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
Expert-data sellers broker a skill the labs cannot yet make for themselves; what they sell belongs to the buyer, so their rent lasts only while the labs need people
Human-data sellers such as Scale AI, Surge AI, Mercor, Handshake, Turing and Invisible find experts (doctors, lawyers, programmers, scientists), pay them to write questions and answers, grade a model's attempts and set down how they do their work, and sell the result to the AI labs to train on. 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 expert time, found quickly and checked well, and that the seller is a broker of it: the work is made to one lab's order and becomes that lab's to keep. Kenneth Arrow, an economist who studied how markets handle knowledge, argued that information is hard to own because copying it is cheap; here the buyer, not the seller, ends up owning it. The leading sellers are now also selling downstream, to ordinary companies rather than labs. Ronald Coase, an economist who studied why some work is done inside firms rather than bought, and Oliver Williamson, an economist who studied contracts and found that whoever has built something useful to only one partner can be squeezed by that partner, explain why the labs pull this layer towards them. This site judges that the move to ordinary companies is the same logic run in reverse: a supplier integrating forward (taking on the next step of the work, closer to the final user) so that it is not at the mercy of a few concentrated buyers.
- Getting cheap and interchangeable
- This site judges that ordinary labelling, the tagging of pictures and text on which the first data firms were built, is a commodity: many sellers, interchangeable work, and prices set by the pay of the workers. 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. For the seller, expert data made to one lab's order fails the first and last of those tests, since the lab owns it and it helps only that lab, and it goes stale as each new model masters what it taught.
- Staying scarce
- This site judges that what stays scarce, for now, is the ability to find, vet and pay many experts in a narrow field quickly, and to check their work well enough that a lab trusts it; Handshake's account of its university network and Mercor's stress on its talent network, matching and speed are both claims to that. 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 here: the benefit of a large expert network is real, but experts can work for several sellers at once and the labs can hire them directly, so the barrier is not yet shown.
- What would make it last
- A seller's graded records, pooled across many buyers, become what Helmer calls a cornered resource (preferential access to something others need), such as checked outcomes of expert work that improve every buyer's model; or a seller becomes so built into a lab's training that switching costs (what a customer would lose by moving to a rival) hold it. Coase and Williamson would expect the labs to bring such work inside as it becomes specific to them. Meta's stake in Scale AI is the largest step towards it this site records, but it stops short: the stake carries no votes, and Meta's filing says it gives Meta no significant influence over Scale. The move to ordinary companies is a second route: if tests and graded records built for one firm improve the service for the next, they pass the tests of Hagiu and Wright and would compound; if each job stays made to one client's order, the sellers have swapped a few lab buyers for many consulting clients, paid for their people's time, and the rent stays thin.
- What would prove this wrong
- A seller keeping or widening its margin (what it keeps from each sale after paying its costs) as more sellers enter and the labs hire experts directly, because something it owns, a pool of experts no rival can reach or a record of graded work, grows more valuable with every lab it serves; or a seller earning a lasting margin from ordinary companies on tests and graded records that carry from one client to the next, which would show the rent outlasting the labs' need for people.
- Binds on (how tight)
- Training data (no reading; a model judges it probably moderate), Post-training (no reading; a model judges it probably moderate), Tests and evaluations (no reading; a model judges it probably moderate)
Holding
Practice tasks for training are made to order for a few labs; they hold value only where the grader checks work that is hard to check
Reinforcement learning trains a model by trial and error: it attempts a task, a grader (a program or a person that scores the attempt) marks it, and the model is nudged towards what scored well. An environment is the practice ground: a copy of a website, a codebase, a spreadsheet or a job, with tasks set inside it and a grader attached. Sellers such as Mechanize, Fleet, Halluminate, Matrices, Applied Compute and Plato build them for the labs. 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 building the practice ground is ordinary software work, so any rent sits in the grader: tasks whose right answer only an expert can judge are scarce, and tasks a machine can check are not.
- Getting cheap and interchangeable
- This site judges that the practice ground itself, a copied website or a sandbox where code runs, is becoming a commodity: open-source tools and the clouds' own services supply it, and a lab can build one in-house. 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. An environment built to one lab's order belongs to that lab, helps only that lab, and goes stale once models master its tasks, since a task every model passes teaches nothing.
- Staying scarce
- This site judges that graders for work whose outcome is hard to judge, such as a legal filing, a clinical call or a long engineering job, stay scarce, together with the experts who can write 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 among environment makers, whose tasks can be rebuilt by a buyer that has seen them.
- What would make it last
- A maker holds graders for a field every lab needs and cannot build, which Helmer calls a cornered resource (preferential access to something others need); or it collects graded outcomes from real work inside firms, which improve with every buyer and go stale slowly. Ronald Coase, an economist who studied why some work is done inside firms rather than bought, and Oliver Williamson, an economist who studied contracts and found that a buyer that depends on an asset built for it can be squeezed by whoever owns it, would expect the labs and compute firms to buy makers whose graders become specific to their training, as the buying so far suggests.
- What would prove this wrong
- Models learning long work whose results are hard to check from practice they set and grade themselves, so the labs stop buying expert-graded environments; or environment makers keeping a lasting margin (what it keeps from each sale after paying its costs) on tasks a machine can already grade.
- Binds on (how tight)
- Post-training (no reading; a model judges it probably moderate), Training data (no reading; a model judges it probably moderate), Tests and evaluations (no reading; a model judges it probably moderate)
Commoditising
Tuning a model for one customer is becoming a feature of the clouds and the serving firms; a rent survives only where the tuned model keeps learning inside the customer's work
Post-training is everything done to a model after its first, general training, changing its weights (the settings, fixed in training, that hold what the model has learned): tuning it on a firm's own examples, training it by trial and error against graders, and, in its newest form, continual learning, in which the model keeps updating as it is used. Sellers include Thinking Machines Lab, Prime Intellect, Adaptive ML, Predibase, Arcee, Unsloth, Trajectory and Baseten. 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. 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 a model tuned to one firm's work is such an asset, but that the tuning itself is a service anyone with chips and open tools (software published for anyone to use and change for free) can offer.
- Getting cheap and interchangeable
- This site judges that a one-off tuning of an open model on a firm's examples is becoming a commodity: open-source tools such as Unsloth do it for free, the clouds and serving firms bundle it, and each new general model narrows the gain. 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.
- Staying scarce
- This site judges that the skill to run reinforcement learning at scale without damaging what the model already knew, and the record of a customer's own work that a continually learning model feeds on, stay scarce. On the tests of Andrei Hagiu and Julian Wright, economists who asked when learning from customers' data keeps rivals out, a customer's record helps only that customer, so it does not compound across customers; but it is specific in Williamson's sense. 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 among the tuning firms today.
- What would make it last
- Continual learning works in production at scale, so a model that has learned a firm's work for months is costly to replace: switching costs in Helmer's sense (what a customer would lose by moving to a rival) for whoever runs it. Ronald Coase, an economist who studied why some work is done inside firms rather than bought, would add that the rent then goes to whoever holds that model, a cloud, a lab, a serving firm or the customer itself, and not necessarily to the firm that did the tuning.
- What would prove this wrong
- Independent tuning firms keeping their margins while the clouds and serving firms bundle tuning with the chips; or, the other way, each new general model wiping out the gains from tuning so firms stop paying for it at all.
- Binds on (how tight)
- Post-training (no reading; a model judges it probably moderate), Continual learning (no reading; a model judges it probably tight), Running models (easing)
Commoditising
Generated data is a feature of whoever owns the model and the chips, not a business of its own
Synthetic data is training data that a model or a program generates: made-up records that mimic real ones, worked problems with their answers, rewritten or translated text. Sellers include Datology, Tonic, Mostly AI, Rendered.ai and Synthesis AI; Gretel was one until NVIDIA bought it in 2025. 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. Kenneth Arrow, an economist who studied how markets handle knowledge, argued that information is hard to own because copying it is cheap. This site judges that generated data is the plainest case of both: its inputs are a model and computing, which the labs and chip makers already own, so a separate seller has little scarce to charge for, unless it starts from data or a way of checking that others lack.
- Getting cheap and interchangeable
- This site judges that generation itself is a commodity whose cost follows the price of running models down. 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, called the inputs a product needs to reach its customers complementary assets; this site judges that the model and the chips a generator needs are held by the labs and chip makers, not by the generator.
- Staying scarce
- This site judges that what stays scarce is the check: generated data is only as good as the test of whether it is right. Where a program can check, as when code runs or a proof holds, generated practice is abundant; where it cannot, people must check it, and that scarcity sits with the expert-data sellers and environment makers, not here. 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 among the generators.
- What would make it last
- A generator becomes the kind of copy a regulator accepts in place of real records, such as privacy-safe versions of health or bank data, turning its rent into a regulated one; or its way of checking generated data proves hard to learn, which Helmer calls process power (a way of working that takes rivals years to learn). Ronald Coase, an economist who studied why some work is done inside firms rather than bought, and Oliver Williamson, an economist who studied contracts and found that a buyer that depends on an asset built for it can be squeezed by whoever owns it, would otherwise expect the owners of models and chips to keep generation inside, as NVIDIA did.
- What would prove this wrong
- An independent seller of generated data keeping a lasting margin (what it keeps from each sale after paying its costs) while the labs' own models generate data ever more cheaply; or the labs paying more for generated data than for data from people in the same field.
- Binds on (how tight)
- Training data (no reading; a model judges it probably moderate), Running models (easing)
Holding
Licensed content earns a rent only for archives no one else holds that keep growing; the rest is priced near what it costs to copy
Data licensing is AI companies paying the owners of text, images, audio or records for the right to train on them or to show them in answers. The middlemen this site tracks include TollBit and ProRata, which charge AI companies for reading publishers' pages; Protege, Created by Humans, Defined.ai and Human Native, which broker datasets; and the RSL Collective, which backs a standard way for sites to state their licence terms to AI crawlers (the programs that read websites for AI companies). 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 an archive sold once becomes, in effect, part of every model trained on it, while content that keeps coming, such as a forum where people keep writing or a daily news service, can be sold again each year.
- Getting cheap and interchangeable
- This site judges that ordinary web text is already a commodity for training: it has been read, copied into many open collections, and cannot be taken back. Kenneth Arrow, an economist who studied how markets handle knowledge, argued that information is hard to own because copying it is cheap. Andrei Hagiu and Julian Wright, economists who asked when data keeps rivals out, found it does so only when it is proprietary, keeps its value and is hard to copy; an archive sold to many labs fails the first test, and old text fails the second once it is in the models.
- Staying scarce
- This site judges that archives which are unique, legally clean and still growing stay scarce: a platform's ongoing conversations, a publisher's daily reporting, a firm's records of how its experts work. The power belongs to the owners of those archives, most of whom sit outside this layer; the brokers in it earn a fee. 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 the brokers today.
- What would make it last
- Courts or lawmakers settle that training on content needs the owner's permission, turning the owners' rent into a regulated one; or a licensing standard such as RSL becomes what AI crawlers must honour, which could give the brokers network economies in Helmer's sense (each publisher and each AI buyer that joins makes the market more useful to the others). Ronald Coase, an economist who studied why some work is done inside firms rather than bought, would expect the largest buyers to strike direct deals with the largest archives and leave the brokers the long tail.
- What would prove this wrong
- Owners of large, still-growing archives failing to win lasting payments because AI companies train on public, generated or scraped material instead; or ordinary web pages fetching licence prices as high as unique archives do.
- Binds on (how tight)
- Training data (no reading; a model judges it probably moderate)
Commoditising
Testing AI is being folded into the labs' and clouds' platforms; an independent tester keeps a rent only where buyers need a judge who is not the seller
Evaluation firms test what a model or an agent can do and watch it once it is running, which the industry calls observability: logging what it did and flagging its mistakes. They include Braintrust, LangSmith (from LangChain), Arize, Galileo, Patronus, Vals AI, Confident AI, Atla, and LMArena, which ranks models by public votes. 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 testing tools sit close to the models and chips they test, so the firms that sell those can give testing away as a feature; what a buyer cannot get from the seller is a verdict from someone who is not the seller.
- Getting cheap and interchangeable
- This site judges that testing tools, the logging, the scoring against a set of questions and the dashboards, are becoming a commodity: open-source versions exist and the platforms bundle them. Public benchmarks (standard published tests used to compare models) go stale: once a test is published, models are trained towards it. Kenneth Arrow, an economist who studied how markets handle knowledge, argued that information is hard to own because copying it is cheap, and a published test is information.
- Staying scarce
- This site judges that what stays scarce is a verdict the buyer trusts because the tester does not sell the model: tests for regulated work, for insurers and for buyers choosing between labs, built on private tasks no model has been trained 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; the one that would fit here is branding (buyers paying more for the same thing because of who made it), and this site does not find it proven for any tester yet.
- What would make it last
- Regulators or insurers require an outside tester's sign-off before AI is used in some kinds of work, turning the tester's rent into a regulated one; or one tester's name becomes the mark buyers insist on, which Helmer calls branding. Ronald Coase, an economist who studied why some work is done inside firms rather than bought, and Oliver Williamson, an economist who studied contracts and found that a buyer that depends on an asset built for it can be squeezed by whoever owns it, would otherwise expect the labs and clouds to keep buying testers whose tools become specific to their models, as they have so far.
- What would prove this wrong
- Independent testers growing faster than the labs' and clouds' own testing tools and keeping their margins (what they keep from each sale after paying their costs) as the platforms give testing away.
- Binds on (how tight)
- Tests and evaluations (no reading; a model judges it probably moderate), Agents and tool use (easing)
The companies that matter here
Candidate to be worth a trillion dollars
Scale AI
Show the profile
- Market
- Scale AI sells human data: training data made or checked by people, from simple labelling to expert work, which AI labs use to teach and test their models. It also sells applications: AI systems built on top of models, which its enterprise page says it builds, deploys and runs for businesses and delivers to governments. It calls its data engine the foundation of model building at many of the major frontier labs (the firms building the most capable AI systems); that is its own description. In June 2025 Meta, which also builds frontier models, bought a non-voting minority stake; Meta's quarterly filing says Meta has no significant influence over Scale's operations. The deal also widened Scale's commercial work for Meta, paid money from the investment out to Scale's shareholders, and took Scale's founder, Alexandr Wang, to Meta to work on its AI, while he stayed on Scale's board; by July 2026 Scale described him as its chairman. After a year under an interim chief executive, it named Francis deSouza, a former Google Cloud and Illumina executive, as chief executive from August 2026. Its enterprise page says it builds, deploys and runs AI systems for businesses, names fields such as energy, insurance, health, retail and life sciences, and says its systems work with any leading model or with open-source ones (models whose makers publish them for anyone to run and change). This site reads that as a move from selling data to the labs towards doing the work for ordinary companies. How big the market becomes depends on whether labs keep buying human data from outside firms, and on whether its applications business grows into a second market.
- Customers (concentrated)
- Frontier labs, large businesses and governments. At the time of the Meta deal, a press report in this site's records named Google, Microsoft and OpenAI, all rivals of Meta, among its customers. In July 2026 it named BP and Mayo Clinic as new business customers and said its work for the US and other governments had grown; it lists the US Department of War among the organisations it serves. Its enterprise page carries praise from executives at Howard Hughes and Global Atlantic, both business customers. The sources do not say whether rival labs cut their work with Scale after Meta's investment, and this site has no record of it either way. This site judges the data business is concentrated in a few labs, Meta among them, and that the government and business side is broader.
- Powers
- switching costs: Emerging, not proven. Hamilton Helmer, a strategist, listed seven sources of lasting profit and required each to have both a benefit and a barrier to copying. The benefit is repeat government and business work; this site judges the barrier is the security approval and integration a government buyer would have to redo to move its programmes elsewhere. This site has no measure of customers staying for that reason.
- Under the site's rent rule
- competed away to users. Depends on: Training data (no reading; a model judges it probably moderate), Post-training (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), Capital (slack).
What would have to hold by 2035
- Labs keep buying human data from outside firms rather than building their own expert teams. Ronald Coase, an economist, asked why firms make some things in-house and buy others, and found they buy when using the market costs less than organising the work themselves; a lab keeps buying only while an outside vendor finds and checks experts more cheaply than it could.
- Rival labs keep trusting Scale with their training plans while a rival lab, Meta, holds a sizeable minority stake by a Scale spokesperson's account, and its founder works there. Oliver Williamson, an economist who studied contracts, found that firms bring in-house the work tied closely to their own plans; a lab's data requests reveal what it is trying to teach its models, which is the kind of work a firm is wary of handing to anyone tied to a rival. Meta's stake carries no vote, which is Scale's answer to that worry.
- Its applications business, which its enterprise page presents as building, deploying and running AI systems for businesses and delivering them to governments, grows into a second large market, so Scale does not rest on the few labs that buy most human data. Oliver Williamson found that a supplier whose work is built for a few buyers can be squeezed by them; selling to many firms is the way out, and it works only if what Scale learns on one client's job makes the next one cheaper, rather than each job being priced by its people's time.
- Human expert data stays needed as models learn more from graded outcomes and from data the models generate themselves.
What would show it failing
- Frontier labs other than Meta dropping Scale as a named supplier, or naming rival vendors in its place.
- Labs building their own expert teams or buying data vendors outright; a fall in this site's count of labs buying from the training-input layer would be one sign.
- Its government and business customers staying a small side business beside the lab work.
- Meta taking a larger or voting stake, which would turn Scale into a supplier tied to one lab.
- Its enterprise work staying custom projects priced by staff time, with nothing that carries from one client to the next.
Drafted by claude-opus-5.5 from Scale AI company page, 28 Sep 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), Scale AI announcement, 30 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
Mercor
Show the profile
- Market
- Mercor runs a marketplace that connects experts, such as doctors, bankers and lawyers, with AI labs and businesses that pay them to train and test AI models. The experts write the questions, judge the answers and turn their own work into evaluations (tests a model can be scored and trained against). It says it sits where the labour market meets AI research, and it says its latest money goes to a larger talent network, better matching and faster delivery. 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 all of the shares offered in its late-2025 round without paying any placement agent (a bank paid to find investors), and that declined to disclose its revenue. In this site's records its gross revenue run-rate (the latest month's billings scaled up to a year, before paying its experts, as this site reads 'gross') grew steeply in the first months of 2026; that figure comes from its chief executive and from a press report citing a person who knows its finances, so it is self-reported, not audited. In March 2026 it launched Mercor Enterprise AI for ordinary companies: its co-chief executive wrote that it maps how a firm's staff actually work, turns that into evaluations and rules for the firm's AI agents (programs that carry out several steps of a task on their own), checks every output as the agents run and feeds corrections back, using tools first built for the labs. In July 2026 it agreed to buy Deeptune, a maker of training environments (simulated workplaces where a model practises tasks and is graded), and said the pair would serve companies deploying agents in their own work as well as labs. How big the market becomes depends on how long labs keep paying people to teach models, and whether businesses outside the labs start paying too.
- Customers (concentrated)
- AI labs and, since its March 2026 launch, ordinary companies deploying AI agents; the sources name none. A press headline in this site's records ties its growth to AI labs buying expert data. This site judges its revenue is concentrated in a few frontier labs.
- Powers
- network economies: Emerging, not proven. Hamilton Helmer, a strategist who listed seven sources of lasting profit and required each to have both a benefit and a barrier to copying, counts a network as a power when each new user makes the product better for the rest. More experts could let Mercor fill more lab projects faster, which draws more projects and so more experts. The barrier is weak while experts can sign up with several vendors at once and this site tracks many rival vendors; there is no retention evidence.
- Under the site's rent rule
- competed away to users. Depends on: Training data (no reading; a model judges it probably moderate), Post-training (no reading; a model judges it probably moderate), Tests and evaluations (no reading; a model judges it probably moderate), Capital (slack).
What would have to hold by 2035
- Labs keep buying expert work from outside firms rather than hiring the experts themselves. Ronald Coase, an economist, asked why firms make some things in-house and buy others, and found they buy when using the market costs less than organising the work themselves; this site reads Mercor's case as that recruiting and checking experts across many fields is cheaper done once for many labs.
- Its network compounds. On the tests of Andrei Hagiu and Julian Wright, economists who set out tests for when data really keeps rivals out, among them whether what a firm learns from one customer improves the product for the others, what Mercor learns about which expert does which task well has to improve matching for every lab, not only the lab that paid for the project.
- What it keeps from each dollar a lab pays holds up as rival vendors compete for the same experts and the same projects.
- Businesses outside the labs start paying for expert-built evaluations and training environments, which is where its enterprise product and its Deeptune purchase point. Oliver Williamson, an economist who studied contracts, found that a supplier whose work is built for a few buyers can be squeezed by them; selling to many firms is the way out, and it pays only if what Mercor learns on one firm's agents carries to the next, rather than each job being paid for by its people's time.
- Demand for people to teach models does not fall away as models learn from graded outcomes and from data they generate themselves.
What would show it failing
- Its self-reported run-rate flattening or falling while labs' spending keeps rising.
- Labs moving expert work in-house or buying vendors outright; a fall in this site's count of labs buying from the training-input layer would be one sign.
- The data market splitting among specialist vendors by field, leaving no gain from one broad network.
- No named business customer outside the labs as its enterprise and environments businesses grow.
Drafted by claude-opus-5.5 from Mercor announcement, 9 Jul 2026 (opens in a new tab), Mercor announcement, 26 Mar 2026 (opens in a new tab), Mercor Form D, 18 Feb 2026 (opens in a new tab), Mercor announcement, 27 Oct 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
Handshake AI
Show the profile
- Market
- Handshake AI, launched in 2025, is the part of Handshake that supplies AI labs with expert people to test and judge their models: they write subject-specific questions, assess the answers and grade their quality, in fields from law and finance to virology. Handshake runs the whole job itself, finding and training the experts, checking quality and producing the data on its own software. Its supply comes from its parent: Handshake is a careers network that links students and graduates with universities and employers, and it calls itself the largest early-career network in the world; that is its own description. Handshake AI says it draws experts straight from its university partners, so their degrees are verified rather than self-reported, and that this lets it spend less on recruiting than rivals do. In this site's records Handshake told the press that the unit's gross revenue run-rate (the latest month's billings scaled up to a year) went from nothing to a large figure in little more than a year; that is self-reported, not audited. It also sells to ordinary companies under the name Handshake Enterprise AI: its enterprise page, which carries no date, says it builds AI agents (programs that carry out several steps of a task on their own) inside a client's own systems, lowers what they cost to run, and trains models the client owns. In July 2026 it wrote that evaluations (tests of an AI system on the firm's own work) built with its experts let a firm find which work can move to cheaper models, open-source ones (models published for anyone to run) among them. How big the market becomes depends on how long labs keep paying graduate-level experts to teach models, and on whether ordinary companies pay for the same skill.
- Customers (concentrated)
- Leading AI labs, which the sources do not name; its undated enterprise page says each of the largest frontier labs (the firms building the most capable AI systems) trains on its data. It also sells to ordinary companies; the same page describes unnamed clients, such as a team buried in documents that it says it helped resolve cases faster. This site judges its revenue is concentrated in a few frontier labs.
- Powers
- a cornered resource: Emerging, not proven. Hamilton Helmer, a strategist who listed seven sources of lasting profit and required each to have both a benefit and a barrier to copying, counts preferential access to a valued input as a power. The benefit is cheaper, credential-checked recruiting through long-standing university partnerships. The barrier is unproven: universities are not bound to one partner, and the same graduates can work for rival vendors.
- Under the site's rent rule
- competed away to users. Depends on: Training data (no reading; a model judges it probably moderate), Post-training (no reading; a model judges it probably moderate), Tests and evaluations (no reading; a model judges it probably moderate), Capital (slack).
What would have to hold by 2035
- Labs keep buying expert judgement from outside firms. Ronald Coase, an economist, asked why firms make some things in-house and buy others, and found they buy when using the market costs less than organising the work themselves; Handshake says it reaches graduate experts more cheaply than rivals that spend heavily on recruiting, and a lab keeps buying only while that stays cheaper than doing the work itself.
- Its university network stays a cheaper and more trusted route to experts than rivals' recruiting, and universities keep endorsing the work for their students.
- Its supply compounds. On the tests of Andrei Hagiu and Julian Wright, economists who set out tests for when data really keeps rivals out, among them whether what a firm learns from one customer improves the product for the others, knowing which graduates do which lab task well has to improve the service for every lab, not just one.
- Demand for graduate-level human judgement outlasts models learning from graded outcomes and from data they generate themselves.
- Ordinary companies pay for tests and agents built with its experts, so it does not rest on a few labs. Oliver Williamson, an economist who studied contracts, found that a supplier whose work is built for a few buyers can be squeezed by them; selling to many firms is the way out, and it pays only if what Handshake learns on one client's work carries to the next.
What would show it failing
- University partners limiting the work or signing with rival vendors.
- Labs moving expert work in-house or buying vendors outright; a fall in this site's count of labs buying from the training-input layer would be one sign.
- Its self-reported run-rate flattening while rivals' keep rising.
- Its enterprise work staying a handful of custom projects with no named client.
Drafted by claude-opus-5.5 from Handshake AI company page, 28 Sep 2026 (opens in a new tab), Handshake AI announcement, 13 Jun 2025 (opens in a new tab), Handshake AI announcement, 14 Oct 2025 (opens in a new tab), Handshake AI announcement, 15 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
Surge AI
Show the profile
- Market
- Surge AI sells human data: material written and graded by experts, such as example answers, practice tasks and judgements of a model's answers, which AI labs use to train and test their models. It also lists ready-made data and practice environments (simulated workplaces where a model practises tasks and is graded). Its enterprise page says years of work with leading labs taught it what improves a model, and that it now brings those methods to ordinary companies. It offers to build evaluations (tests of an AI system on the firm's own work and standards), compare models on quality, cost and speed, test systems before launch and watch them once live, find what is holding a system back beyond the model itself, and, where that is not enough, design a training programme and build its data, including post-training (further training of a finished model on narrower material). In this site's records a press report said its yearly revenue for 2024 passed a large figure without it raising outside money; that is reported, not audited, and this site records no funding round for it, though its records also hold a July 2025 press report that Surge was preparing to raise outside money. How big the market becomes depends on how long labs keep paying experts to teach models, and on whether ordinary companies pay for the same work.
- Customers (concentrated)
- AI labs, and now ordinary companies. Press reports in this site's records name Google, OpenAI and Anthropic among the labs it serves, and a research firm estimated that a group of frontier labs (the firms building the most capable AI systems) drove most of its growth; that is an outside estimate, not a company figure. Its enterprise page names no business customer. This site judges its revenue is concentrated in a few frontier labs.
- Powers
- process power: Emerging, not proven. Hamilton Helmer, a strategist who listed seven sources of lasting profit and required each to have both a benefit and a barrier to copying, counts as process power a way of working that yields better quality or lower cost and that rivals cannot copy quickly. Surge says it tests whether its own data improves models and publishes the gains, and that it can return pilot results within a week; this site reads its growth without outside money as a sign of an efficient operation. The barrier is unshown: rivals serve the same labs with the same kinds of experts.
- Under the site's rent rule
- competed away to users. Depends on: Training data (no reading; a model judges it probably moderate), Post-training (no reading; a model judges it probably moderate), Tests and evaluations (no reading; a model judges it probably moderate).
What would have to hold by 2035
- Labs keep buying expert work from outside firms rather than hiring the experts themselves. Ronald Coase, an economist, asked why firms make some things in-house and buy others, and found they buy when using the market costs less than organising the work themselves; a lab keeps buying from Surge only while it finds and checks experts more cheaply than the lab could.
- Ordinary companies pay for tests and training built on their own work, as its enterprise page is built to win, so it does not rest on the few labs that buy most expert data. Oliver Williamson, an economist who studied contracts, found that a supplier whose work is built for a few buyers can be squeezed by them; selling to many firms is the way out.
- What it learns compounds. On the tests of Andrei Hagiu and Julian Wright, economists who set out tests for when data really keeps rivals out, among them whether what a firm learns from one customer improves the product for the others, what Surge learns testing one firm's systems has to make the next job better or cheaper; otherwise its enterprise work is consulting paid for its people's time.
- Its gains stay its own. David Teece, an economist who studied who profits from a new idea, found that the gains go to whoever holds the assets needed to bring it to market; here the labs, which own the models and the finished data, hold them, so Surge keeps a share only while its experts and checking are hard to replace.
- Demand for expert judgement outlasts models learning from graded outcomes and from data they generate themselves.
What would show it failing
- The labs this site's records name as its buyers naming other vendors in its place, or moving expert work in-house; a fall in this site's count of labs buying from the training-input layer (the firms that sell labs the data, practice tasks and tests models learn from) would be one sign.
- No named business customer outside the labs as its enterprise work grows.
- Its enterprise work staying custom projects priced by staff time, with nothing that carries from one client to the next.
- It raising outside money on terms that suggest its own earnings no longer fund its growth; the July 2025 press report in this site's records that it was preparing to raise is the first sign to watch.
Drafted by claude-opus-5.5 from Surge AI company page, 28 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.