placed by this site.

Read by: FDA-authorised AI-enabled devices, growth ●consistent with normal; FDA-authorised devices built on a large language model ●consistent with normal; US state AI bills introduced (legislative year) ◐emerging

Expected: Byrne Hobart: “AI grows more unpopular even as more people use it, giving politicians on both the left and the right reason to campaign against it.” can't be tested yet.

Their barriers by domain: agents 53; general 43 44 46 47 55; law 48 49 50 51 52; medicine 45; safety 54

• Diffusion: adaptation: They say these limits act across every stage; this site draws them where firms reorganise. Placed by this site.

Read by: Circular financing scale ▲concentrating; Cloud backlog (remaining performance obligations) ▲concentrating

Their barriers by domain: general 57 58 59 61 63; science 56 62; software engineering 60

• Diffusion: adaptation: They say these limits act across every stage; this site draws them where firms reorganise. Placed by this site.

Read by: 50%/80% horizon ratio ●consistent with normal; Human interventions on 4–8 hour agent tasks (self-reported) ◐emerging

Expected: Byrne Hobart: “A formal pause looks unlikely; instead the biggest labs, loosely in step, slow the public release of their most powerful models.” can't be tested yet.

Their barriers by domain: agents 66 67 69; general 64 68; safety 65

• Diffusion: adaptation: They say these limits act across every stage; this site draws them where firms reorganise. Placed by this site.

Read by: Application layer's share of enterprise generative-AI spend (Menlo) ◐emerging; Consumer surplus from generative AI (willingness to accept) ▼dispersing; Lab concentration of routed tokens (OpenRouter, HHI) ◐emerging; Lab recoupment ratio ◐emerging; Price per unit of capability (cheapest METR-hour on OpenRouter) ◐emerging; Semiconductors' share of stack operating income ▲concentrating; US labour productivity, year on year ●consistent with normal; US total factor productivity, private nonfarm business ●consistent with normal; Public cloud software valuation multiples (Clouded Judgement) (not published)

Their barriers by domain: general 70 72 74 75; law 71 73; media 77; software engineering 76

• Diffusion: adaptation: They say these limits act across every stage; this site draws them where firms reorganise. Placed by this site.

Readings with no status: AI capability on science tasks, not the pace of science itself: best score on FrontierMath's hardest, research-level maths problems 100.0% (2026-09-29); best AI score on SciCode, code written to solve research problems across the sciences 66.9% (2026-09-22); best AI score on CritPt, unpublished research-level physics problems 36.0% (2026-10-28)

Their barriers by domain: science 78 79 80 81 82 83 84 85

• Capability / invention: Their reasons AI cannot simply hurry its own spread; this site draws them where AI is made. Placed by this site.

Read by: Agent-workdays per human workday in frontier research (self-reported) ◐emerging; Human interventions on 4–8 hour agent tasks (self-reported) ◐emerging

Expected: Benjamin Jones: “Whole projects stay far behind single skills, with the best score on real freelance projects below 50%.” tonight's reading fits both sides.

Their barriers by domain: general 86 87 88 89

Someone has to check an AI system's work and answer for it when it is wrong.

• Innovation: products and applications: A product that cannot be checked cannot be sold for work that matters.
• Diffusion: early adoption: Firms adopt where checking the work costs less than doing it.

Read by: Agent reliability across repeated runs (tau-bench pass^k) ◐emerging

Readings with no status: Checking AI's work sold as its own category, events in the latest twelve months: 4 (2026-08-27); ai evaluation market, events in the latest twelve months: 1 (2026-03-09); Insurers writing AI's mistakes out of standard policies, events in the latest twelve months: 4 (2026-07-20)

Expected: This site, extending Gillian Hadfield and Andrew Koh: “Checking AI's work is sold as a category of its own.” tonight's reading fits both sides. This site, extending Gillian Hadfield and Andrew Koh: “Insurers write AI's mistakes out of their standard policies, so the cover has to be bought and priced on its own.” tonight's reading fits both sides. Byrne Hobart: “Firms increasingly buy AI's work through a party with assets and a licence that can be held liable, such as an established service firm.” can't be tested yet.

The statistics that count output miss much of what free and cheap AI tools do.

• Diffusion: early adoption: Firms invest where they can see a return, and much of it goes unmeasured.
• Diffusion: adaptation: Reorganisation shows up in the numbers late, if at all.

Read by: Executives reporting no AI impact on their own firm ●consistent with normal; Enterprise pilots with measurable P&L impact ●consistent with normal; Customer-support productivity uplift from a generative-AI assistant ●consistent with normal; Chatbot effect on earnings and hours in exposed occupations (Denmark) ●consistent with normal

Expected: Diane Coyle and John Lourenze Poquiz: “Official output statistics miss most of AI's gains, which come as better quality and saved time that no price records.” can't be tested yet.

Who does the work, and how the income from it is split between workers and owners.

• Diffusion: adaptation: Jobs and pay change slowly, through hiring, training and contracts.

Read by: US labour share of income, year on year ◐emerging; Employment in AI-exposed occupations, year on year ●consistent with normal

Expected: Brynjolfsson, Chandar and Chen: “In the jobs most exposed to AI, workers in their early twenties fall behind workers over fifty, while in the least exposed jobs they do not.” holding so far. Betsey Stevenson: “Work in the jobs most exposed to AI grows more intense before it grows shorter.” can't be tested yet. Alex Imas: “Care, teaching, hospitality and other work where the person is part of what is bought take a rising share of spending and jobs as AI spreads.” holding so far. Pascual Restrepo: “Labour's share of income falls as AI takes on more tasks, even as wages rise.” holding so far. Arvind Narayanan and Sayash Kapoor: “AI does not wipe out software engineering jobs, and demand for engineers may rise as software gets cheaper to make.” can't be tested yet.

The web that AI learns from, and the traffic that pays for it.

• Diffusion: adaptation: If answers replace visits, the sites that produce new information lose the income that funds it.

Readings with no status: AI training's share of crawler requests (Cloudflare) 52.0% (2026-06-30); pages Anthropic's crawler takes for each visitor it sends back (Cloudflare) 50,000× (2025-08-07); pages OpenAI's crawler takes for each visitor it sends back (Cloudflare) 887× (2025-08-07)

Expected: Joseph Stiglitz and Màxim Ventura-Bolet: “Publishers lose the visits that pay for what they produce.” can't be tested yet.