What to build, future by future
The site lists businesses that could be built where the value chain is short of something. This page takes a position on each: build it, build it on a condition, or don't build it alone, with one line of reasoning. Then it asks what each future would do to that business. A future here is one square of the grid on the outlook page: how far AI goes, crossed with how the rules settle. Only the squares a named writer argues for are used. A second group asks who ends up owning the models, where they run and who they deal for; each of those is a summary of what a writer the outlook already holds argues.
Choose a future and the businesses it strengthens fill in, the ones it weakens turn to a dashed outline, any it breaks are struck through, and the rest fade. A business that some future strengthens and none weakens is marked. Open any row for the problem it solves, the profit the site's rent rule gives it, what would overturn the call, and the reason behind each future's effect. Every call and every effect is a model's judgement, not a reading, and each is labelled. Further down the page ties the site's own predictions of where the shortage moves to these businesses, and sets out the assumptions the calls rest on.
A few words recur. An agent is an AI program that carries out a task on its own, such as filing a claim or booking a purchase, rather than only answering questions. A lab is a company that builds the largest AI models, such as OpenAI or Anthropic, and a cloud is a company that rents out computing, such as Amazon or Microsoft; both can bundle extra services into what they already sell. A roll-up is a firm that buys many small businesses in one trade and runs them together, and a system of record is the software where a firm keeps the official copy of its work, such as a law firm's case files.
The value chain is the run of businesses from power and chips to the products people use. A rent is profit beyond what it takes to keep a business going, and the rent rule is how this site judges who would keep it and how large it would be; the businesses page sets it out. An open model is one whose maker publishes it for anyone to run, and tuning is adjusting such a model for one kind of work. An identity firm is a company that confirms who a person or a program is before another system lets it in.
The chain, and what is scarce on it
Each tile is one part of the value chain, from power on the left to applications on the right, and it never moves. Its shade says how short that part is of the thing it most needs. Now: What the site reads tonight, and a model's judgement where it has no reading. No future changes it. Transition: The years in which AI is capable but firms are still fitting it into how they work. Shortages here come from the pace of adoption, not the pace of the models. Mature state: A hypothesis of what stays scarce once firms have fitted AI in. Nothing here is a forecast with a date. A reading from tonight's scorecard is drawn in the scorecard's colours; a model's judgement is drawn in ink with a dashed edge. A tile leads to that part's entry in the list further down.
The map shows Now, with no future chosen.
1 Physical inputs and compute supply
2 Models and capability production
3 Data, training and verification inputs
4 Runtime, serving and orchestration
5 Trust, control and transaction infrastructure
6 Deployment and organizational diffusion
7 Applications and operating models
Build
Scarce in most futures, with a profit worth the fight.2Independent sign-off for AI in regulated workFirst betno future here weakens itStart here. Where a rule asks for an outside sign-off, the seller cannot be the one who gives it.unchangedunchangedunchangedstronger herestronger hereunchangedstronger hereunchangedstronger hereunchangedunchangedunchanged
The problem
Regulated firms cannot use a model until someone other than the seller vouches for it.
The profit, by the site's rule
A monopoly-like profit, kept by the maker.
What would overturn this call
Regulators accept the seller's own tests.
What it builds up: the backing.
Future by future
- Steady progress, courts and insurers price the risk: stronger. Courts and insurers ask who vouched for the system, and the seller cannot vouch for itself.
- Steady progress, oversight comes first: stronger. Registration before dealing gives a regulator a point at which to ask for an outside sign-off.
- AI automates research, oversight comes first: stronger. A race slowed until systems can be checked makes an outside check worth more.
- Labs charge for their lead and move up the stack: stronger. The more a firm depends on one seller, the more it needs someone else to vouch for what is sold.
Sits on 3.8 Evals and assurance. The full record: what it sells first, what it builds up and what would prove it wrong
A model's judgement, not a reading: the call, the line under the name, what would overturn it and each future's effect. The problem, the profit and its place in the order of bets are the site's own record.
1Outcome verification with pay-per-verified-result billingThird betno future here weakens itStrong on its own. The more work agents do, the scarcer the proof that it was done right.unchangedunchangedunchangedstronger hereunchangedstronger hereunchangedunchangedunchangedunchangedstronger hereunchanged
The problem
Nobody can cheaply prove that an agent finished a task correctly, so buyers keep humans checking every result.
The profit, by the site's rule
A fat profit, kept by the maker.
What would overturn this call
The labs ship self-checking that buyers accept.
What it builds up: the test, the yardstick.
Future by future
- Steady progress, courts and insurers price the risk: stronger. Once a payer is liable for an agent's mistakes, it will pay for proof that the work was done right.
- AI automates research, the rules stay unsettled: stronger. More agent work with fewer people watching makes an outside check worth more, unless agents come to check themselves.
- Agents deal in markets for people and firms: stronger. A buyer's agent needs proof that the work was done before it pays.
Sits on 5.4 Outcome verification. The full record: what it sells first, what it builds up and what would prove it wrong
A model's judgement, not a reading: the call, the line under the name, what would overturn it and each future's effect. The problem, the profit and its place in the order of bets are the site's own record.
5Tamper-proof audit trail for what agents didno future here weakens itUnglamorous and durable. Whoever pays, licenses or investigates starts from the record.unchangedunchangedunchangedstronger herestronger herestronger hereunchangedunchangedunchangedunchangedstronger hereunchanged
The problem
After an agent acts, nobody can reconstruct what it saw, decided and changed.
The profit, by the site's rule
A fat profit, kept by the maker.
What would overturn this call
The audit record becomes a free feature of the software that runs the agent.
What it builds up: the record, the backing.
Future by future
- Steady progress, courts and insurers price the risk: stronger. A claim cannot be settled without a record of what the agent did.
- Steady progress, oversight comes first: stronger. Registration is only useful if what a registered agent did can be read back afterwards.
- AI automates research, the rules stay unsettled: stronger. Faster agents doing more work leave more to reconstruct after the fact.
- Agents deal in markets for people and firms: stronger. A deal between agents is settled from the record.
Sits on 5.3 Provenance and audit. The full record: what it sells first, what it builds up and what would prove it wrong
A model's judgement, not a reading: the call, the line under the name, what would overturn it and each future's effect. The problem, the profit and its place in the order of bets are the site's own record.
4Know-your-agent and delegated-authority registryA monopoly-like prize: every counterparty has to ask whoever holds the registry. The fight is with the identity firms, so take a regulated niche first.weaker hereunchangedunchangedstronger herestronger hereunchangedunchangedunchangedunchangedunchangedstronger hereunchanged
The problem
A counterparty cannot tell which agent acts for whom, and under what limits.
The profit, by the site's rule
A monopoly-like profit, kept by incumbent firms.
What would overturn this call
An identity firm bundles it for free.
What it builds up: the record, the backing.
Future by future
- Capability stalls, the rules stay unsettled: weaker. If whole jobs stay out of reach, agents keep working inside a person's own session, and no separate identity is needed.
- Steady progress, courts and insurers price the risk: stronger. A liable party needs to know which agent acted for whom before it pays or sues.
- Steady progress, oversight comes first: stronger. Registration before dealing is this business written into law.
- Agents deal in markets for people and firms: stronger. A stranger's agent has to show whom it acts for before anyone deals with it.
Sits on 5.2 Identity and authority. The full record: what it sells first, what it builds up and what would prove it wrong
A model's judgement, not a reading: the call, the line under the name, what would overturn it and each future's effect. The problem, the profit and its place in the order of bets are the site's own record.
13A trade's AI system of record, with the approvals only that trade can giveA strong application bet. It holds if the records and the approvals stay when the models change.unchangedunchangedunchangedstronger hereunchangedweaker hereunchangedunchangedweaker hereunchangedunchangedunchanged
The problem
A general model can draft the work but cannot hold the trade's records or sign for it.
The profit, by the site's rule
A fat profit, kept by the maker.
What would overturn this call
A general model is allowed to hold the records and sign.
What it builds up: the record, the backing.
Future by future
- Steady progress, courts and insurers price the risk: stronger. The trade's own approvals are what a court or an insurer is likeliest to recognise.
- AI automates research, the rules stay unsettled: weaker. Models that improve this fast are the case in which customers move to the labs' own products.
- Labs charge for their lead and move up the stack: weaker. Labs that move into their customers' work compete for the same place.
Sits on 7.3 Vertical workflow software. The full record: what it sells first, what it builds up and what would prove it wrong
A model's judgement, not a reading: the call, the line under the name, what would overturn it and each future's effect. The problem, the profit and its place in the order of bets are the site's own record.
15A licensed firm that sells the finished job and stands behind itThe endgameThe endgame. It needs a licence and a balance sheet, so arrive through sign-off and verification.weaker hereweaker hereunchangedstronger herestronger hereunchangedunchangedunchangedunchangedunchangedunchangedunchanged
The problem
Buyers want one accountable party when consequential work goes wrong.
The profit, by the site's rule
A fat profit, kept by the maker.
What would overturn this call
Regulators license the software itself, with no firm behind it.
What it builds up: the backing.
Future by future
- Capability stalls, the rules stay unsettled: weaker. A firm that sells the finished job needs agents that can finish it.
- The money stalls, the rules stay unsettled: weaker. Its own first need is capital, and there is less of it.
- Steady progress, courts and insurers price the risk: stronger. Buyers want one accountable party, and a licence makes the firm that party.
- Steady progress, oversight comes first: stronger. Oversight that decides who may act favours firms that already hold a licence.
Sits on 7.4 Autonomous services. The full record: what it sells first, what it builds up and what would prove it wrong
A model's judgement, not a reading: the call, the line under the name, what would overturn it and each future's effect. The problem, the profit and its place in the order of bets are the site's own record.
21A firm's own map of what its data means, kept on its own servers for every agent to readThe system of record for the whole firm and not for a single trade: what the firm's data means, written down once, on its own servers, for every agent to read. Build it where data cannot leave the building; elsewhere the labs and clouds are likely to offer a memory of their own.stronger hereunchangedunchangedunchangedunchangedunchangedunchangedstronger hereweaker hereunchangedunchangedstronger here
The problem
What a firm knows is scattered across systems whose tables no document explains, so every agent has to be taught the firm again.
The profit, by the site's rule
A moderate profit, kept by the maker.
What would overturn this call
The labs, the clouds or the vendors that already hold the firm's data ship a model of the firm as a built-in feature, and buyers accept it.
What it builds up: the record, the choice.
Future by future
- Capability stalls, the rules stay unsettled: stronger. A model that will not learn the firm on its own needs the firm written down.
- Open models are good enough for most work: stronger. A record that works with any model is worth more when good models are free to swap.
- Labs charge for their lead and move up the stack: weaker. Labs that move into their customers' work offer the memory themselves.
- Governments decide which models a firm may use: stronger. A firm whose permitted models can change needs its knowledge kept apart from any one of them.
Sits on 4.4 Context and memory. The full record: what it sells first, what it builds up and what would prove it wrong
A model's judgement, not a reading: the call, the line under the name, what would overturn it and each future's effect. The problem, the profit and its place in the order of bets are the site's own record.
Build on a condition
Worth building as a way in, or once one thing is true.3A model that keeps learning inside one customer's workSecond betA good way in with a moderate profit. Build it for what it learns inside the customer's work, which the customer would lose by leaving.stronger hereunchangedunchangedunchangedunchangedweaker hereunchangedstronger hereweaker herestronger hereunchangedunchanged
The problem
Everyday work leaves no reusable signal, so models stop improving once deployed.
The profit, by the site's rule
A moderate profit, kept by the maker.
What would overturn this call
The labs offer learning inside the customer themselves.
What it builds up: the record.
Future by future
- Capability stalls, the rules stay unsettled: stronger. With no better general model coming, what a model learns inside one customer keeps its value.
- AI automates research, the rules stay unsettled: weaker. When each model speeds up the next, a new release is likelier to overtake what the customer's model learned.
- Open models are good enough for most work: stronger. When the base model is free, what a model learns inside one customer is the part worth paying for.
- Labs charge for their lead and move up the stack: weaker. Labs that move into their customers' work offer the learning themselves.
- Everyday work moves onto devices: stronger. A model that runs beside the customer's data can learn from it there.
Sits on 3.7 Post-training and learning infrastructure. The full record: what it sells first, what it builds up and what would prove it wrong
A model's judgement, not a reading: the call, the line under the name, what would overturn it and each future's effect. The problem, the profit and its place in the order of bets are the site's own record.
11Deployment that leaves behind the platform the customer runs onA quick way into a customer today that may not last. If agents do the fitting themselves, there is no scarce labour left to charge for.stronger hereunchangedstronger hereunchangedunchangedweaker hereunchangedunchangedweaker hereunchangedunchangedunchanged
The problem
Engineers who can turn messy data and unwritten habits into working systems are scarce.
The profit, by the site's rule
A moderate profit, kept by the maker.
What would overturn this call
Agents fit themselves into a firm without engineers.
What it builds up: the record.
Future by future
- Capability stalls, the rules stay unsettled: stronger. With no better model coming, the gains come from fitting what exists into firms.
- Steady progress, the rules stay unsettled: stronger. Adoption at the pace of past technologies means years of fitting work.
- AI automates research, the rules stay unsettled: weaker. Agents that can do the fitting remove the labour this business sells.
- Labs charge for their lead and move up the stack: weaker. Labs that deploy inside their customers' work do this fitting themselves.
Sits on 6.1 Technical implementation. The full record: what it sells first, what it builds up and what would prove it wrong
A model's judgement, not a reading: the call, the line under the name, what would overturn it and each future's effect. The problem, the profit and its place in the order of bets are the site's own record.
6AI warranty backed by a pool of loss datano future here weakens itToo early alone: there is too little loss data to price yet. Strong once verification and the audit trail supply it.unchangedunchangedunchangedstronger hereunchangedunchangedunchangedunchangedunchangedunchangedstronger hereunchanged
The problem
Agent failures cannot be priced, so procurement stalls or insurers exclude them.
The profit, by the site's rule
A fat profit, kept by the maker.
What would overturn this call
Insurers price agent risk from their own claims first.
What it builds up: the backing.
Future by future
- Steady progress, courts and insurers price the risk: stronger. Once rules settle who pays, the risk can be priced, though the same future lets insurers price it themselves.
- Agents deal in markets for people and firms: stronger. A firm that lets an agent spend for it wants someone standing behind the agent.
Sits on 5.6 Liability and risk transfer. The full record: what it sells first, what it builds up and what would prove it wrong
A model's judgement, not a reading: the call, the line under the name, what would overturn it and each future's effect. The problem, the profit and its place in the order of bets are the site's own record.
18Routing that keeps a firm inside approved models and fails over when one is cut offno future here weakens itIts worth turns on how far governments decide which models a firm may use. Where they do, this business keeps a firm inside the line, though the routing itself can be copied.unchangedunchangedunchangedunchangedunchangedunchangedunchangedunchangedunchangedunchangedunchangedstronger here
The problem
Access to models splits by country and supplier, and a banned supplier breaks the product.
The profit, by the site's rule
A fat profit, kept by the maker.
What would overturn this call
One supplier is approved everywhere.
What it builds up: the choice.
Future by future
- Governments decide which models a firm may use: stronger. A firm barred from a supplier has to stay inside the models it is allowed and switch when one is cut off.
Sits on 4.3 Routing and gateways. The full record: what it sells first, what it builds up and what would prove it wrong
A model's judgement, not a reading: the call, the line under the name, what would overturn it and each future's effect. The problem, the profit and its place in the order of bets are the site's own record.
20Incident response and forensics for agent failuresno future here weakens itA natural next product for whoever already holds the audit trail. Demand waits on failures that are public.unchangedunchangedunchangedstronger herestronger herestronger hereunchangedunchangedunchangedunchangedunchangedunchanged
The problem
After a public agent failure, nobody can say quickly what went wrong or who is liable.
The profit, by the site's rule
A moderate profit, kept by the maker.
What would overturn this call
Failures stay rare or stay private.
What it builds up: the backing.
Future by future
- Steady progress, courts and insurers price the risk: stronger. Each claim needs someone to say what went wrong.
- Steady progress, oversight comes first: stronger. A registered agent has a named party behind it, who must answer when it fails.
- AI automates research, the rules stay unsettled: stronger. Faster agents doing more work fail in more places, and each failure has to be explained.
Sits on 5.1 Security. The full record: what it sells first, what it builds up and what would prove it wrong
A model's judgement, not a reading: the call, the line under the name, what would overturn it and each future's effect. The problem, the profit and its place in the order of bets are the site's own record.
14An AI roll-up that moves its clients to fixed feesAn ownership play more than an AI one: the asset is the client and the AI is bought in. It works only while clients accept fixed fees.unchangedweaker herestronger hereunchangedunchangedunchangedunchangedunchangedunchangedunchangedunchangedunchanged
The problem
A firm billed by the hour hands AI's saving straight back to the client at the next re-bid.
The profit, by the site's rule
A moderate profit, kept by the maker.
What would overturn this call
Clients bid the saving away at renewal.
Future by future
- The money stalls, the rules stay unsettled: weaker. Buying firms takes capital, and there is less of it.
- Steady progress, the rules stay unsettled: stronger. Slow adoption leaves rivals billing by the hour for longer.
Sits on 7.6 AI-native rollups. The full record: what it sells first, what it builds up and what would prove it wrong
A model's judgement, not a reading: the call, the line under the name, what would overturn it and each future's effect. The problem, the profit and its place in the order of bets are the site's own record.
16Moving computing to where power is availablePower is short today. It takes capital and sites, and the profit follows whoever controls the connections to the power grid.weaker hereweaker hereunchangedunchangedunchangedstronger hereweaker herestronger hereunchangedweaker hereunchangedunchanged
The problem
A site already wired to the grid is scarcer than the building or the money.
The profit, by the site's rule
A moderate profit, kept by incumbent firms.
What would overturn this call
Connections to the power grid catch up with demand.
What it builds up: the choice.
Future by future
- Capability stalls, the rules stay unsettled: weaker. If capability stalls, demand for new computing eases, and so does the premium for power.
- The money stalls, the rules stay unsettled: weaker. A funding stall stops the building that makes sites scarce.
- AI automates research, the rules stay unsettled: stronger. Research run by models needs far more computing, and power is what limits it.
- AI automates research, oversight comes first: weaker. A deliberate slowdown eases the rush for new computing, and with it the premium for power.
- Open models are good enough for most work: stronger. Free models move the cost to the computing that runs them.
- Everyday work moves onto devices: weaker. Work done on the device is not rented from a data centre.
Sits on 1.6 Compute utilization and control. The full record: what it sells first, what it builds up and what would prove it wrong
A model's judgement, not a reading: the call, the line under the name, what would overturn it and each future's effect. The problem, the profit and its place in the order of bets are the site's own record.
9Escrow that pays an agent only when its work passes a checkA monopoly-like prize that exists only if agents come to deal with each other. Build it when that trade is visible, and beat the payment firms on proof, not on payments.weaker hereunchangedunchangedunchangedunchangedunchangedunchangedunchangedunchangedunchangedstronger hereunchanged
The problem
Agents can find and order work from other agents, but nothing holds the money until the work is proven done.
The profit, by the site's rule
A monopoly-like profit, kept by incumbent firms.
What would overturn this call
Payment networks themselves hold the money until the work is checked.
What it builds up: the backing, the yardstick.
Future by future
- Capability stalls, the rules stay unsettled: weaker. Agents that cannot finish whole jobs do not buy from other agents.
- Agents deal in markets for people and firms: stronger. Agents that order work for people have to pay for it, and pay on proof.
Sits on 5.5 Settlement and billing. The full record: what it sells first, what it builds up and what would prove it wrong
A model's judgement, not a reading: the call, the line under the name, what would overturn it and each future's effect. The problem, the profit and its place in the order of bets are the site's own record.
10The standard way agents get permission into business softwareWorth most as the standard; without that, what is left is each customer's wiring. The software vendors and the labs want the same place.unchangedunchangedunchangedunchangedunchangedunchangedunchangedunchangedweaker hereunchangedstronger hereunchanged
The problem
Every agent needs its own negotiated, audited link into each system a firm runs.
The profit, by the site's rule
A moderate profit, kept by incumbent firms.
What would overturn this call
A lab's protocol becomes the default.
What it builds up: the record, the choice.
Future by future
- Labs charge for their lead and move up the stack: weaker. A lab built into its customers' work can make its own protocol the default.
- Agents deal in markets for people and firms: stronger. Agents that deal across firms need a way into each firm's systems.
Sits on 4.6 Tools and actuation. The full record: what it sells first, what it builds up and what would prove it wrong
A model's judgement, not a reading: the call, the line under the name, what would overturn it and each future's effect. The problem, the profit and its place in the order of bets are the site's own record.
7Graded expert-work records pooled across buyersOne kind of buyer, the labs. Worth building while they buy graded work from outside.unchangedweaker hereunchangedunchangedunchangedunchangedunchangedunchangedunchangedunchangedunchangedunchanged
The problem
Expert data is made to order and belongs to the buyer, so the seller keeps nothing when the job ends.
The profit, by the site's rule
A moderate profit, kept by the maker.
What would overturn this call
Labs bring expert work in-house.
Future by future
- The money stalls, the rules stay unsettled: weaker. Labs are the buyers of graded expert work, and a lab short of money buys less of it.
Sits on 3.1 Data origination. The full record: what it sells first, what it builds up and what would prove it wrong
A model's judgement, not a reading: the call, the line under the name, what would overturn it and each future's effect. The problem, the profit and its place in the order of bets are the site's own record.
Don't build alone
Useful work whose profit is not kept. A feature of something else.8A router that chooses models from pooled outcomesUseful and unprofitable. The saving goes to the user, so make it a feature.unchangedunchangedunchangedunchangedunchangedunchangedunchangedstronger hereweaker hereunchangedunchangedweaker here
The problem
Each firm guesses which model and venue to use for each task, and pays for the guess.
The profit, by the site's rule
Competed away to users.
What would overturn this call
Pooled outcome data proves impossible to copy.
What it builds up: the yardstick, the choice.
Future by future
- Open models are good enough for most work: stronger. With many good models to choose from, the choice is worth more.
- Labs charge for their lead and move up the stack: weaker. A customer built into one lab's products has less left to route.
- Governments decide which models a firm may use: weaker. A firm allowed only some models has less left to route.
Sits on 4.3 Routing and gateways. The full record: what it sells first, what it builds up and what would prove it wrong
A model's judgement, not a reading: the call, the line under the name, what would overturn it and each future's effect. The problem, the profit and its place in the order of bets are the site's own record.
12Capturing how a firm really works, for its agentsA thin profit that weakens once agents learn a firm's habits on their own. Do it inside deployment or the system of record.stronger hereunchangedstronger hereunchangedunchangedweaker hereunchangedunchangedunchangedunchangedunchangedunchanged
The problem
The knowledge an agent needs sits in people's heads, not in any system.
The profit, by the site's rule
A thin profit, kept by the maker.
What would overturn this call
Unwritten knowledge stays out of reach of any model.
What it builds up: the record.
Future by future
- Capability stalls, the rules stay unsettled: stronger. Firms pay to have their unwritten rules written down when the model will not learn them alone.
- Steady progress, the rules stay unsettled: stronger. Slow adoption keeps the knowledge in people's heads for longer.
- AI automates research, the rules stay unsettled: weaker. Better models learn a firm's habits from its records without the service.
Sits on 6.2 Context engineering. The full record: what it sells first, what it builds up and what would prove it wrong
A model's judgement, not a reading: the call, the line under the name, what would overturn it and each future's effect. The problem, the profit and its place in the order of bets are the site's own record.
19Tuning open models to frontier quality for one industryThe skill is scarce today and the profit is not kept. Tuning ends up a feature of a model that learns inside the customer, or of a trade's system.unchangedunchangedunchangedunchangedunchangedweaker hereunchangedstronger hereunchangedunchangedunchangedstronger here
The problem
Open weights are free but turning them into a reliable product takes scarce skill.
The profit, by the site's rule
Competed away to users.
What would overturn this call
Open models stay hard to tune for years.
Future by future
- AI automates research, the rules stay unsettled: weaker. If each model speeds up the next, the leading labs pull away from open models.
- Open models are good enough for most work: stronger. Tuning is the work that turns a free model into a product.
- Governments decide which models a firm may use: stronger. A firm shut out of the closed models needs an open one tuned to its work.
Sits on 3.7 Post-training and learning infrastructure. The full record: what it sells first, what it builds up and what would prove it wrong
A model's judgement, not a reading: the call, the line under the name, what would overturn it and each future's effect. The problem, the profit and its place in the order of bets are the site's own record.
17Training and placing the engineers who deploy AIA good service firm, not a durable one.stronger hereunchangedstronger hereunchangedunchangedbreaks hereunchangedunchangedunchangedunchangedunchangedunchanged
The problem
Deployment talent, not models, keeps pilots from reaching production.
The profit, by the site's rule
Competed away to users.
What would overturn this call
Agents never learn to do the fitting.
Future by future
- Capability stalls, the rules stay unsettled: stronger. Spreading what already works takes people who can deploy it.
- Steady progress, the rules stay unsettled: stronger. Adoption at the pace of past technologies means years of work for people who can deploy.
- AI automates research, the rules stay unsettled: breaks. If agents do the deployment, there is nobody left to train.
Sits on 6.5 Talent production. The full record: what it sells first, what it builds up and what would prove it wrong
A model's judgement, not a reading: the call, the line under the name, what would overturn it and each future's effect. The problem, the profit and its place in the order of bets are the site's own record.
The ground under today's companies
Every part of the chain, sorted by where its shortage is heading between now and the mature state with no future chosen. Open a part for the reason, what each future does to it, and the companies the site places there, each of which can be put to Ask.
Ground that gets scarcer
Shorter at the mature state than today.1.1Power and sitesSites already connected to the power grid: tight now (tonight’s reading of Electricity generation, a low-confidence score), probably severe at the mature state.
Connections to the power grid take longer to build than anything else in the chain, so the shortage eases last and least. In the transition: probably severe.
- Capability stalls, the rules stay unsettled: eases. Demand for new computing eases, and with it the queue for power.
- The money stalls, the rules stay unsettled: eases. Building slows, and the queue for connections shortens.
- AI automates research, the rules stay unsettled: tightens. Research run by models needs far more computing, and power is what limits it.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading, except the word for now, which is tonight’s reading of Electricity generation, the nearest gauge the site scores: what is scarce, each word, the reason and each future's move.
1.6Compute utilization and controlbusiness 16Knowing where each job should run: probably easing now, probably moderate at the mature state.
Few firms need it while computing is simply short; it matters more once supply is uneven by place and price. In the transition: probably moderate.
No future here is judged to move it.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
3.6Reward and verification systemsA trusted way to grade work: probably moderate now, probably tight at the mature state.
Grading gets harder as tasks get longer and less clear-cut. In the transition: probably tight.
No future here is judged to move it.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
3.8Evals and assurancebusiness 2A judge who is not the seller: probably moderate now, probably tight at the mature state.
The more that rides on a model, the more a buyer needs someone other than its maker to vouch for it. In the transition: probably tight.
- Steady progress, courts and insurers price the risk: tightens. Courts and insurers ask who vouched for the system.
- Steady progress, oversight comes first: tightens. Registration gives a regulator a point at which to ask for an outside test.
- AI automates research, oversight comes first: tightens. The race is slowed until systems can be checked.
- Labs charge for their lead and move up the stack: tightens. A buyer tied to one seller needs a judge who is not that seller.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
4.4Context and memorybusiness 21A firm's own knowledge, kept fresh and permissioned: probably moderate now, probably tight at the mature state.
Every deployment needs it, and it has to be rebuilt firm by firm. In the transition: probably tight.
- AI automates research, the rules stay unsettled: tightens. Faster agents are held back by access to the firm's records, not by skill.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
5.1Securitybusiness 20Defences against agents being turned: probably moderate now, probably tight at the mature state.
Attacks grow with what agents are allowed to do. In the transition: probably tight.
- Steady progress, courts and insurers price the risk: tightens. Each claim needs someone to say how the agent was turned or went wrong.
- Steady progress, oversight comes first: tightens. A registered agent that is turned is its registrant's problem.
- AI automates research, the rules stay unsettled: tightens. Faster agents doing more work can be turned in more places.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
5.3Provenance and auditbusiness 5A record of what an agent did that nobody can alter: probably moderate now, probably tight at the mature state.
Every dispute and every review starts from the record, and few agents keep one. In the transition: probably tight.
- Steady progress, courts and insurers price the risk: tightens. A claim is settled from the record.
- Steady progress, oversight comes first: tightens. A registered agent has to be able to show what it did.
- AI automates research, the rules stay unsettled: tightens. Faster agents doing more work leave more to reconstruct after the fact.
- Agents deal in markets for people and firms: tightens. A deal between agents is settled from the record.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
5.4Outcome verificationbusiness 1Proof that an agent's work was done right: probably moderate now, probably severe at the mature state.
Producing work gets cheaper while checking it does not. In the transition: probably tight.
- Steady progress, courts and insurers price the risk: tightens. A payer who is liable wants proof.
- AI automates research, the rules stay unsettled: tightens. More work is done with fewer people watching.
- Agents deal in markets for people and firms: tightens. Payment waits on proof.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
5.5Settlement and billingbusiness 9A way to pay an agent only when the work is done: probably slack now, probably moderate at the mature state.
Few agents pay for anything today; the need grows only if they start to. In the transition: probably easing.
- Capability stalls, the rules stay unsettled: eases. Agents that cannot finish whole jobs do not buy from each other.
- Agents deal in markets for people and firms: tightens. Agents that buy have to pay.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
5.6Liability and risk transferbusiness 6Someone solvent to stand behind an agent's work: probably tight now, probably severe at the mature state.
Insurers are excluding AI from standard cover while firms put more work in agents' hands. In the transition: probably tight.
- Agents deal in markets for people and firms: tightens. A firm that lets an agent deal for it wants someone standing behind the agent.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
7.2Functional enterprise applicationsA place in a department's daily work: probably slack now, probably easing at the mature state.
The products are plentiful; what tightens slightly is room in a department's routine. In the transition: probably easing.
No future here is judged to move it.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
7.3Vertical workflow softwarebusiness 13A trade's records and approvals: probably easing now, probably moderate at the mature state.
A general model can draft the work but cannot hold the trade's records or sign for it. In the transition: probably moderate.
No future here is judged to move it.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
7.4Autonomous servicesbusiness 15A licence, and someone accountable for the result: probably easing now, probably tight at the mature state.
Licences are limited by law, and accountable firms cannot be created quickly. In the transition: probably moderate.
- Steady progress, courts and insurers price the risk: tightens. Buyers want one accountable party, and there are few.
- Steady progress, oversight comes first: tightens. Oversight favours firms that are already accountable, and there are few of them.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
7.6AI-native rollupsbusiness 14Owners ready to sell, and clients on fixed fees: probably easing now, probably moderate at the mature state.
The firms worth buying are finite, and buyers are arriving. In the transition: probably moderate.
No future here is judged to move it.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
7.7Consumer AIA daily habit: probably slack now, probably easing at the mature state.
Assistants are free and plentiful; the habit a few products hold is what is hard to win. In the transition: probably slack.
No future here is judged to move it.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
Ground that holds
About as short later as it is today.1.3Semiconductor supply chainCapacity at the newest chip-making processes: moderate now (tonight’s reading of Chip foundries, a low-confidence score), probably moderate at the mature state.
A plant for the newest process takes years to build and longer to run well, so capacity grows without catching demand. In the transition: probably moderate.
- The money stalls, the rules stay unsettled: eases. Orders for chips are cut.
- AI automates research, the rules stay unsettled: tightens. The chips for that computing have to be made.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading, except the word for now, which is tonight’s reading of Chip foundries, the nearest gauge the site scores: what is scarce, each word, the reason and each future's move.
2.1General-purpose modelsA lead over the next-best model: easing now (tonight’s reading of Building frontier models), probably easing at the mature state.
Open models chase the frontier within months, so a lead lasts only while it is renewed. In the transition: probably easing.
- Capability stalls, the rules stay unsettled: eases. With no new frontier, the followers catch up.
- AI automates research, the rules stay unsettled: tightens. When each model speeds up the next, the leading labs pull away.
- AI automates research, oversight comes first: tightens. The leading lab sits on the committee that slows the race, so its lead is kept.
- Open models are good enough for most work: eases. Where an open model is good enough, the lead earns less.
- Labs charge for their lead and move up the stack: tightens. Customers built into a lab's products cannot easily leave it.
- Governments decide which models a firm may use: tightens. Inside each fence fewer makers are allowed to sell, so those that are face less competition.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading, except the word for now, which is tonight’s reading of Building frontier models, the nearest gauge the site scores: what is scarce, each word, the reason and each future's move.
2.2Modality-native modelsData for sound, images and science: probably easing now, probably easing at the mature state.
General models keep absorbing other kinds of input, which leaves a specialist's lead narrow and steady. In the transition: probably easing.
No future here is judged to move it.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
2.3Specialized modelsA field's own data and trust: probably easing now, probably easing at the mature state.
A field's data is hard to gather, but each better general model needs less of it. In the transition: probably easing.
No future here is judged to move it.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
2.4Model compression and adaptationThe skill of shrinking a model without losing quality: probably easing now, probably easing at the mature state.
As more work moves to cheaper and smaller models, the skill is wanted more before tools spread it. In the transition: probably moderate.
- Everyday work moves onto devices: tightens. Shrinking a model to fit the device is the work.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
3.2Data rights and exchangeRights to archives nobody else holds: probably moderate now, probably moderate at the mature state.
An archive is licensed and stays its owner's, so the shortage neither clears nor grows. In the transition: probably moderate.
No future here is judged to move it.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
3.3Data transformationClean, traceable data from live work: probably easing now, probably easing at the mature state.
The cleaning is being automated by the same models it feeds. In the transition: probably moderate.
No future here is judged to move it.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
3.4Synthetic dataLittle: whoever owns the model can make it: probably slack now, probably slack at the mature state.
A model's owner can generate it at will, so nothing here stays short. In the transition: probably slack.
No future here is judged to move it.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
4.3Routing and gatewaysbusiness 8business 18A record of which choice worked: probably slack now, probably slack at the mature state.
Open tools, the clouds and the labs give the choosing away; only the record of where traffic went is hard to copy. In the transition: probably slack.
- Governments decide which models a firm may use: tightens. Keeping a firm inside the models it is allowed is work the free tools do not do.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
4.6Tools and actuationbusiness 10Permission into the software that holds a firm's records: probably moderate now, probably moderate at the mature state.
Each system has an owner who decides who gets in, and that does not change quickly. In the transition: probably moderate.
- Agents deal in markets for people and firms: tightens. Dealing across firms needs a way into each firm's systems.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
6.3Managed AI operationsPeople to keep agents running well: probably moderate now, probably moderate at the mature state.
Running agents is new work that peaks while firms are still learning it. In the transition: probably tight.
No future here is judged to move it.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
7.1Horizontal copilotsLittle: the labs sell their own: probably slack now, probably slack at the mature state.
The labs sell assistants directly, and one is much like another. In the transition: probably slack.
No future here is judged to move it.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
Ground that eases
Shorter today than at the mature state.1.2Data-centre infrastructureHeavy electrical gear, and the trades who install it: tight now (tonight’s reading of Data-centre buildings), probably moderate at the mature state.
Factories for the gear and trained crews can be added in a few years, faster than connections to the power grid. In the transition: probably tight.
- The money stalls, the rules stay unsettled: eases. Less is built, so the gear and the crews are easier to get.
- AI automates research, the rules stay unsettled: tightens. The buildings and their power gear have to be built for that computing.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading, except the word for now, which is tonight’s reading of Data-centre buildings, the nearest gauge the site scores: what is scarce, each word, the reason and each future's move.
1.4Memory, storage and networkingThe fast stacked memory that sits beside an AI chip: probably moderate now, probably easing at the mature state.
Memory is a cyclical business whose makers add capacity when prices rise. In the transition: probably moderate.
No future here is judged to move it.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
1.5Compute providersRentable computing at a price that works: probably moderate now, probably easing at the mature state.
Many firms are building capacity to rent, and what they rent out is much alike. In the transition: probably moderate.
- The money stalls, the rules stay unsettled: eases. Capacity built for demand that did not pay is rented cheaply.
- AI automates research, the rules stay unsettled: tightens. Demand for computing outruns what is being built.
- Open models are good enough for most work: tightens. The cost moves to the computing that runs the model.
- Everyday work moves onto devices: eases. Everyday work leaves the data centre.
- Governments decide which models a firm may use: tightens. A state that fears being cut off runs models at home, which needs computing there.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
3.1Data originationbusiness 7Experts whose judgement can be checked: probably moderate now, probably easing at the mature state.
Experts in narrow fields are hard to recruit in quantity; the need falls as models learn from their own use. In the transition: probably moderate.
- The money stalls, the rules stay unsettled: eases. A lab short of money buys less expert work.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
3.5Training environmentsRealistic practice tasks that can be graded: probably moderate now, probably easing at the mature state.
Many new sellers have entered, and building environments is a skill that spreads quickly. In the transition: probably moderate.
No future here is judged to move it.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
3.7Post-training and learning infrastructurebusiness 3business 19A model that keeps learning without forgetting: probably tight now, probably moderate at the mature state.
Deployment makes the lack of learning felt before research closes the gap. In the transition: probably tight.
- Open models are good enough for most work: tightens. Tuning is where a free model is made to fit.
- Governments decide which models a firm may use: tightens. A firm shut out of the closed models has an open one tuned instead.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
4.1Inference supplyLittle: serving is priced like a commodity: easing now (tonight’s reading of Running models), probably slack at the mature state.
Prices for running a model have fallen steadily, and many firms offer the same service. In the transition: probably easing.
- Everyday work moves onto devices: eases. Work done on the device is not bought from a host.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading, except the word for now, which is tonight’s reading of Running models, the nearest gauge the site scores: what is scarce, each word, the reason and each future's move.
4.2Inference optimizationEngineers who can make serving cheaper: probably moderate now, probably easing at the mature state.
Each new model and chip needs the work done again, though the methods are published and copied. In the transition: probably moderate.
- Open models are good enough for most work: tightens. Running a free model cheaply is the skill left to compete on.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
4.5Agent runtimesLittle: the plumbing is mostly free: easing now (tonight’s reading of Agents and tool use), probably slack at the mature state.
The labs and open projects give the software away. In the transition: probably easing.
No future here is judged to move it.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading, except the word for now, which is tonight’s reading of Agents and tool use, the nearest gauge the site scores: what is scarce, each word, the reason and each future's move.
6.1Technical implementationbusiness 11Engineers who can fit AI into a firm: probably tight now, probably easing at the mature state.
Much of the fitting is done once at each firm, and tools and agents take over more of it. In the transition: probably tight.
- Capability stalls, the rules stay unsettled: tightens. The gains come from fitting what exists, so the people who do it stay short for longer.
- Steady progress, the rules stay unsettled: tightens. Adoption at the pace of past technologies means years of fitting work.
- AI automates research, the rules stay unsettled: eases. Agents do more of the fitting themselves.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
6.2Context engineeringbusiness 12A firm's unwritten rules, written down: probably tight now, probably moderate at the mature state.
The knowledge sits in people's heads, and models slowly get better at learning it from records. In the transition: probably tight.
- Capability stalls, the rules stay unsettled: tightens. A model that will not learn a firm's habits needs them written down.
- Steady progress, the rules stay unsettled: tightens. Slow adoption keeps the knowledge in people's heads for longer.
- AI automates research, the rules stay unsettled: eases. Better models learn a firm's habits from its records.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
6.4Adoption and changeManagers willing to rebuild work around agents: probably tight now, probably moderate at the mature state.
Firms change at the pace of their people, which is slower than the models. In the transition: probably tight.
- Capability stalls, the rules stay unsettled: tightens. Spreading what already works is the whole of the gain.
- Steady progress, the rules stay unsettled: tightens. Adoption at the pace of past technologies keeps the reorganising of work the limit for longer.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
6.5Talent productionbusiness 17People trained to deploy AI: probably tight now, probably easing at the mature state.
People can be trained faster than firms can be rebuilt, so this shortage passes first. In the transition: probably moderate.
- Capability stalls, the rules stay unsettled: tightens. Spreading what already works takes people who can deploy it.
- Steady progress, the rules stay unsettled: tightens. Years of fitting work need people who can deploy.
- AI automates research, the rules stay unsettled: eases. There is less deployment work left for people.
The companies the site places here, each with a link to have Ask reason it through.
A model's judgement, not a reading: what is scarce, each word, the reason and each future's move.
Where the shortage goes next
The site's migration page predicts where the shortage sits now and where it moves, and tests each prediction nightly where it can. Each prediction and its state tonight are the site's own; the line tying it to businesses is a model's judgement.
- NowThe site's test: holding tonight
Data-centre buildings and their power hookups are the bottleneck.
While buildings and their power hookups bind, a business that finds where computing can run is selling into the shortage. A model's judgement.
On the map: 1.1 Power and sites, 1.2 Data-centre infrastructure.
Businesses that sell into it: 16 Moving computing to where power is available.
- NowThe site's test: holding tonight
The site reads money as plentiful, so money alone is no advantage; the businesses that need most of it are weakened in the future where funding stalls. A model's judgement.
- NowThe site's test: cannot be tested yet
Memory and chip packaging are short.
No business here sells into memory or packaging; if they are short, the makers of memory and the firms that package chips are paid. A model's judgement.
On the map: 1.4 Memory, storage and networking, 1.3 Semiconductor supply chain.
- NextThe site's test: failing tonight
If the shortage moves here, graded expert work and the means of grading it are what the labs buy. A model's judgement.
On the map: 3.5 Training environments, 3.1 Data origination, 3.6 Reward and verification systems.
Businesses that sell into it: 7 Graded expert-work records pooled across buyers.
- WatchThe site's test: cannot be tested yet
If proof of what an agent did becomes the bottleneck, these businesses are the ones that sell it; the site cannot yet test whether it has. A model's judgement.
On the map: 5.4 Outcome verification, 5.3 Provenance and audit, 5.2 Identity and authority, 5.5 Settlement and billing, 5.6 Liability and risk transfer.
Businesses that sell into it: 1 Outcome verification with pay-per-verified-result billing; 4 Know-your-agent and delegated-authority registry; 5 Tamper-proof audit trail for what agents did; 6 AI warranty backed by a pool of loss data; 9 Escrow that pays an agent only when its work passes a check; 20 Incident response and forensics for agent failures.
What a durable business builds
These are five things the site's owner holds that a better model does not replace. Building one is not enough to keep a profit: the calls above say which businesses do. Which business builds which is a model's judgement.
| It builds | What that is | Why it lasts | Businesses that build it, a model's judgement |
|---|---|---|---|
| The record | The customer's own state, permissions, traces and results. | Models and suppliers change; the firm's memory has to stay. | 5 Tamper-proof audit trail for what agents did; 4 Know-your-agent and delegated-authority registry; 13 A trade's AI system of record, with the approvals only that trade can give; 21 A firm's own map of what its data means, kept on its own servers for every agent to read; 3 A model that keeps learning inside one customer's work; 11 Deployment that leaves behind the platform the customer runs on; 10 The standard way agents get permission into business software; 12 Capturing how a firm really works, for its agents |
| The test | The customer's own definition of good work, checked against real cases. | Better output raises the need to know whether it is good. | 1 Outcome verification with pay-per-verified-result billing |
| The backing | Someone accountable, an audit trail, a warranty. | Work with consequences needs a solvent party behind it. | 2 Independent sign-off for AI in regulated work; 5 Tamper-proof audit trail for what agents did; 4 Know-your-agent and delegated-authority registry; 13 A trade's AI system of record, with the approvals only that trade can give; 15 A licensed firm that sells the finished job and stands behind it; 6 AI warranty backed by a pool of loss data; 20 Incident response and forensics for agent failures; 9 Escrow that pays an agent only when its work passes a check |
| The yardstick | Cost, speed, quality and reliability measured on the same terms. | Buyers keep asking for proof that things improved. | 1 Outcome verification with pay-per-verified-result billing; 9 Escrow that pays an agent only when its work passes a check; 8 A router that chooses models from pooled outcomes |
| The choice | The freedom to switch models, places to run and suppliers. | Work differs and suppliers change. | 21 A firm's own map of what its data means, kept on its own servers for every agent to read; 18 Routing that keeps a firm inside approved models and fails over when one is cut off; 16 Moving computing to where power is available; 10 The standard way agents get permission into business software; 8 A router that chooses models from pooled outcomes |
The planning assumptions, not readings: what the calls take for granted
The planning assumptions of the site's owner, reworded by a model. They are what the calls on this page take for granted, not readings and not findings. Where one follows a writer the site introduces elsewhere, it says so.
About customers
- Customers want the problem solved, not another tool. Exception: Technical buyers who want control. Weakened by: Self-serve tools outgrowing done-for-you services in the same trade.
- Customers want fewer suppliers to hold accountable. Exception: Where the best part from each supplier is worth the trouble. Weakened by: Firms adding AI suppliers faster than they consolidate them.
- The way in is the customer's most pressing problem. Exception: Where a regulator dictates the order. Weakened by: Products that began with a minor convenience taking the main budget later.
- When something important fails, customers want someone solvent to put it right. Exception: Low-stakes work. Weakened by: Buyers accepting agents with no named party behind them.
- Customers want to be able to leave a model or supplier, even when they stay. Exception: When a supplier is far ahead. Weakened by: Firms giving up the means to switch: a single lab's tools throughout, and no other supplier tested.
- Customers expect better economics over time, though the bill may still rise if they use more. Weakened by: Prices holding while quality stalls.
About the systems
- Producing an answer and proving it right stay separate jobs. Weakened by: Agents becoming cheap and reliable at checking their own work.
- What a firm knows is local and goes stale, so it has to be kept fresh and tied to results. After Friedrich Hayek. Weakened by: Agents learning a firm's unwritten rules from its records alone.
- An agent's right to act stays limited: capability does not confer permission. Weakened by: Firms letting agents sign, pay and commit with no limit set, and no losses following.
- Physics still matters: energy, delay, memory and distance limit where AI can run. Weakened by: Efficiency gains outrunning demand for years.
- Different work needs different models and places to run. Weakened by: A single supplier being best on price and quality for most tasks.
About markets and institutions
- Adoption lags capability: budgets, rules and trust move slower than the models. After Arvind Narayanan and Sayash Kapoor. Weakened by: Firms' use of AI rising as fast as the models improve.
- Competition moves the shortage and does not remove it: when a part becomes plentiful, the profit moves to what is still scarce. After the site's migration essay, which follows Eliyahu Goldratt and David Ricardo. Weakened by: Profits falling across the whole chain at once.
- Responsibility has to land on someone. Weakened by: A court or regulator treating an agent's act as nobody's.
- Customers differ, so shared infrastructure will sit beside local set-up and judgement. Weakened by: A single configuration serving most firms in a trade.
- Evidence has to change decisions: a rating with no source, no test and no review date becomes decoration. Weakened by: Calls on this page being kept after the readings under them have moved.
The table: every business in every future, each cell a model's judgement
| Business | Capability stalls, the rules stay unsettled | The money stalls, the rules stay unsettled | Steady progress, the rules stay unsettled | Steady progress, courts and insurers price the risk | Steady progress, oversight comes first | AI automates research, the rules stay unsettled | AI automates research, oversight comes first | Open models are good enough for most work | Labs charge for their lead and move up the stack | Everyday work moves onto devices | Agents deal in markets for people and firms | Governments decide which models a firm may use |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 2 Independent sign-off for AI in regulated work | unchanged | unchanged | unchanged | stronger | stronger | unchanged | stronger | unchanged | stronger | unchanged | unchanged | unchanged |
| 1 Outcome verification with pay-per-verified-result billing | unchanged | unchanged | unchanged | stronger | unchanged | stronger | unchanged | unchanged | unchanged | unchanged | stronger | unchanged |
| 5 Tamper-proof audit trail for what agents did | unchanged | unchanged | unchanged | stronger | stronger | stronger | unchanged | unchanged | unchanged | unchanged | stronger | unchanged |
| 4 Know-your-agent and delegated-authority registry | weaker | unchanged | unchanged | stronger | stronger | unchanged | unchanged | unchanged | unchanged | unchanged | stronger | unchanged |
| 13 A trade's AI system of record, with the approvals only that trade can give | unchanged | unchanged | unchanged | stronger | unchanged | weaker | unchanged | unchanged | weaker | unchanged | unchanged | unchanged |
| 15 A licensed firm that sells the finished job and stands behind it | weaker | weaker | unchanged | stronger | stronger | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged |
| 21 A firm's own map of what its data means, kept on its own servers for every agent to read | stronger | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | stronger | weaker | unchanged | unchanged | stronger |
| 3 A model that keeps learning inside one customer's work | stronger | unchanged | unchanged | unchanged | unchanged | weaker | unchanged | stronger | weaker | stronger | unchanged | unchanged |
| 11 Deployment that leaves behind the platform the customer runs on | stronger | unchanged | stronger | unchanged | unchanged | weaker | unchanged | unchanged | weaker | unchanged | unchanged | unchanged |
| 6 AI warranty backed by a pool of loss data | unchanged | unchanged | unchanged | stronger | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | stronger | unchanged |
| 18 Routing that keeps a firm inside approved models and fails over when one is cut off | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | stronger |
| 20 Incident response and forensics for agent failures | unchanged | unchanged | unchanged | stronger | stronger | stronger | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged |
| 14 An AI roll-up that moves its clients to fixed fees | unchanged | weaker | stronger | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged |
| 16 Moving computing to where power is available | weaker | weaker | unchanged | unchanged | unchanged | stronger | weaker | stronger | unchanged | weaker | unchanged | unchanged |
| 9 Escrow that pays an agent only when its work passes a check | weaker | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | stronger | unchanged |
| 10 The standard way agents get permission into business software | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | weaker | unchanged | stronger | unchanged |
| 7 Graded expert-work records pooled across buyers | unchanged | weaker | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged |
| 8 A router that chooses models from pooled outcomes | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged | stronger | weaker | unchanged | unchanged | weaker |
| 12 Capturing how a firm really works, for its agents | stronger | unchanged | stronger | unchanged | unchanged | weaker | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged |
| 19 Tuning open models to frontier quality for one industry | unchanged | unchanged | unchanged | unchanged | unchanged | weaker | unchanged | stronger | unchanged | unchanged | unchanged | stronger |
| 17 Training and placing the engineers who deploy AI | stronger | unchanged | stronger | unchanged | unchanged | breaks | unchanged | unchanged | unchanged | unchanged | unchanged | unchanged |
A model's judgement, not a reading. Made 2026-10-09 by Claude, a model made by Anthropic (claude-opus-5-5). One model read each published business beside the site's readings of the value chain and each future the outlook's grid fills, gave the business a call with one line of reasoning and the one thing that would undo it, and said in words what each future does to it. A future it left out changes nothing. The same model then said, for each part of the map in scope, what is scarce there, how short it is now, in the transition and at the mature state, and which futures tighten or ease it, each with a reason. The second group of futures are each a model's summary of what named writers on the outlook argue, and were judged the same way. It also tied each of the site's migration predictions to parts of the map and to businesses, and said which of the owner's lasting things each business builds. Reviewed by separate reviewing runs of the same model, which changed calls, takes, words and reasons, withdrew some effects and moves and added others; no person reviewed each one. That covers every call, effect, word and move on this page, except the predictions and their states under “where the shortage goes next”, which are the site's own and tested nightly, and the lasting things and the planning assumptions, which are the site's owner's. No status, tally or answer on the site reads a judgement made here.