So far, the profit is staying with the chip makers
The AI industry is a stack of layers, from chips and data centres at the bottom to the apps people use at the top. This page follows where the profit settles, layer by layer. Gross profit is what is left of revenue after paying to deliver the product; where a layer publishes none, it is left out rather than guessed.
Where each layer's economics are heading, what is getting cheap and what stays scarce, is on the value chain.
No gauge here is scored for direction yet.
Each square is a published gauge that this site scores for direction: whether profit, sales or financing in that layer are gathering with fewer firms (concentrating) or spreading out to more firms or to buyers (dispersing). A gauge reads too early or unclear until it has enough readings to show a direction. The word beside each layer is a vote of its gauges, counting gauges that share a source once, so it is not a tally of the squares.
It does not show how far a gauge has moved, only its direction, and a layer with few squares is thinly measured, not quiet. Gauges of how fast AI is spreading sit on the other lens.
Every gauge and its reading
StartSomething every buyer of AI needs, such as a chip, a model or an app
Is it scarce?
Scarce means buyers want more of it than its makers can supply at today's price.
If noProfit competed awayMany firms can supply it, so they undercut each other until the price is close to the cost. The saving passes to whoever buys it.
- If yes
Can whoever holds it keep rivals off?
A rival is kept off when copying the thing takes years, costs too much, or needs something only the holder has.
If noProfit for a whileThe holder is paid well today, but the high price draws in rivals and substitutes. As they arrive the price falls and the money moves on to the next scarce thing.
- If yes to bothProfit stays here
Buyers must pay the holder's price, and the profit collects with the holder for as long as both answers stay yes.
What keeps rivals off- A design that takes years to match
- Factories and power connections that take years to build
- Customers whose own work is already built on the product
- Being the place a customer opens first
The gauges on this page do not ask these questions directly. They watch for the result: a layer reads as concentrating when its profit or its sales are gathering with fewer firms, and as dispersing when they are spreading out to more firms or to buyers. This is a drawing of an idea, with no number of its own. It is the rule argued in the essay on where profit moves, from Ricardo on rent and Teece on who profits from an invention; the charts on this page show where profit sits now, which is evidence about the rule and not a test of it.
- Place one
The model
Money arrivingOnly a handful of labs can build the best models, and buyers pay for the gap to the next best.
- the labs' share of the stack's gross profit, estimated6.6% 2026-06-30
- how far open models trail the best closed one7.7 months 2026-10-26
- what closed models charge over open ones at the same skill0.28× 2026-10-19
- open models' share of routed tokens77.6% 2026-11-03
- Money arriving if the labs' share of gross profit is above a year before:
- yes
- Moving on if open models trail the best closed one by under three months:
- not yet
Who is paid here: Frontier labs, Neolabs (founded 2023+), Open-weight & Chinese labs
What would squeeze it: Open models closing the gap, and cheaper models good enough for most work.
- Place two
Running it
Money arrivingServing models quickly, cheaply and reliably, at the scale buyers need.
- inference firms' run-rate growth, median802.6% 2026-09-30
- the labs' run-rate growth, median408.7% 2026-06-30
- rent for NVIDIA's most-rented AI chip, against a year before29.2% 2026-10-26
- Money arriving if inference firms' run-rates grow faster than the labs':
- yes
Who is paid here: Inference serving, Routing, Agent infrastructure
What would squeeze it: Routers and open serving software that make one provider as good as another.
- Place three
The firm's own knowledge
Money arrivingA firm's own records, permissions and ways of working, which no model can learn from the public web.
- application firms' run-rate growth, median562.9% 2026-09-30
- companies above the model layer bought by labs and compute firms in a year46 2026-06-30
- Salesforce's yearly revenue from its AI agents, including Slackbot (company-stated, a floor)$1.5B 2026-07-31
- Money arriving if application firms' run-rates grow faster than the labs':
- yes
- Moving on if labs and compute firms buy more of the layers above them than a year before:
- can't be tested yet
Who is paid here: Vertical AI by domain, Coding agents & dev tools, Deployment / FDE / implementation services, Evals
What would squeeze it: Labs buying or copying the firms that hold it, and general assistants that pick up a workflow from a few connectors.
- Place four
The front door
Paid here, looseningThe assistant a customer opens first, which chooses the suppliers behind it.
- concentration of the visits chatbots send to websites (StatCounter, CC BY-SA)0.37 index 2026-10-31
- people using Google's Gemini app in a month (company-stated)950,000,000 users 2026-06-30
- paid seats for Copilot inside Microsoft's office tools (company-stated, a floor)30,000,000 seats 2026-06-30
- Money arriving if one or two assistants send most of the visits:
- yes
- Moving on if that concentration is below a year before:
- yes
Who is paid here: Consumer AI
What would squeeze it: Customers switching assistants easily, and operating systems or browsers that open first.
The rule is the one the migrating bottleneck essay sets out: the money goes to whatever cannot be routed around, and moves when that is fixed. Applied up the stack, it passes from the model to running it, to a firm's own knowledge, to the customer's front door, as positions in what people expect argue. While the build-out lasts, the chip makers are paid first wherever the money settles: both sides expect this
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).
The numbers
| quarter | part | share | record |
|---|---|---|---|
| 2026-06-30 | Chips | 72.6% | derived row |
| 2026-06-30 | Cloud | 20.8% | derived row |
| 2026-06-30 | Labs (estimated) | 6.6% | derived row |
| 2026-03-31 | Chips | 73.6% | derived row |
| 2026-03-31 | Cloud | 21.6% | derived row |
| 2026-03-31 | Labs (estimated) | 4.8% | derived row |
| 2025-12-31 | Chips | 71.8% | derived row |
| 2025-12-31 | Cloud | 24.5% | derived row |
| 2025-12-31 | Labs (estimated) | 3.7% | derived row |
| 2025-09-30 | Chips | 69.6% | derived row |
| 2025-09-30 | Cloud | 27.8% | derived row |
| 2025-09-30 | Labs (estimated) | 2.6% | derived row |
| 2025-06-30 | Chips | 65.7% | derived row |
| 2025-06-30 | Cloud | 32.1% | derived row |
| 2025-06-30 | Labs (estimated) | 2.2% | derived row |
| 2025-03-31 | Chips | 63.9% | derived row |
| 2025-03-31 | Cloud | 34.6% | derived row |
| 2025-03-31 | Labs (estimated) | 1.5% | derived row |
The same shares as the columns in the figures of gross profit and operating income on this page, read at their ends: where each part stood in the newest quarter on file and in the same quarter a year before. Gross profit is revenue less the cost of delivering the product; operating income also takes off research, selling and running costs. A share can rise while the money itself falls, and the reverse, because every part is measured against the total of the parts drawn. Apps, and labs in the operating measure, file nothing usable and are left out, so these are shares of what can be measured, not of the whole industry.
The labs' share is this site's estimate, not a filing: for OpenAI and Anthropic, the latest run-rate reported before the quarter ended, divided across the year's quarters, times Epoch AI's estimate of what is left after the cost of running the models. A run-rate is a recent month or quarter of sales multiplied up to a year. The run-rate used can be some months older than the quarter, so the labs' two dots are not exactly a year apart and the newest one is likely too low. The run-rates used are dated: Anthropic, May 2025 for Q2 2025 and May 2026 for Q2 2026; OpenAI, June 2025 for Q2 2025 and February 2026 for Q2 2026. In the gross-profit rows chips are the whole of NVIDIA and AMD and cloud is a single company's cloud segment; in the operating-income rows both are the segments named beneath the last figure on this page.
The numbers
| Measure | Part | A year before | Newest quarter | Change |
|---|---|---|---|---|
| gross profit | Chips | 65.7% Q2 2025 | 72.6% Q2 2026 | +7.0 pts |
| gross profit | Cloud | 32.1% Q2 2025 | 20.8% Q2 2026 | −11.3 pts |
| gross profit | Labs (estimated) | 2.2% Q2 2025 | 6.6% Q2 2026 | +4.3 pts |
| operating income | Chip segments | 52.9% Q2 2025 | 61.0% Q2 2026 | +8.1 pts |
| operating income | Cloud segments | 47.1% Q2 2025 | 39.0% Q2 2026 | −8.1 pts |
The layers, from the chips up
Why it matters: whichever layer holds the scarce part keeps the profit, and the four places above say where that is moving.
The chip makers, the big cloud companies and the newer rented-computing firms, data centres and the power they draw, memory and networking.
- Rent concentration in semiconductorsconcentrating
- Cloud backlog (remaining performance obligations)concentrating
- Cloud segment operating marginsconcentrating
The labs that build AI models: the leaders such as OpenAI and Anthropic, newer labs founded since 2023, and labs that publish their models openly, many of them Chinese.
- METR 50% time horizonfaster than normal
- METR 80% time horizonfaster than normal
- 50%/80% horizon ratioconsistent with normal
What models are trained and tested on: work by paid human experts, practice tasks for training by trial and error, generated data, licensed content, and tests.
- Expert-data market run-rate (Mercor)emerging
- Continual-learning ladder (highest production rung)emerging
- Public text data exhaustion year (Epoch projection)emerging
The businesses that run models for others, route requests between them, and supply the plumbing for AI agents.
The layer in full →The products people use and the firms that put AI to work: coding assistants, apps built for one industry, consumer apps, AI-native firms that sell finished work, firms that train the people who deploy AI, and firms that buy up service businesses to run them with AI.
- Developer productivity uplift (METR RCTs)consistent with normal
- Waymo paid rides per weekconsistent with normal
- FDA-authorised devices built on a large language modelconsistent with normal
The companies and people using the products. Tracked through its readings; no company list.
- Share of work hours assisted by generative AIemerging
- Customer-support productivity uplift from a generative-AI assistantconsistent with normal
- US firms using AI (Census BTOS)consistent with normal
Workers and consumers: wages, jobs, and the value users get without paying for it. Tracked through its readings; no company list.
- US labour productivity, year on yearconsistent with normal
- Employment in AI-exposed occupations, year on yearconsistent with normal
- Chatbot effect on earnings and hours in exposed occupations (Denmark)consistent with normal
Among the listed companies that file a segment
Within one router's published top table
The index is the Herfindahl index, HHI for short: square each firm's share of a market and add the squares. It reaches the right end when a single firm holds everything. A segment is a division a company reports separately; a router is a service that passes a request to whichever model the buyer picks, its top table is the list of most-used models it publishes each day, and tokens are the pieces of text a model reads and writes. The chip and cloud rows each count only a handful of listed companies' filed segments and say nothing of firms that file none. The lab and model rows count the text passing through a single router, OpenRouter, in its published top table; large buyers mostly go direct to the labs and never appear there. The rows are read on different dates, shown beside each, and what each index counts is in the table beneath. An index over a few firms cannot read low however evenly they share: the block at the left of each row is the part it cannot reach. Read each dot against its own block, and do not rank one row against another.
The numbers
| Index | Layer | Reading | Counts | As of | Reads now | What it counts |
|---|---|---|---|---|---|---|
| Concentration among covered chip filers (HHI) | Compute & physical | 0.87 | 2 firms | June 2026 | stable | Herfindahl index of data-centre revenue between NVIDIA (Data Center segment) and AMD (Data Center segment), per calendar quarter, from filed segment disclosures. Derived metric hhi_by_layer, semis slice. |
| Concentration among covered cloud filers (HHI) | Compute & physical | 0.35 | 3 firms | June 2026 | stable | Herfindahl index of segment revenue between AWS, Google Cloud and Microsoft Intelligent Cloud, per calendar quarter, from filed segment disclosures. Derived metric hhi_by_layer, cloud slice. |
| Lab concentration of routed tokens (OpenRouter, HHI) | Model | 0.24 | 12 labs | November 2026 | emerging | Herfindahl index of prompt-plus-completion tokens by lab across the models in OpenRouter's daily top table, each model counted toward the lab in its OpenRouter id, last reading each week. Derived metric lab_token_hhi. One router's traffic and the listed slice only. |
| Model-layer token concentration on OpenRouter (HHI, top table) | Model | 0.14 | 20 models | November 2026 | emerging | Herfindahl index of prompt-plus-completion tokens across the models in OpenRouter's daily top table, derived metric model_token_hhi_top10; one router's traffic, the listed slice only. |
Spending ahead of sales
Why it matters: a lab's own profit depends on its sales catching up with what it has raised and signed up to pay. These figures set the two side by side, and show how much is promised between the labs and their own suppliers.
Too little on record to compare
- xAI: the latest figure on record, $428M a year, is from September 2025, too old to set beside the others. A further $20.0B was raised after that date.
- DeepSeek: sells $1.0B a year as of September 2026 (a press report, graded as likely by Epoch AI). Only 1 funding round is on record, so no multiple is drawn.
- MiniMax: sells $800M a year as of August 2026 (stated by the company, graded as likely by Epoch AI). Only 1 funding round is on record, so no multiple is drawn.
- Zhipu AI: sells $1.6B a year as of August 2026 (stated by the company and reported by the press, graded as confident by Epoch AI). Only 1 funding round is on record, so no multiple is drawn.
- Cohere and Moonshot AI: a run-rate is on record but no funding round, so nothing is drawn.
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.
The numbers
| Lab | Run-rate | Dated | Who said it | Paid in by then | Rounds on record | Raised after that date | Run-rate over money paid in | Signed up to pay suppliers | Deals |
|---|---|---|---|---|---|---|---|---|---|
| OpenAI | $50.0B | September 2026 | a press report, graded as likely by Epoch AI | $183B | 11 | — | 0.27× | $730B | 8 |
| Anthropic | $65.0B | July 2026 | a press report, graded as likely by Epoch AI | $139B | 14 | — | 0.47× | $194B | 4 |
| xAI | $428M | September 2025 | a press report, graded as likely by Epoch AI | $17.0B | 3 | $20.0B | not drawn: too little on record | no priced deal on the ledger | — |
| DeepSeek | $1.0B | September 2026 | a press report, graded as likely by Epoch AI | $7.5B | 1 | $15.9B | not drawn: too little on record | no priced deal on the ledger | — |
| Mistral AI | $1.0B | September 2026 | stated by the company, graded as confident by Epoch AI | $3.0B | 5 | $3.5B | 0.33× | no priced deal on the ledger | — |
| MiniMax | $800M | August 2026 | stated by the company, graded as likely by Epoch AI | $619M | 1 | — | not drawn: too little on record | no priced deal on the ledger | — |
| Zhipu AI | $1.6B | August 2026 | stated by the company and reported by the press, graded as confident by Epoch AI | $558M | 1 | — | not drawn: too little on record | no priced deal on the ledger | — |
Thickness is the full amount written in the deals, not money paid. OpenAI is a party to 68.5% of everything drawn. The ledger records no direction, so no arrow is drawn.
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.
Every pairing and the full amount written
| Supplier or investor | Other side | Kind | Full amount written | Deals | Of which press-reported | Of which disputed |
|---|---|---|---|---|---|---|
| Oracle | OpenAI | a promise to buy | $300B | 1 | 1 | 1 |
| Microsoft | OpenAI | a promise to buy | $250B | 1 | — | — |
| Amazon | OpenAI | a promise to buy | $138B | 2 | — | — |
| NVIDIA | OpenAI | a stake or a guarantee | $135B | 2 | 1 | 1 |
| Amazon | Anthropic | a promise to buy | $100B | 1 | — | — |
| SoftBank Group | OpenAI | a stake or a guarantee | $60.0B | 2 | — | — |
| Amazon | OpenAI | a stake or a guarantee | $50.0B | 1 | — | — |
| Nscale | Anthropic | a promise to buy | $44.6B | 1 | — | — |
| Alphabet | Anthropic | a stake or a guarantee | $40.0B | 1 | 1 | — |
| CoreWeave | Meta | a promise to buy | $35.2B | 2 | — | — |
| Microsoft | Anthropic | a promise to buy | $30.0B | 1 | — | — |
| Nebius | Meta | a promise to buy | $29.9B | 3 | — | — |
| Amazon | Anthropic | a stake or a guarantee | $29.0B | 2 | — | — |
| CoreWeave | OpenAI | a promise to buy | $22.4B | 3 | — | — |
| Cerebras | OpenAI | a promise to buy | $20.0B | 1 | — | — |
| TeraWulf | Anthropic | a promise to buy | $19.0B | 1 | — | — |
| Microsoft | Nebius | a promise to buy | $17.4B | 1 | — | — |
| TeraWulf | Fluidstack | a promise to buy | $16.2B | 3 | — | — |
| Applied Digital | CoreWeave | a promise to buy | $11.0B | 2 | — | — |
| NVIDIA | Anthropic | a stake or a guarantee | $10.0B | 1 | — | — |
| Core Scientific | CoreWeave | a promise to buy | $9.9B | 5 | — | — |
| IREN | Microsoft | a promise to buy | $9.7B | 1 | — | — |
| Hut 8 | Fluidstack | a promise to buy | $7.0B | 1 | — | — |
| CoreWeave | NVIDIA | a stake or a guarantee | $6.3B | 1 | — | — |
| Cipher Mining | Amazon | a promise to buy | $5.5B | 1 | — | — |
| Microsoft | Anthropic | a stake or a guarantee | $5.0B | 1 | — | — |
| NVIDIA | Sharon AI | a promise to buy | $4.9B | 1 | — | — |
| Cipher Mining | Fluidstack | a promise to buy | $3.8B | 2 | — | — |
| IREN | NVIDIA | a promise to buy | $3.4B | 1 | — | — |
| NVIDIA | Nscale | a stake or a guarantee | $2.6B | 3 | — | — |
| NVIDIA | IREN | a stake or a guarantee | $2.1B | 1 | — | — |
| NVIDIA | Nebius | a stake or a guarantee | $2.0B | 1 | — | — |
| NVIDIA | CoreWeave | a stake or a guarantee | $2.0B | 1 | — | — |
| Microsoft | G42 | a stake or a guarantee | $1.5B | 1 | — | — |
| NVIDIA | Nscale | a promise to buy | $1.2B | 1 | — | — |
| Microsoft | Mistral AI | a stake or a guarantee | $16.3M | 1 | 1 | — |
Derived (margin_stack_share_by_layer) from: SEC EDGAR 10-Q/10-K XBRL instances (segment data). SEC EDGAR, inline XBRL filings (segment disclosures).
The numbers
| quarter | part | share | record |
|---|---|---|---|
| 2026-06-30 | Chip segments | 61.0% | derived row |
| 2026-06-30 | Cloud segments | 39.0% | derived row |
| 2026-03-31 | Chip segments | 61.4% | derived row |
| 2026-03-31 | Cloud segments | 38.6% | derived row |
| 2025-12-31 | Chip segments | 59.1% | derived row |
| 2025-12-31 | Cloud segments | 40.9% | derived row |
| 2025-09-30 | Chip segments | 56.4% | derived row |
| 2025-09-30 | Cloud segments | 43.6% | derived row |
| 2025-06-30 | Chip segments | 52.9% | derived row |
| 2025-06-30 | Cloud segments | 47.1% | derived row |
| 2025-03-31 | Chip segments | 48.1% | derived row |
| 2025-03-31 | Cloud segments | 51.9% | derived row |
| 2024-12-31 | Chip segments | 52.5% | derived row |
| 2024-12-31 | Cloud segments | 47.5% | derived row |
| 2024-09-30 | Chip segments | 50.2% | derived row |
| 2024-09-30 | Cloud segments | 49.8% | derived row |
| 2024-06-30 | Chip segments | 49.1% | derived row |
| 2024-06-30 | Cloud segments | 50.9% | derived row |
| 2024-03-31 | Chip segments | 47.0% | derived row |
| 2024-03-31 | Cloud segments | 53.0% | derived row |
Every layer's reading
As of 2026-11-04: compute & physical concentrating, model mixed, training input emerging, serving & orchestration emerging, deployment & application emerging, adopters (no capture rule yet), labour & consumers dispersing (one reading).
Recent changes
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.
What would change each reading
- Rent concentration in semiconductors: Four fiscal quarters; a move under one point of margin is noise. Rising margin = rent concentrating at the chip layer.
- Cloud backlog (remaining performance obligations): Four quarters; moves under $10B are noise at this scale. Growing backlog concentrates future cloud revenue in a few contracts.
- Frontier-lab revenue run-rates: Four reports; moves under $2B are reporting noise. Rising run-rate concentrates revenue at the two frontier labs.
- Lab recoupment ratio: Three points; a move under 0.05 is noise. A rising ratio means the lab is recouping capital faster than it raises it.
- Consumer surplus from generative AI (willingness to accept): Two annual waves so far, so the trend is one step; moves under $10B are within survey error. Rising surplus disperses value away from producers to consumers.
- Application layer's share of enterprise generative-AI spend (Menlo): Two annual waves, so the trend is one step; a move under five points is within survey error. Rising application share = value concentrating at the application layer.
- Semiconductors' share of stack operating income: Four calendar quarters; a move under two points is within fiscal-calendar noise. Rising semis share = rent concentrating at the chip layer; falling = migrating up.
- Chips' share of the stack's gross profit: Four calendar quarters, the same rule as the operating-income stack; a move under two points is fiscal-calendar noise. Rising = profit concentrating at the chip layer; falling = migrating up the stack.
- Cloud segment operating margins: Four fiscal quarters; two points of margin is noise. Rising cloud margin under rising capex means the layer is keeping surplus.
- Circular financing scale: Three quarters of the trailing-year flow; moves under $20B are one small deal. Rising commitments concentrate the stack's financing on a few counterparties, and unlike the cumulative total this series can fall.
- Concentration among covered chip filers (HHI): Four calendar quarters; a move under 200 HHI points is noise. Rising = the covered chip layer concentrating further.
- Concentration among covered cloud filers (HHI): Four calendar quarters; a move under 200 HHI points is noise.
- Lab concentration of routed tokens (OpenRouter, HHI): Four weekly readings; under 0.02 in the index is table churn. Rising = routed tokens concentrating on fewer labs.
- Price per unit of capability (cheapest METR-hour on OpenRouter): Four price change-points; a move under two cents per million tokens per horizon-hour is noise. Rising = labs keeping the surplus from capability gains.
- Price per unit of capability (cheapest near-frontier model, Artificial Analysis): Four price change-points; a move under $0.0006 per million tokens per index point (about a tenth of the 11 Sep 2026 level) is one model's repricing. Rising = labs keeping the surplus from capability gains.
- Frontier labs' share of tracked venture dollars: Four quarters; a move under five points is a single round's timing. Rising = capital concentrating in the frontier labs.
- Vertical versus horizontal application funding: Four quarters; a move under 0.2x is round timing. Rising = capital concentrating in vertical applications.
- Frontier lab valuation to run-rate multiple: Three valuation events; a move under 3x is noise in press-reported run-rates. Rising = the market pricing more durable rents into the labs.
- Lab vertical integration (acquisitions and acquihires, trailing four quarters): Four quarters. A yearly count of rare events varies between two years with no trend by about the square root of twice its level, which at the December 2025 count after the September 2026 search is about seven, so a smaller move is stable. More deals per year concentrate the application layer into the labs.
- Open-weights gap to the closed frontier (ECI points): Four open-weights releases; a move under two ECI points is inside the index's confidence intervals. Widening gap = rents concentrating in the closed labs.
- Non-Nvidia share of AI compute shipped (H100-equivalents): Four quarters; under two points is inside Epoch's estimate ranges. Rising non-Nvidia share = the chip layer's rent dispersing.
- Hyperscaler capex to AI revenue (trailing year): Four quarters; a move under 0.3x is run-rate noise. Rising = capex pulling further ahead of revenue (the bet growing); falling = revenue catching the buildout.
- Hyperscaler server useful lives (10-K estimates): Two annual readings; a change under half a year is rounding. A shorter life means faster hardware turnover and thinner margins for the layer, so shortening reads as dispersing.
- Neocloud borrowing cost (CoreWeave facility rates): Three quarter-ends; moves under a percentage point are facility mix. Rising rates concentrate the compute layer's returns with its lenders.
- Enterprise multi-model usage share (Menlo): Annual waves; a move under five points is survey noise. More multi-homing disperses the model layer's pricing power.
- Startups' share of enterprise AI application revenue: Two annual waves, so the trend is one step; under five points is survey error. Rising startup share = value dispersing away from incumbents.
- Frontier lab gross margin (estimated): Year on year; a move under five points is noise. A rising margin means the lab layer keeps more of what it sells.
- GPU rental price per hour: Eight weekly readings, about two months. A move of a quarter an hour or less is noise in a median across some forty providers whose panel changes week to week (a third of four-week moves in the first year exceeded fifteen cents); rising rents mean the owners of GPUs keep more of the value.
- Model-layer token concentration on OpenRouter (HHI, top table): Weekly readings; moves under 0.02 in the index are table churn. A rising index means routed tokens concentrating on fewer models.
- Labs buying from the training-input layer (enter bell): Four quarters; a one-point move is one vendor's single disclosure. Rising = the labs buying more from the training-input layer.
- Application-company revenue run-rates: The rule compares three readings from one compiler (Dealroom) for the largest application company, Anysphere (Cursor); moves under $300M are reporting noise. Other publishers reporting the same company (TechCrunch, Forbes) are the second source. Rising run-rate means the application layer is keeping more of the surplus.
- Inference-layer revenue run-rates: The rule compares three estimates from one estimator (Sacra) for the largest serving firm; moves under $100M are estimation noise. Other publishers reporting the same firm (the Fireworks blog, CNBC) are the second source. Rising run-rate means the serving layer is keeping more of what passes through it.