Skip to content
Slow Variables
Ask

Capture / layer 5 · The value chain

Deployment & application

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. Dependency tier 4.

How the economics move

The site's judgement of each part of this layer, read against the readings below and the theory on the value chain page: what is getting cheap and interchangeable, what stays scarce, and what would prove it wrong.

What the powers mean

Hamilton Helmer, a strategy consultant and investor who studied why some firms keep earning more than their investors could get elsewhere, found that an advantage lasts only when a benefit sits behind a barrier that stops rivals copying it, and named the kinds that recur:

a cornered resource
:
preferential access to something others need: a patent, a licence, a scarce input, a rare team
branding
:
buyers pay more for the same thing because of what they believe about who made it
counter-positioning
:
a newcomer's model the incumbent could copy only by damaging its own business
network economies
:
each user makes the product more valuable to the others, so the leader's lead feeds itself
process power
:
a way of working that improves cost or quality and takes a rival years to learn
scale economies
:
unit costs fall as volume grows, and a rival cannot match them without matching the volume
switching costs
:
a customer would lose money, time or work by moving to a rival
Holding

Fitting AI into a firm earns a services margin that grows with headcount; a rent lasts only where the deployer leaves behind software the customer runs its work on

Deployment firms send engineers into a customer to connect AI models to its data, systems and daily work. The industry calls them forward-deployed engineers: staff who sit with the customer rather than at head office. This site judges Palantir the largest listed firm doing this work, though Palantir describes itself as selling software platforms. The labs have set up their own deployment arms, among them the OpenAI Deployment Company and a venture Anthropic formed with Blackstone and other investors (reports in this site's records differ on who the others are) to serve companies owned by private-equity firms (investment firms that buy whole companies); and consultancies sell the same work. David Ricardo, an English economist who asked in 1817 why landlords grew rich when bread grew dear, found that the owner of the scarcest input collects a rent: profit above what it takes to keep a business going, earned because something holds rivals off. This site judges that the scarce input here is people who understand both the models and the customer's business, and that selling their hours earns a wage-like margin, because each new customer needs more of them. Oliver Williamson, an economist who studied contracts, found that an asset built for one buyer, worth far less anywhere else, ties that buyer to whoever holds it. This site judges that the rent appears only when the deployment leaves behind such an asset: software and a map of the customer's data that its work comes to run on.

Getting cheap and interchangeable
This site judges that generic integration, wiring a model to a firm's documents and common business software, is becoming a commodity: the clouds and labs ship connectors (links that let a model read and act in other software) and tools for building agents (AI programs that carry out tasks on their own), and each new model needs less hand-fitting. William Nordhaus, an economist who measured how much of the value of new inventions their makers keep as profit, found that it is a small part; most goes to buyers. David Teece, an economist who studied why the firm that invents something often fails to keep the returns, called the inputs a product needs to reach its customers complementary assets. This site judges that the deployer owns neither the model nor the customer's records, so on generic work it earns an ordinary services return.
Staying scarce
This site judges that what stays scarce is engineers who can turn a firm's messy data and unwritten habits into working systems, and, once that is done, the map of the customer's data and operations that the system runs on. Hamilton Helmer, a strategy consultant who studied why some firms keep earning more than their investors could expect elsewhere, found that a lasting advantage is a benefit behind a barrier. Switching costs (what a customer would lose by moving to a rival) are the likeliest here, and Palantir's expanding customers fit them, but its filing reports expansion, not how many customers leave, so this site does not count the power proven. The rival view holds that these engineers are scarce enough to collect a scarcity rent themselves, not a wage-like margin.
What would make it last
A deployer's work ends in its own platform, which the customer's operations come to run on, so replacing it means rebuilding the work, not repointing it: switching costs in Helmer's sense. Ronald Coase, an economist who studied why some work is done inside firms rather than bought, and Williamson would add that a lab owning the deployer keeps its model inside that platform, which is why the labs are buying consultancies. The rent stays with services margins if the models learn to fit themselves to a firm's work.
What would prove this wrong
Deployment firms keeping or widening their margins while the models need less fitting with each release and the labs give deployment away to win model sales; or, the other way, deployment staff becoming easy to hire while customers still stay with the firm that first fitted their systems.
Binds on (how tight)
Deployment platforms (no reading; a model judges it probably easing), Agents and tool use (easing), Applications (no reading; a model judges it probably slack)
Commoditising

Coding agents are among the fastest-growing AI apps, but the labs sell their own and a lab's owner now holds Cursor; an independent agent's margin is a lease on the labs' models

A coding agent is an AI tool that writes, edits and tests software, from suggesting the next line to carrying out a whole task on its own. The independent sellers include Cognition, Replit, Lovable, Factory and Augment. The labs and their owners sell their own: Anthropic's Claude Code, OpenAI's Codex, and Cursor, made by Anysphere, which SpaceX, the owner of the lab xAI, bought in 2026 and which now works with xAI on its Grok models; and Microsoft sells coding tools through GitHub, a site where developers keep and share code. This site judges that the independent agents mostly run on models bought from the labs, paying by the token (a fragment of a word), though some train models of their own. David Ricardo, an English economist who asked in 1817 why landlords grew rich when bread grew dear, found that the owner of the scarcest input collects a rent: profit above what it takes to keep a business going, earned because something holds rivals off. This site judges that the scarce input in a coding agent is the model underneath, which the labs own, so an independent agent keeps a rent only for what it adds: its editor (the program in which developers write and read code), its grip on a firm's code and habits, and its own models where it trains them.

Getting cheap and interchangeable
This site judges that the agent's wrapper, the editor, the chat window and the loop that lets a model run and test code, is becoming a commodity: several sellers offer much the same, open-source versions exist, and the labs bundle their own with their models. William Nordhaus, an economist who measured how much of the value of new inventions their makers keep as profit, found that it is a small part; most goes to buyers as lower prices and better products. David Teece, an economist who studied why the firm that invents something often fails to keep the returns, found that the firm holding the scarce inputs a product needs usually keeps the profit; here the lab holds the model, the scarcest input.
Staying scarce
This site judges that what stays scarce is the best coding model at any moment, and, for an agent maker, the record of how developers accept, reject and correct its work, which could train its own models. Andrei Hagiu and Julian Wright, economists who asked when learning from customers' data keeps rivals out, found it does so only when the data is proprietary, keeps its value and improves the product for every customer; a record of accepted code may pass the first and last tests, but goes stale as models improve. Hamilton Helmer, a strategy consultant who studied why some firms keep earning more than their investors could expect elsewhere, found that a lasting advantage is a benefit behind a barrier. Switching costs (what a customer would lose by moving to a rival) are the candidate, but this site judges that a firm can try a rival agent alongside the one it has, and no seller publishes figures on customers leaving, so this site does not count them proven.
What would make it last
An agent becomes the place a firm's software is planned, reviewed and run, holding its rules, permissions and history, so leaving means rebuilding how its engineers work: switching costs in Helmer's sense. Or its record of developers' corrections trains models of its own good enough to cut what it pays the labs. Ronald Coase, an economist who studied why some work is done inside firms rather than bought, would otherwise expect the owners of models to bring the agents inside, as SpaceX did with Cursor.
What would prove this wrong
An independent coding agent keeping its customers and its margin through a year in which the labs cut their own agents' prices and ship better models to everyone at once, with customers not moving to the labs' tools.
Binds on (how tight)
Agents and tool use (easing), Applications (no reading; a model judges it probably slack), Running models (easing)
Holding

A trade's own AI app keeps a rent where it holds the trade's workflow, records and approvals; where it only packages a general model, each lab release takes the margin

Vertical AI apps are built for one trade: Harvey and Legora for lawyers, OpenEvidence, Abridge and Ambience for doctors, Hebbia and Rogo for finance, Sierra and Decagon for customer service. They buy models from the labs and wrap them in the forms, checks and records of one kind of work. David Ricardo, an English economist who asked in 1817 why landlords grew rich when bread grew dear, found that the owner of the scarcest input collects a rent: profit above what it takes to keep a business going, earned because something holds rivals off. David Teece, an economist who studied why the firm that invents something often fails to keep the returns, found that the profit often goes to whoever holds the complementary assets: the distribution, the trade's trust and the systems a product needs to reach its buyers. This site judges that a vertical app earns a rent only where it holds such assets that a lab cannot copy by training a better general model.

Getting cheap and interchangeable
This site judges that a vertical app that adds only prompts (the written instructions given to a model) and a trade's vocabulary to a general model is becoming a commodity, since each general release learns more trades at once. William Nordhaus, an economist who measured how much of the value of new inventions their makers keep as profit, found that it is a small part; most goes to buyers as lower prices and better products. Kenneth Arrow, an economist who studied how markets handle knowledge, argued that information is hard to own because copying it is cheap; a clever way of asking a model a legal question is information of that kind.
Staying scarce
This site judges that what stays scarce is the place in the trade's daily work (the tool a lawyer drafts in or a doctor's notes are written by), the trade's own records the app is allowed to see, licensed content the trade trusts, such as the medical journals OpenEvidence draws on, and approvals that only a regulator or a licensed professional can give. Andrei Hagiu and Julian Wright, economists who asked when learning from customers' data keeps rivals out, found it does so only when the data is proprietary, keeps its value and improves the product for every customer; much of a trade's data belongs to each client and helps only that client. Hamilton Helmer, a strategy consultant who studied why some firms keep earning more than their investors could expect elsewhere, found that a lasting advantage is a benefit behind a barrier. Switching costs (what a customer would lose by moving to a rival) and branding (buyers paying more because of the name) are the candidates, but these firms are young and publish no figures on customers leaving, so this site counts neither proven.
What would make it last
A vertical app becomes the system of record for its trade, the software that holds a firm's matters, patients or cases, so switching costs in Helmer's sense hold customers when the labs catch up; or a regulator's approval, or the trade's liability rules, make the app's checked record something a lab cannot simply match, turning its rent into a regulated one. Ronald Coase, an economist who studied why some work is done inside firms rather than bought, would otherwise expect the labs to keep buying the apps whose trades they want.
What would prove this wrong
Vertical apps' prices and margins falling each time the labs release a better general model, with customers switching to the labs' own specialist products; or, the other way, a vertical app keeping its customers and prices after a lab ships a rival product for the same trade.
Binds on (how tight)
Applications (no reading; a model judges it probably slack), Training data (no reading; a model judges it probably moderate), Tests and evaluations (no reading; a model judges it probably moderate)
Holding

In consumer AI the rent goes to whoever owns the daily habit, and users keep most of the value

Consumer AI is the assistants and apps people use for themselves: OpenAI's ChatGPT, Meta AI inside Meta's apps, Google's Gemini in its app and in Search, Perplexity's answer engine, and companion apps such as Character.AI and Replika, which act as a friend or partner. Many people use them free; some pay a subscription, and advertising may pay for the rest. David Ricardo, an English economist who asked in 1817 why landlords grew rich when bread grew dear, found that the owner of the scarcest input collects a rent: profit above what it takes to keep a business going, earned because something holds rivals off. This site judges that in consumer AI the scarce input is a user's habit, the app someone opens without thinking, and that it is held by firms that already own people's phones, feeds and searches, plus ChatGPT, which built one of its own. A newer kind of consumer AI, the personal agent (an assistant that carries out tasks for its user, such as booking, shopping, filling in forms and paying, rather than only answering), adds a second question beside the habit: who stands where the user pays.

Getting cheap and interchangeable
This site judges that answering an everyday question is a commodity: many free assistants do it well enough, and the price to the user is zero. William Nordhaus, an economist who measured how much of the value of new inventions their makers keep as profit, found that it is a small part; most goes to users as lower prices and better products. On his finding, the gap between what users would pay and what they are charged is the value they keep. The rival view holds that the labs are durable, high-margin businesses whose customers mostly do not switch, which would leave more of the value with the makers.
Staying scarce
This site judges that what stays scarce is the user's attention and habit, the personal history an assistant builds up about each user, and the payment and advertising systems that turn attention into money. David Teece, an economist who studied why the firm that invents something often fails to keep the returns, called those inputs complementary assets; Meta and Google held theirs before AI arrived. Hamilton Helmer, a strategy consultant who studied why some firms keep earning more than their investors could expect elsewhere, found that a lasting advantage is a benefit behind a barrier. Network economies (each user makes the product more useful to the others) and scale economies hold in the platforms' existing businesses, but this site finds none proven for AI assistants themselves, where one user's chat does not make another's better and people use several at once. This site judges that agents which act, not just answer, could move both the habit and the point of payment: if people hand their errands to one agent, the firm running it stands between them and every shop they buy from, and could take a cut of each purchase, as card networks and app stores do, or charge sellers to be chosen. Whether that rent goes to the platforms that already own the habit, such as Meta with Muse and Manus, or to newcomers such as Instinct is open; one tester's early rankings are not evidence of lasting use. Because Meta says Muse keeps users' conversations with it, and the data in their Muse workspace, out of its ad systems, this site judges that, for now, the source suggests any money Muse makes would come from subscriptions or purchases rather than advertising.
What would make it last
An assistant's memory of each user becomes costly to leave, turning a habit into switching costs in Helmer's sense (what a user would lose by moving to a rival); or assistants learn what people want to buy early enough that advertisers pay more to reach them than in search, turning attention into the kind of rent search earned. It fails to convert if memories become easy to carry from one assistant to another. A third route opens with agents that buy for their users: the agent becomes the place where payment happens, and its owner takes a fee on each purchase, a rent on the sale rather than on attention. That fails to convert if users can move their saved cards, logins and preferences to a rival agent easily, or if shops and payment firms such as Stripe keep the fee.
What would prove this wrong
A consumer AI app with no existing platform behind it, such as a personal agent like Instinct, taking a large, paying audience from Meta, Google and ChatGPT and holding it; or consumer AI firms' revenue growing as fast as the value users say they get.
Binds on (how tight)
Applications (no reading; a model judges it probably slack), Running models (easing), Deployment platforms (no reading; a model judges it probably easing)
Holding

A rollup keeps AI's saving only where the firms it buys are paid for results rather than hours; its asset is the customer, and the AI is bought in

A rollup buys many small firms in one trade, such as accounting, property management or home services, and runs them as one. The groups this site files as AI rollups, among them Thrive Holdings, Long Lake, Crete, Dwelly and Titan, and funds such as General Catalyst that have set money aside to build them, plan to run the firms they buy with AI and keep the saving. David Ricardo, an English economist who asked in 1817 why landlords grew rich when bread grew dear, found that the owner of the scarcest input collects a rent: profit above what it takes to keep a business going, earned because something holds rivals off. Ronald Coase, an economist who studied why some work is done inside firms rather than bought, and Oliver Williamson, who studied contracts and found that whoever depends on an asset owned by someone else can be squeezed by its owner, would read a rollup as a bet that owning the customer, the licence and the trust is worth more than selling software to the firm that owns them. This site judges that the bet pays only where the firm's prices do not fall when its costs do.

Getting cheap and interchangeable
This site judges that the AI a rollup uses is bought from the labs and software firms like any other customer's, so it gives no lasting edge over a rival firm that buys the same tools. William Nordhaus, an economist who measured how much of the value of new inventions their makers keep as profit, found that it is a small part; most goes to buyers. In a trade where clients re-bid their contracts, a saving becomes a lower price once rivals match it. The rival view holds that AI lets one centre see and run many unrelated businesses, an edge that would outlast the tools themselves.
Staying scarce
This site judges that what stays scarce is the customer relationship, a licence or professional's signature where the trade needs one, and prices that are set by something other than the firm's costs. David Teece, an economist who studied why the firm that invents something often fails to keep the returns, called such inputs complementary assets and found that their owners often keep the profit; the rollup buys them rather than building them. Hamilton Helmer, a strategy consultant who studied why some firms keep earning more than their investors could expect elsewhere, found that a lasting advantage is a benefit behind a barrier; this site finds none proven for any AI rollup yet.
What would make it last
A rollup turns its running of many firms into a way of working rivals take years to learn, what Helmer calls process power, or grows large enough in one trade and place that its costs per client fall below any single firm's, his scale economies. Both need the saving to outlast the next time clients re-bid, which in trades billed by the hour means moving clients to fixed fees first.
What would prove this wrong
AI rollups in trades billed by the hour keeping higher margins for years after buying, without moving clients to fixed fees; or rollups' margins falling back to their trades' ordinary level as rivals that did not merge buy the same AI tools.
Binds on (how tight)
Applications (no reading; a model judges it probably slack), Agents and tool use (easing), Capital (slack)

The companies that matter here

Already worth over a trillion dollars

Meta Platforms

Show the profile
Market
Meta runs Facebook, Instagram, Messenger, WhatsApp and Threads, and its 10-K (the annual report US companies file with their regulator) says it earns substantially all its revenue from advertising on Facebook and Instagram. Its AI work sits on both sides of that business. It builds its own models. It released the Llama family openly, but its newest flagship, Muse Spark, released in April 2026, is its first frontier model without open weights; the 10-K says it expects to train a mix of open and closed models from here (an open model is one whose trained settings are published so anyone can run it). It offers Meta AI, an assistant available inside its apps, as a stand-alone app, on its AI glasses and on the web. And it uses AI to choose which posts and ads each person sees, which the 10-K credits with better ad targeting and measurement. It states a goal of building superintelligence, which it defines as AI that surpasses human intelligence. To pay for this it spends heavily on data centres: this site's records from its filings show its capital spending (money spent on buildings and machines) growing much faster than its revenue in 2025 and again so far in 2026, and its 10-K lists large commitments, mostly for data centres and for computing capacity rented from other firms. In October 2025 it entered an arrangement to co-develop a data-centre campus in Louisiana, in which it holds a minority interest. In June 2025 it bought a non-voting minority stake in Scale AI, a seller of training data made by people; its filings say it has no significant influence over Scale's operations, and Scale's founder joined Meta to work on its AI. This site's records also hold Meta agreeing to buy Manus and buying Limitless in December 2025 and Stilla in September 2026, and buying or hiring teams from other small AI firms in March 2026. Manus, maker of a general-purpose AI agent (software that carries out tasks such as research and building things for its user, rather than only answering), announced in December 2025 that it was joining Meta, but in April 2026 China's top economic planner, which reviews foreign investment, blocked the deal and ordered it withdrawn. In September 2026 Meta launched Muse, a personal agent that Meta says runs in its own walled-off computer in Meta's cloud with its own web browser, books travel, fills in forms and buys things after checking with the user, pays through Link, a checkout service built by Stripe, is free for most uses with paid plans for more, and does not share users' conversations with Muse, or the data in their Muse workspace, with Meta's ad systems. The question is whether AI adds to or eats into the advertising business that pays for it.
Customers (diffuse)
Advertisers pay for substantially all of it, the 10-K says. It also says most of its revenue comes from marketers advertising on Facebook and Instagram, that they make no long-term commitments, and that the loss of marketers or cuts in their spending could seriously harm the business. It names no large customer. The users it reaches are counted in billions of daily visitors, a figure Meta measures itself and warns may be inaccurate. Meta AI earns no direct revenue that the 10-K reports; in September 2026 Meta began selling its models and agents to businesses and developers through a new line, Meta Enterprise Platform. This site judges the paying base is diffuse, spread across many advertisers.
Powers
  • network economies: Hamilton Helmer, a strategy consultant who studied why some firms keep earning more than their rivals, found that a lasting advantage needs both a benefit and a barrier to copying. In Meta's social apps the benefit is that each user makes the apps more useful to friends and more valuable to advertisers; the barrier is that a rival must assemble a comparable audience first. The 10-K shows the audience still growing in 2025, with more ads shown and a higher average price per ad, so this site counts the power proven in its social apps. It finds no evidence yet that the power extends to Meta AI itself, where one person's chat does not improve another's.
  • scale economies: Emerging, not proven. This site judges that the cost of training models and building ad-ranking systems is spread over Meta's very large audience and ad sales, so each user costs it less than a smaller rival; but the barrier holds only against newcomers, not against Google or the frontier labs (the firms building the most capable AI systems), which spend at similar scale and are its main AI rivals.
  • counter-positioning: Emerging, not proven, and weakening. Counter-positioning is a newcomer's model the incumbent could copy only by damaging its own business. The benefit is that Meta can give AI away inside its apps and let advertising pay; the barrier is that a lab living on subscriptions and paid access to its models would give up that revenue by doing the same. But the labs can add advertising too, and in September 2026 Meta itself began selling its models and agents to businesses and developers through Meta Enterprise Platform, on top of paid Muse plans, the same paid access a lab lives on; so this site does not count the barrier as proven.
Under the site's rent rule
a fat profit, kept by the company itself. Depends on: AI chips (moderate), Data-centre buildings (tight), Electricity generation (tight), Building frontier models (easing), Post-training (no reading; a model judges it probably moderate), Running models (easing), Applications (no reading; a model judges it probably slack).

What would have to hold by 2035

  • Advertising keeps paying for the AI build-out. The 10-K says its AI work will need more spending on infrastructure and staff, and this site's records show capital spending rising far faster than revenue; ad revenue must keep growing fast enough to carry those costs and the losses of Reality Labs, its glasses and headsets division, which the 10-K says depends on profits earned elsewhere. In the second quarter of 2026 its operating income fell from a year earlier even as ad revenue grew.
  • AI keeps making its ads work better, so advertisers keep paying more per ad; the 10-K reports more ads shown and a higher average price per ad in 2025, and credits its AI with better targeting and measurement.
  • People keep turning to Meta AI inside apps they already use rather than moving their questions, shopping and time to assistants from OpenAI or Google. William Nordhaus, an economist who measured how much of the value of new inventions their makers keep, found that most goes to users; this site judges a free assistant only pays Meta if the attention it holds can be sold to advertisers.
  • Its own models stay close enough to the best to power its products, so it need not pay rival labs for the capability; the 10-K lists competitors whose AI may be better or cheaper among its risks, and its latest 10-Q already lists rising third-party AI token costs, spending on other providers' models, among the reasons its costs grew.
  • Regulators let it keep showing personalised ads. The 10-K says the European Commission ruled in April 2025 that its offer of a paid subscription with no ads did not comply with EU rules, a decision it has appealed.
  • Its minority stake in Scale AI, which carries no vote, and its purchases of small AI firms add skills and products without drawing new limits from competition authorities, whose past scrutiny of its acquisitions the 10-K describes.
  • People let Muse act for them, booking and buying rather than only chatting, and Meta earns from that work through subscriptions or a share of what Muse buys, since Meta says Muse keeps users' conversations with it, and the data in their Muse workspace, out of its ad systems; for now, the source suggests, that leaves subscriptions and purchases rather than advertising. This site judges that an agent shopping for its user could draw spending away from the ads that pay Meta today, so the agent must earn at least what it displaces.
  • Meta's own agents do not depend on buying agent makers that regulators can block; China blocked its purchase of Manus, the agent maker it agreed to buy in December 2025, and ordered the deal withdrawn.

What would show it failing

  • Capital spending rising while advertising revenue growth slows for several quarters in a row.
  • Meta AI's use growing without any rise in ad prices, ads shown or time spent in its apps.
  • Meta licensing another lab's model to run its main products, a sign its own models fell behind.
  • A regulator's order forcing it to show much less personalised ads across Europe or the United States.
  • Its data-centre venture partners or cloud suppliers renegotiating terms because Meta cut its capacity plans.
  • People using Muse to shop and book while time spent and ads shown in Meta's apps fall, with no subscription or purchase revenue from Muse making up the gap.
  • A regulator blocking another of its AI purchases, a sign that buying agent makers cannot fill gaps in Meta's own agents.

Drafted by claude-opus-5.5 from Meta Platforms announcement, 29 Dec 2025 (opens in a new tab), Meta Platforms annual report (10-K), 29 Jan 2026 (opens in a new tab), Meta Platforms quarterly report (10-Q), 30 Jul 2026 (opens in a new tab), Meta Platforms announcement, 28 Sep 2026 (opens in a new tab), TechCrunch article, 27 Apr 2026 (opens in a new tab), Meta Platforms announcement, 8 Sep 2026 (opens in a new tab), The Decoder article, 8 Apr 2026 (opens in a new tab), Meta Platforms quarterly report (10-Q), 31 Jul 2025 (opens in a new tab), Scale AI announcement, 12 Jun 2025 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it. Drafted by a Claude model, made by Anthropic, which competes with, sells to or buys from companies in this layer.

Candidate to be worth a trillion dollars

Palantir

Show the profile
Market
Palantir sells software platforms that connect a large organisation's data to its daily operations, and its 10-K (the annual report US companies file with their regulator) names four: Gotham, Foundry, Apollo and AIP. AIP, its AI platform, connects customers securely to large language models (AI systems trained on huge amounts of text) made by other firms, and gives them tools to build agents (AI programs that carry out tasks on their own) on top. At the heart of its software is what it calls the Ontology, a map of how a customer's data relates to its real operations. It runs short workshops, which it calls bootcamps, that build working tools on a customer's own data within days, and it says it embeds its staff directly with customers. It warns that pilots given at little or no cost may never turn into paying contracts. Government customers brought in most of its revenue in 2025 and commercial customers the rest, and its revenue from US commercial customers rose steeply that year. How big the market becomes depends on whether fitting AI to a firm's operations stays a job for a software platform like Palantir's, or passes to the labs (the firms building the most capable AI systems), which have set up deployment arms of their own.
Customers (moderate)
Governments and large companies. The 10-K says no single customer reached the threshold at which the filing reports a customer's share of revenue in any year from 2023 to 2025, and that the average revenue from its twenty largest customers rose in 2025 over 2024. But its risk factors warn that a limited number of customers account for a substantial portion of its revenue, because existing customers have widened their use, and its notes show one customer owing it a large share of its receivables (money customers owe for work already billed) at the end of both 2024 and 2025. Many of its contracts can be ended by the customer at will, and the US federal government cannot take up contract options more than a year ahead. This site judges the base is moderately concentrated: no single buyer reaches the reporting threshold, but a limited number of customers, led by governments, carry much of it.
Powers
  • switching costs: Emerging, not proven. Hamilton Helmer, a strategy consultant who studied why some firms keep earning more than their rivals, found that a lasting advantage needs both a benefit and a barrier to copying. The benefit is repeat, widening business from customers whose operations come to run on its Ontology; the barrier is the work a customer would have to redo to rebuild that map elsewhere. The 10-K reports customers expanding but gives no figure for how many leave, so this site does not count the barrier as proven; under its written rules that makes the rent a scarcity rent and its durability short until a filing shows customers staying.
Under the site's rent rule
a thin profit, kept by the company itself. Depends on: Running models (easing), Agents and tool use (easing), Deployment platforms (no reading; a model judges it probably easing), Applications (no reading; a model judges it probably slack).

What would have to hold by 2035

  • Its platform, not the model, stays where customers' work runs. AIP uses models made by others; this site judges that helps Palantir while the models are interchangeable, and hurts it if one lab's model becomes the product customers ask for by name.
  • The labs' own deployment arms do not take its commercial customers. This site's records hold the OpenAI Deployment Company agreeing to buy Tomoro in May 2026 and Northslope in July 2026, both firms that put AI to work inside large companies, and Poolside, a lab, buying Fern Labs, a deployment firm, in November 2025. David Teece, an economist who studied why inventors often fail to keep the returns, found that profit tends to go to whoever holds the assets a product needs to reach buyers; here that is the customer relationship and the map of its data, which this site judges Palantir holds today for its own customers, and which the labs' deployment arms are trying to win.
  • Its margins hold as it sends more staff into customers. Its contribution margin, its own measure of revenue left after the cost of serving and winning customers, rose in both its government and commercial segments in 2025. This site judges that a business that must grow its engineers with each customer earns a services margin (the thinner margin of a firm that sells its staff's time) rather than a software margin (the wider one of a firm that sells the same code again at little extra cost).
  • Its short pilots and bootcamps keep turning into long, paid contracts, which the 10-K warns may not happen.
  • Governments and its other largest customers keep renewing their contracts, although many can be ended at will, budgets can stall, and the 10-K warns that a limited number of customers carry a substantial portion of its revenue.

What would show it failing

  • Large commercial customers moving their AI work to a lab's deployment arm or a lab's own platform.
  • Its contribution margin falling as the staff it embeds with customers grows.
  • Revenue from existing customers flattening while pilots fail to convert, which the 10-K names as a risk.
  • A large government customer ending contracts for convenience or letting options lapse.

Drafted by claude-opus-5.5 from Palantir annual report (10-K), 17 Feb 2026 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it. Drafted by a Claude model, made by Anthropic, which competes with, sells to or buys from companies in this layer.

Candidate to be worth a trillion dollars

Cognition

Show the profile
Market
Cognition makes Devin, a coding agent: an AI tool that writes, tests and fixes software on its own, working from a task an engineer sets. It also owns Windsurf, a code editor (the program in which developers write and read code), which this site's records show it agreed to buy in July 2025. It calls itself an independent agent lab: it says it works with all the major labs (the firms that build the most capable AI systems) and uses the model best suited to each task, and it has begun training models of its own, which it says customers choose for cost and speed. It says its revenue run-rate (a recent month's revenue scaled up to a year) grew sharply between May and September 2026; that figure is self-reported, from its own posts, not audited. It has opened offices in several countries to work beside customers. How big the market becomes depends on whether firms keep paying an independent agent maker once the labs sell their own coding agents directly.
Customers (moderate)
Large companies and government. It names Citi, Mercedes-Benz, Goldman Sachs, Santander, Dell, NVIDIA, GE Aerospace, the US Army and the US Navy, and says systems integrators (firms that assemble technology for clients), among them Infosys and Cognizant, build Devin into the work they deliver. It gives no revenue by customer. This site judges the base is moderately concentrated in large firms.
Powers
  • switching costs: Emerging, not proven. Hamilton Helmer, a strategy consultant who studied why some firms keep earning more than their rivals, found that a lasting advantage needs both a benefit and a barrier to copying. The benefit is repeat business as Devin is wired into a firm's code, tools and approval steps; the barrier is the setup a customer would redo to move. But a rival agent can be tried alongside it, Cognition itself offers every lab's model, and it publishes no figures on customers leaving, so this site does not count the barrier as proven.
Under the site's rent rule
a thin profit, kept by platforms. Depends on: Running models (easing), Post-training (no reading; a model judges it probably moderate), Agents and tool use (easing), Applications (no reading; a model judges it probably slack), Capital (slack).

What would have to hold by 2035

  • Being independent of any one lab stays worth paying for. Its pitch is that customers get the best model for each task; this site judges that holds while the labs' models differ task by task, and weakens if one lab's agent is best at most coding work.
  • Its own models carry a growing share of its work, so that less of each sale goes to buying tokens (the units of text a model reads and writes, which the labs charge for) from the labs whose agents compete with it, among them OpenAI's Codex and Anthropic's Claude Code.
  • Its work inside large firms, where it sits beside engineering teams and systems integrators, holds customers better than a lab selling the same agent directly. David Teece, an economist who studied why inventors often fail to keep the returns, found that profit tends to go to whoever holds the assets a product needs to reach buyers. Here the labs hold the model and Cognition holds the enterprise relationship; this site judges the model the harder of the two to replace, so the labs, not Cognition, hold the asset that decides who keeps the gain.
  • Its reported growth is real revenue that renews; its run-rate is self-reported, and a run-rate built on a few fast months can fall as fast.

What would show it failing

  • Large customers replacing Devin with a lab's own coding agent after a new model release.
  • Price cuts on the labs' own agents followed by Cognition cutting its prices or its margin.
  • Its own models losing use inside Devin to models bought from the labs.
  • A lab buying a rival agent maker and bundling it with its models for large firms.

Drafted by claude-opus-5.5 from Anthropic announcement, 12 Feb 2026 (opens in a new tab), Cognition announcement, 27 May 2026 (opens in a new tab), Cognition announcement, 8 Sep 2026 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it. Drafted by a Claude model, made by Anthropic, which competes with, sells to or buys from companies in this layer.

Candidate to be worth a trillion dollars

Anysphere

Show the profile
Market
Anysphere makes Cursor, a code editor (the program in which developers write and read code) built around AI that suggests code, answers questions about a codebase and runs agents (AI programs that carry out whole tasks on their own). It is now part of SpaceX, the rocket and satellite company that also owns xAI, a lab building frontier models (the most capable AI systems of their time). SpaceX's 10-Q (the quarterly report US companies file with their regulator) says it took an option to buy Cursor in April 2026, exercised it in June 2026 and signed a merger agreement paying in its own shares, subject to regulators' approval; it also says that alongside the option it agreed to give Cursor computing capacity and to work with it on improving xAI's Grok models and developing new ones jointly. Cursor's own post says the purchase was completed in August 2026. Cursor says the deal gives it access to what it calls the largest fleet of AI chips in the world, to train stronger models that cost less to run, and that same month it released a Grok model, the family xAI builds, made with SpaceX. Before the deal it said in November 2025 that its in-house models generated more code than almost any other model in the world, and that its revenue had passed a large milestone; press reports in this site's records, citing people familiar with the company, show its run-rate (a recent month's revenue scaled up to a year) rising sharply between February and June 2026. These figures are self-reported or sourced to people close to it, not audited. Because Cursor is now a business line inside SpaceX, the question of a trillion-dollar company applies to SpaceX and xAI as a whole; this profile judges Cursor as a business line whose fate depends on the conditions in this site's xAI profile.
Customers (diffuse)
Individual developers and engineering teams; it says millions of developers and many large engineering organisations use it. It gives no revenue by customer. This site judges the base is diffuse, though it cannot tell how much of the revenue large companies bring.
Powers
  • switching costs: Emerging, not proven. Hamilton Helmer, a strategy consultant who studied why some firms keep earning more than their rivals, found that a lasting advantage needs both a benefit and a barrier to copying. The benefit is the daily habit of developers who live in its editor and the settings and history a team builds up there; this site judges the barrier weak, because rivals offer much the same and a developer can try another editor alongside it, and it has no figures on users leaving.
Under the site's rent rule
a thin profit, kept by the company itself. Depends on: AI chips (moderate), Data-centre buildings (tight), Post-training (no reading; a model judges it probably moderate), Running models (easing), Agents and tool use (easing), Applications (no reading; a model judges it probably slack).

What would have to hold by 2035

  • The conditions in this site's xAI profile hold, since Cursor's owner now carries xAI's data-centre costs, legal cases and losses, and the value of Cursor to its owner rests on them.
  • Owning its models and chips lowers what it pays per task below what rivals pay the labs, so it can cut prices or keep a margin that an agent renting models from a lab cannot. That is the point of the merger in Cursor's own words; this site has no figures yet that show it. David Teece, an economist who studied why inventors often fail to keep the returns, found that profit tends to go to whoever holds the assets a product needs to reach buyers; because Cursor's owner now holds the chips and co-develops the models, this site judges that Cursor sits with the holder of those assets, unlike Cognition and Lovable, which rent their models from the labs.
  • Its record of how developers accept or reject its suggestions keeps improving its models. This site's records hold Cursor describing, in September 2025, how it rolls out new models to users through the day and trains on what happens. Andrei Hagiu and Julian Wright, economists who set out tests for when learning from customers' data keeps rivals out, found that it does so only when the data is proprietary, keeps its value and improves the product for every customer; this record may pass the first and last, but goes stale as models improve.
  • Customers who want other labs' models stay after Cursor's owner became a rival lab, and this site judges those labs must keep selling Cursor their models on fair terms for that to hold.

What would show it failing

  • Large customers moving to the labs' own coding agents after the SpaceX purchase.
  • Rival labs withdrawing their best models from Cursor, or Cursor offering only xAI's.
  • Cursor's models falling behind the labs' coding models on the tasks its users care about.
  • SpaceX's filings showing the AI segment's losses widening without Cursor's revenue growing.

Drafted by claude-opus-5.5 from Anysphere announcement, 14 Aug 2026 (opens in a new tab), Anysphere announcement, 13 Nov 2025 (opens in a new tab), SpaceX quarterly report (10-Q), 4 Aug 2026 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it. Drafted by a Claude model, made by Anthropic, which competes with, sells to or buys from companies in this layer.

Candidate to be worth a trillion dollars

Lovable

Show the profile
Market
Lovable lets people who cannot code build websites and apps by describing what they want to an AI in plain words, and then host them (run them on its own servers so others can use them); it has added hosting, databases, logins and payments so the result can be a working business, not only a mock-up. It says people have created a very large number of projects since it launched in November 2024 and that its users' apps draw a very large number of visits each month. It says it draws on several models, matching each part of a job to the one best suited, and post-trains open-source models (further trains freely published models on its own data). Its Form D (the notice a private company files with the US regulator when it sells shares without a public offering) shows a Delaware company formed in 2023 that had sold most of the shares offered in its mid-2026 round without paying any placement agent (a bank paid to find investors), and that declined to disclose its revenue. Its revenue figures in this site's records come from the company itself and rose sharply between mid-2025 and mid-2026; they are not audited. How big the market becomes depends on how many people and firms start making their own software, and whether the labs (the firms building the most capable AI systems) let a middleman keep that work.
Customers (diffuse)
Founders, small businesses and teams inside large companies; it names Adidas, NVIDIA, Deutsche Telekom, Zendesk, Klarna and Checkr among the firms using it, and says its reach inside large US companies has grown. It gives no revenue by customer. This site judges the base is diffuse, spread across many small subscribers and a growing number of company accounts.
Powers
  • switching costs: Emerging, not proven. Hamilton Helmer, a strategy consultant who studied why some firms keep earning more than their rivals, found that a lasting advantage needs both a benefit and a barrier to copying. The benefit is that a business built and hosted on Lovable, with its data, logins and payments there, keeps paying; the barrier is the work of moving a running business elsewhere. Lovable publishes no figures on users leaving, so this site does not count the barrier as proven.
Under the site's rent rule
a thin profit, kept by platforms. Depends on: Running models (easing), Post-training (no reading; a model judges it probably moderate), Agents and tool use (easing), Applications (no reading; a model judges it probably slack), Capital (slack).

What would have to hold by 2035

  • Apps built on it stay on it. Once a business runs there, Lovable earns from hosting it, not only from building it; this site judges that is what could turn a tool bought for a project into a service kept for years.
  • Its own models, post-trained from open ones, carry more of the work, so that less of each sale goes to buying tokens (the units of text a model reads and writes, which the labs charge for) from the labs.
  • Its data on which apps succeed improves the product for all users, as it says. Andrei Hagiu and Julian Wright, economists who set out tests for when learning from customers' data keeps rivals out, found that it does so only when the data is proprietary, keeps its value and improves the product for every customer; Lovable's claim passes the last test on its own account, and this site has no way yet to check it.
  • The labs do not sell the same thing to the same people. This site judges that a lab could add app building and hosting to its own chat product, and Lovable must stay easier or more complete than that to keep its users.
  • Its growth is real revenue that renews; its figures are self-reported, and many of its users are people trying an idea, who may not stay.

What would show it failing

  • Paying users leaving in large numbers after a lab adds app building and hosting to its chat product.
  • Its prices falling toward the cost of the tokens it buys.
  • Firms moving apps built on Lovable to their own systems or a cloud once they grow.
  • Its own models failing to take work from the labs' models inside its product.

Drafted by claude-opus-5.5 from Lovable Form D, 17 Aug 2026 (opens in a new tab), Lovable announcement, 18 Dec 2025 (opens in a new tab), Lovable announcement, 12 Aug 2026 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it. Drafted by a Claude model, made by Anthropic, which competes with, sells to or buys from companies in this layer.

Candidate to be worth a trillion dollars

Harvey

Show the profile
Market
Harvey sells AI software for lawyers: tools to research, draft, review contracts and documents, and store a firm's files, plus agents (AI programs that carry out multi-step work on their own) that customers build and run on it. It says customers had built a very large number of custom agents on it by March 2026, and that its legal engineers (staff who combine legal training with technical skill) work beside customers to build and improve them. In August 2026 this site's records hold Harvey saying it had worked with Fireworks, a firm that runs and tunes models, to post-train (further train for legal work) a model starting from an open one, as a research preview; by September 2026 it described this as its first post-trained open-weight model (one whose trained settings are published). It also published its own test of legal work, in which, by its account in May 2026, the best general models finished few tasks end to end. Its Form D (the notice a private company files with the US regulator when it sells shares without a public offering) shows a Delaware company formed in 2022 that sold shares from late August 2026 without paying any placement agent (a bank paid to find investors) and declined to disclose its revenue. Its annual recurring revenue (what customers have signed up to pay in a year) grew sharply through 2026 according to its co-founder and press reports; those figures are self-reported, not audited. How big the market becomes depends on whether law firms pay for a legal specialist once general models learn the law.
Customers (moderate)
Law firms and companies' in-house legal teams. It says most of the largest US law firms by revenue and many of the world's biggest companies use it, and names NBCUniversal, HSBC, DLA Piper International, Corrs Chambers Westgarth and McCann FitzGerald among recent customers. OpenAI's venture fund is among its investors. It gives no revenue by customer. This site judges the base is moderately concentrated in large firms.
Powers
  • switching costs: Emerging, not proven. Hamilton Helmer, a strategy consultant who studied why some firms keep earning more than their rivals, found that a lasting advantage needs both a benefit and a barrier to copying. The benefit is repeat business as firms store their files on Harvey and run their own agents there; the barrier is the work of rebuilding those agents and moving those files. Harvey's own pitch is that firms own their intelligence, which lowers the barrier, and it publishes no figures on customers leaving, so this site does not count it proven.
Under the site's rent rule
a thin profit, kept by incumbent firms. Depends on: Running models (easing), Post-training (no reading; a model judges it probably moderate), Tests and evaluations (no reading; a model judges it probably moderate), Training data (no reading; a model judges it probably moderate), Applications (no reading; a model judges it probably slack).

What would have to hold by 2035

  • Law firms keep paying a legal specialist rather than using a general model directly. Anthropic said early in 2026 that its Cowork product has plugins (add-ons that give its model the tools and instructions for one job) for roles such as legal work, and Microsoft's 10-K (the annual report US companies file with their regulator) describes selling its Copilot assistant and agents to businesses for gains in productivity; Harvey must stay better at a firm's real work than general tools like these.
  • Its legal test shows what it claims and others confirm it: that general models still fail most full legal tasks. It is a vendor's test, and this site has seen no one else repeat it.
  • The firms' own records, which Harvey stores but does not own, keep it inside their work. David Teece, an economist who studied why inventors often fail to keep the returns, found that profit tends to go to whoever holds the assets a product needs to reach its buyers; here the law firms hold their clients, their files and their reputations, so this site judges they, more than Harvey, are placed to keep the gain.
  • Its own models, post-trained from open ones, cut what it pays the labs for tokens (the units of text a model reads and writes), so its margin does not shrink each time a lab raises its price or sells a legal tool of its own.
  • Law firms keep buying tools that cut hours. This site judges that where a firm bills its clients by the hour, faster work lowers its bills, so it may pay Harvey less for the saving.

What would show it failing

  • Large law firms moving work from Harvey to a lab's own legal product after a new model release.
  • An independent test showing general models doing full legal tasks as well as Harvey's agents.
  • Harvey cutting prices as general models improve, or firms buying fewer seats (the per-user licences they pay for).
  • Firms taking their stored files and agents to a rival at little cost.

Drafted by claude-opus-5.5 from Anthropic announcement, 12 Feb 2026 (opens in a new tab), Harvey announcement, 25 Mar 2026 (opens in a new tab), Harvey announcement, 9 Sep 2026 (opens in a new tab), Harvey Form D, 11 Sep 2026 (opens in a new tab), Microsoft Corp annual report (10-K), 29 Jul 2026 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it. Drafted by a Claude model, made by Anthropic, which competes with, sells to or buys from companies in this layer.

Candidate to be worth a trillion dollars

OpenEvidence

Show the profile
Market
OpenEvidence answers doctors' clinical questions with AI, citing medical research, and offers a research agent that reads many studies and writes a report. Its answers draw on content from partners it names, among them the New England Journal of Medicine, the American Medical Association's JAMA journals, the National Comprehensive Cancer Network and the Cochrane Library. It offers its research agent free to verified US clinicians, and neither of its press releases says who pays for the service. It calls itself the fastest-growing application for physicians in history and says, in March 2026, that most practising US physicians use it daily; these are its own descriptions, made in press releases it issued itself. This site's records hold the company saying its run-rate (a recent month's revenue scaled up to a year) passed a large milestone in 2025, and a research firm's estimate that it rose sharply by July 2026; neither is audited. How big the market becomes depends on whether a free tool with doctors' daily attention can charge someone for it without losing that trust.
Customers (diffuse)
Its users are doctors and other clinicians; the sources do not say that they pay, or who does, so this site cannot name its paying customers or say how concentrated they are. Alex Brogan judged the base diffuse, since it is spread across many clinicians and institutions.
Powers
  • a cornered resource: Emerging, not proven. Hamilton Helmer, a strategy consultant who studied why some firms keep earning more than their rivals, found that a lasting advantage needs both a benefit and a barrier to copying; a cornered resource is preferential access to something others need. The benefit is answers grounded in leading journals' full text; the barrier would be that rivals cannot get the same licences. The releases do not say the licences are exclusive, and the journals could license others, so this site does not count the barrier as proven.
Under the site's rent rule
a thin profit, kept by incumbent firms. Depends on: Running models (easing), Training data (no reading; a model judges it probably moderate), Tests and evaluations (no reading; a model judges it probably moderate), Applications (no reading; a model judges it probably slack).

What would have to hold by 2035

  • It finds a way to earn from doctors' attention that does not cost their trust; the releases do not say how it earns today.
  • Its journal partnerships stay open to it on terms it can afford, and are not matched by a lab. David Teece, an economist who studied why inventors often fail to keep the returns, found that profit tends to go to whoever holds the assets a product needs to reach buyers; here the journals own the content, so this site judges they can raise their price or sell to rivals as the content grows more valuable.
  • Doctors keep choosing a specialist over general assistants. Microsoft's 10-K (the annual report US companies file with their regulator) describes a cloud business, once called Nuance Healthcare, that sells AI to the health care industry; this site's records hold OpenAI buying Torch, a health-care technology start-up, in January 2026; and Anthropic said early in 2026 that it had widened its reach into health care. A general model grounded in the same research would compete directly.
  • Regulators keep treating it as a reference tool for doctors rather than a medical device needing approval, or it wins approval if they do not.
  • Its reported use and revenue are real and lasting; both rest on its own releases and on estimates.

What would show it failing

  • A leading journal licensing its full text to a lab or a rival clinical tool.
  • Doctors' use of OpenEvidence flattening after a lab launches a clinical product.
  • Its revenue model drawing complaints from doctors or regulators that it biases answers.
  • A regulator ruling that clinical-answer tools need approval as medical devices, without OpenEvidence holding one.

Drafted by claude-opus-5.5 from Anthropic announcement, 12 Feb 2026 (opens in a new tab), Microsoft Corp annual report (10-K), 29 Jul 2026 (opens in a new tab), OpenEvidence announcement, 12 Mar 2026 (opens in a new tab), OpenEvidence announcement, 15 Jul 2025 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it. Drafted by a Claude model, made by Anthropic, which competes with, sells to or buys from companies in this layer.

Already worth over a trillion dollars

Apple

Show the profile
Market
Apple designs and sells the iPhone, Mac, iPad, wearables and home devices, and sells services on them: the App Store, cloud storage, subscriptions, advertising, AppleCare and payments. Its 10-K (the annual report US companies file with their regulator), for the year to September 2025, says it designs nearly the entire solution for its products, from hardware and operating system to apps and services, and relies on outsourcing partners, mostly in Asia, to build them. The 10-K mentions artificial intelligence mainly among its risks: safety, copyright and regulation. Apple's machine-learning research site describes the models behind Apple Intelligence, its AI features: a small model that runs on the device itself, tuned for the chips Apple designs (Apple silicon), and a larger model that runs on Private Cloud Compute, Apple's own server platform, which it presents as part of its commitment to users' privacy. A software framework lets app developers build on these models. Apple says both models match or beat open models of similar size (models whose trained settings are published so anyone can run them); that is its own test, and a comparison with models of similar size, not with the frontier models (the most capable AI systems, from firms such as OpenAI and Google). The market map this site uses places Apple in edge inference: running a trained model on the user's own device rather than in a data centre. This site judges that Apple's AI market is not selling AI itself but keeping people on its devices and in its services: the 10-K shows its Services gross margin (the share of sales left after the direct cost of providing them) far above its Products gross margin, and rising again in 2025. How big this becomes depends on whether AI makes people buy new devices and stay in its services, or moves their daily habit to assistants that live in other companies' apps.
Customers (diffuse)
People and businesses that buy its devices, directly and through indirect channels such as mobile carriers and other resellers; the 10-K says most of its net sales in 2025 went through indirect channels, and that its carriers accounted for a large share of its trade receivables (money customers owe it). The United States is its largest market; its sales in Greater China (China mainland, Hong Kong and Taiwan) fell in 2025. It also earns revenue from licensing arrangements with Google and other companies to offer their search services on its devices, and a commission on some purchases made through its App Store. The 10-K names no single customer as a large share of its net sales; the only single-customer exposure it discloses is in trade receivables. David Teece, an economist who studied why the firm that invents something often fails to keep the returns, called the inputs a product needs to reach its buyers complementary assets, and found that whoever holds them often keeps the profit; this site judges that Apple holds its own: the devices, the chips it designs, the App Store and the habits of the people who carry them.
Powers
  • network economies: Emerging, not proven. Hamilton Helmer, a strategy consultant who studied why some firms keep earning more than their rivals, required each lasting advantage to have both a benefit and a barrier to copying. The 10-K describes the loop: customers' decisions to buy its hardware depend in part on third-party apps, and developers weigh Apple's market position and expected sales when deciding whether to build for it. It does not show the barrier: it says its minority share of the global smartphone, computer, tablet and wearables markets can push developers toward rivals with larger shares, and that some rivals have large bases of active devices of their own. This site judges that the benefit shows in a Services gross margin far above its Products margin, but the filing does not tie that margin to the loop. Under the site's written rules that makes Apple's rent (profit above what it would earn in a fully competitive market) a scarcity rent, one that lasts while an input is short or a firm is first, and its durability short until a filing shows the barrier. This site finds no evidence that the loop extends to AI features, where one person's use does not improve another's.
  • switching costs: Emerging, not proven. This site judges that an assistant which knows a user's messages, photos and habits on the device could make leaving costly. The barrier is untested: the 10-K reports no AI revenue or use, and assistants from other firms can run on the same phone.
Under the site's rent rule
a thin profit, kept by the company itself. Depends on: Chip foundries (moderate), Memory (no reading), Running models (easing), Applications (no reading; a model judges it probably slack), Deployment platforms (no reading; a model judges it probably easing).

What would have to hold by 2035

  • AI makes people keep buying and replacing its devices and keeps them in its services; the 10-K says competitors imitate its features and cut prices aggressively, and that Apple holds a minority share of the global markets for smartphones, personal computers, tablets and wearables.
  • Running AI on the device stays good enough for everyday requests, so that its own chips, rather than rented data-centre chips, carry most of the work; Apple compares its small model with open models of similar size, not with the frontier models.
  • When a request needs a stronger model than its own, Apple keeps the user inside its own interface and on its own terms, rather than users opening ChatGPT, Google's Gemini or another lab's app directly.
  • Its revenue from letting Google and others offer search on its devices survives the remedies a US court ordered against Google in September 2025; the 10-K says that order may still change in further proceedings or on appeal, and that a reversal on appeal could bring remedies barring Google from offering Apple commercial terms for search placement. This site judges it must also survive any shift of people's questions from search to AI assistants.
  • Courts and regulators leave it room to charge on its App Store; the 10-K describes a US court order barring it from charging a commission on certain purchases and a fine from the European Commission in April 2025 under the EU's Digital Markets Act. This site judges that if AI agents (assistants that act for their users, booking and buying) become where people pay, Apple keeps a share only while those purchases still pass through its devices and store.
  • Its suppliers keep delivering; the 10-K says it relies on single-source partners for many components, some of them custom parts available from only one source, and on partners in Asia for final assembly of substantially all its hardware.

What would show it failing

  • Its Services gross margin falling for several years in a row.
  • Device sales falling, or people keeping their phones longer, after its AI features launch.
  • Its own models falling so far behind that it hands most AI requests to another firm's model, with that firm's brand shown to the user and much of the value paid to it.
  • Appeals ending with Google barred from offering Apple commercial terms for search placement, or search revenue shrinking as people ask AI assistants instead.
  • Regulators forcing it to allow other app stores and payment systems widely enough that its store commission falls.

Drafted by claude-opus-5.5 from Apple annual report (10-K), 31 Oct 2025 (opens in a new tab), Apple announcement, 17 Jul 2025 (opens in a new tab); reviewed by separate reviewing runs of the same model, which checked its claims against the sources it cites; no person reviewed it. Drafted by a Claude model, made by Anthropic, which competes with, sells to or buys from companies in this layer.

Indicators

Application layer's share of enterprise generative-AI spend (Menlo)
emerging
51.4%as of 2025-12-31
Vertical versus horizontal application funding
unclear
2.03×as of 2026-09-30
Startups' share of enterprise AI application revenue
emerging
63.0%as of 2025-12-31
Application-company revenue run-rates
emerging
$4.0Bas of 2026-06-05
Developer productivity uplift (METR RCTs)
consistent with normal
-4.0%as of 2026-02-24
Enterprise pilots with measurable P&L impact
consistent with normal
5.0%as of 2025-06-30
FDA-authorised devices built on a large language model
consistent with normal
1as of 2025-12-23
Frontier models completing expert legal tasks end to end (Harvey LAB)
emerging
7.1%as of 2026-05-26
Waymo paid rides per week
consistent with normalstale since 2026-03-27
500,000 rides per weekas of 2026-03-27

Sub-layers and venture flow

  1. 18. Deployment / FDE / implementation services

    No primary rounds on file.

  2. 19. Talent production / AI education & FDE training

    No primary rounds on file.

  3. 20. Coding agents & dev tools

    Coding agents & dev tools: primary-round equity by quartereach segment links to its first round's record
    Q1 2025 · Form D filings · $20.2M · reportedQ2 2025 · Form D filings · $193M · reportedQ4 2025 · Form D filings · $425M · reportedQ3 2026 · Form D filings · $382M · reported
    Form D filings

    Derived (venture_dollars_by_source) from: SEC EDGAR Form D (private offerings). SEC EDGAR Form D data sets.

    The numbers
    quartersourcekinddollarsrecord
    Q3 2026Form D filingsequity$382Mreported
    Q4 2025Form D filingsequity$425Mreported
    Q2 2025Form D filingsequity$193Mreported
    Q1 2025Form D filingsequity$20.2Mreported

  4. 21. Vertical AI by domain

    Vertical AI by domain: primary-round equity by quartereach segment links to its first round's record
    Q1 2023 · Form D filings · $28.8M · reportedQ2 2023 · Form D filings · $25.8M · reportedQ4 2023 · Form D filings · $66.0M · reportedQ1 2024 · Form D filings · $141M · reportedQ2 2024 · Form D filings · $114M · reportedQ3 2024 · Form D filings · $99.0M · reportedQ4 2024 · Form D filings · $127M · reportedQ1 2025 · Form D filings · $300M · reportedQ2 2025 · Form D filings · $496M · reportedQ3 2025 · Form D filings · $81.4M · reportedQ4 2025 · Form D filings · $150M · reportedQ1 2026 · Form D filings · $692M · reportedQ2 2026 · Form D filings · $247M · reportedQ3 2026 · Form D filings · $550M · reported
    Form D filings

    Derived (venture_dollars_by_source) from: SEC EDGAR Form D (private offerings). SEC EDGAR Form D data sets.

    The numbers
    quartersourcekinddollarsrecord
    Q3 2026Form D filingsequity$550Mreported
    Q2 2026Form D filingsequity$247Mreported
    Q1 2026Form D filingsequity$692Mreported
    Q4 2025Form D filingsequity$150Mreported
    Q3 2025Form D filingsequity$81.4Mreported
    Q2 2025Form D filingsequity$496Mreported
    Q1 2025Form D filingsequity$300Mreported
    Q4 2024Form D filingsequity$127Mreported
    Q3 2024Form D filingsequity$99.0Mreported
    Q2 2024Form D filingsequity$114Mreported
    Q1 2024Form D filingsequity$141Mreported
    Q4 2023Form D filingsequity$66.0Mreported
    Q2 2023Form D filingsequity$25.8Mreported
    Q1 2023Form D filingsequity$28.8Mreported

  5. 22. Autonomous services / AI-native service firms

    No primary rounds on file.

  6. 23. Consumer AI

    No primary rounds on file.

  7. 24. AI-native rollups / buyout vehicles

    No primary rounds on file.

Primary-round dollars per calendar quarter for each sub-layer, each strip on its own scale: Form D amounts sold where the entity files, Epoch's press-compiled rounds otherwise.

Crosswalk to the diffusion lens

  • Products (same valve) — Uplift RCTs, pilot→production and app-layer margin net of inference measure the same valve.
  • Adaptation (leak from) — Deployment/FDE services capture part of the surplus that adaptation creates.
  • Adaptation (same valve) — Adaptation measured inside deployed products: regulated devices built on language models and paid robotaxi rides.