AI’s New Economics
What Keeps Its Value — and Who Captures It
Globalization reset the market for reproducible physical labor; AI is doing the same for cognitive labor. The fear is rational. But AI does not erase pricing power uniformly; it relocates it. Where that power goes, and who captures the resulting margin, is the whole question. The answer runs in three parts:
- AI makes repeatable cognitive labor abundant. Work that can be procedurally reproduced has its price competed downward; judgment and the tacit remainder — which are harder and costlier to reproduce — retain their value.
- In an abundant-supply world, demand is the scarce thing. Value shifts to whoever owns the customer relationship — the aggregators racing to control the agent interface — while every layer beneath risks being commoditized.
- The defense is sovereignty. Own the operating memory you depend on, make models contestable, and keep direct relationships with your customers — so that no supplier controls your memory, no lab makes you dependent, and no aggregator takes your market.
This is the map for where value flows; the sections below walk it.
The supply shock
The mechanism is a growing abundance of scalable cognitive labor. AI supplies “employees” that can be replicated quickly, often cost less per unit of well-specified work, and have generally become more capable over time.
The line is reproducibility, not seniority.1 The pressure reaches the paralegal drafting the contract and the senior associate reviewing it, the bookkeeper closing the ledger and the analyst building the model. Wherever work is well-specified, digitally legible, and procedurally reproducible — however senior the hands doing it — a machine can increasingly perform more of it at lower cost, competing away its price.2 This looks less like a temporary wave than a new sea level.
Within cognitive work, two sources of scarcity provide safe harbor because they are hard to reproduce. Judgment and taste are the first — knowing what’s worth doing, telling good from good-enough, and standing behind the result. AI can provide the cognitive labor to do the work. However, it cannot yet decide what the work should be, when it is done, whether it is done well, or own the mistakes.
Tacit knowledge is the second form of safe harbor — the hard-won expertise that lives in the hands and habits of the people and teams doing the work. Parts of it can be elicited, validated, and converted into operating memory; what remains embodied is harder and costlier to reproduce. TSMC is the standing example: its advantage cannot be reduced to published process descriptions. It includes proprietary recipes, yield learning, supplier coordination, and accumulated process know-how no rival can copy from a blueprint.3
As AI commoditizes the reproducible and the explicit, judgment and the tacit remainder grow more valuable, not less, because they remain harder to replicate. When anyone can generate a competent draft, the scarce skill is knowing which draft is right and standing behind it. Value flows to whatever stays scarce.4
That is where the cognitive-supply story ends: two difficult-to-reproduce inputs remain scarce.5 But surviving a flood is not the same as getting paid for it.
The demand grab
The most valuable position in an AI economy isn’t on the supply side at all. It’s owning demand — and the story starts, as Ben Thompson’s aggregation theory did, with text.6
The internet made publishing nearly free: for the first time, anyone could reach the whole world without owning a printing press or a distribution deal, and the supply of text stopped being scarce. So the value moved from producing the writing to controlling the demand for it; Google aggregated the world’s readers by owning discovery. In a flood of free text, search was how you found anything worth reading, so every publisher had to compete for that attention on Google’s terms. The publishers were commoditized; Google captured the value. (Amazon did the same to merchants, Uber to drivers.) Once supply is abundant, whoever owns the demand owns the game.
AI is now doing to many standardized forms of cognitive work what the internet did to text: collapsing the cost of supply and handing more power to whoever aggregates demand. And it can go further this time. Google still needed publishers to produce the writing; an aggregator drawing on abundant, lower-cost cognitive work needs fewer particular suppliers. Being aggregated used to mean thin margins. Increasingly, it means less leverage.
Who owns demand today? The incumbents: Google owns search, Apple the phone, Microsoft the enterprise. But the agent interface puts those positions in play: it is a new discovery layer — how buyers find the right output amid abundant cognitive work. If many standardized tasks begin routing through agent-mediated conversations, incumbent search boxes and home screens stop being where some demand forms — and interaction shifts are how aggregators fall. Desktop-to-mobile cost Microsoft the consumer market; search-to-agent could cost Google part of the search market.
This economy-spanning opportunity is why the AI labs, already running leading agent interfaces, have a real shot at owning demand. And they have existential reasons to take it: their models are commoditizing under them, so their escape is to stop being an input to someone else’s software and become the interface more cognitive tasks run through.
The incumbents are running the same race from the other side, wiring models into every demand surface they own. Whoever wins pairs a capable model with a dominant agent interface — the model is the commoditizing half, the interface the scarce one — to aggregate demand across many cognitive tasks. A model-plus-aggregator, built from either direction, gains the power to reprice layers beneath it.7
So the supply-side threat is that AI makes your work cheap. The demand-side threat is that the winner of the interface race owns your customer — and a customer the aggregator owns is one it can cut you off from. Which raises the question the rest turns on: if owning demand is how value gets captured, how does anyone who isn’t the aggregator keep a share?
The defense: sovereignty through ownership
The answer is ownership — the path to sovereignty, the freedom to keep operating, or switch, without a supplier’s leave. You already have it over your phone number: you can carry it to any carrier, so no carrier can hold it hostage — they compete on price and service instead. Note that you don’t run your own cell towers; you depend on a carrier completely, you’re just not trapped by one. Most companies have no such freedom over the data they run on.
Ownership matters because promises alone do not neutralize dependence. Put your data in a supplier’s system, train your team on it, build workflows around it, and the supplier gains leverage because leaving is hard.8 Contracts help, but they are incomplete and eventually reach renegotiation. Architecture and tested portability make the commitment credible: hold what you depend on in a form you can move, and the trap becomes materially weaker. Custody is one barrier cheap agents cannot route around, which is why ownership is more than exit tooling alone.9
Sovereignty has three parts, one for each way a layer above can capture you:
- Own your operating memory — your data and, now, the records, rules, workflow state, evals, examples, and accumulated judgment your agents need to act correctly. Owning it does not mean running your own servers, any more than owning your phone number means running cell towers; it means keeping it in a form you can pick up and move. Never let a supplier control your memory.
- Make models contestable — Models and harnesses should be replaceable components pointed at your operating memory, where your evals and operating specification live under your control. Never become dependent on one lab or runtime.
- Keep your customer — You must keep a direct relationship with your customer — the aggregator’s discovery layer can be a channel, never your only path to them. Never give the aggregator the opportunity to take your market.
Buyers at every level must fight for this sovereignty: people own their context, companies their data, vendors their judgment and their customers’ trust. Every layer buys from the one above and sells to the one below, and the same question returns at each — what must I own so the layer above can’t reprice me?
Why would any supplier go along with this? Because giving up lock-in and enabling sovereignty can raise its profit — provided the value was ever the work, not the captivity. The threat of the trap is what keeps customers from committing.10 And because every company needs its data back and its models swappable, buyer demand can turn “hand it over” from a concession into a market. Incumbents such as Oracle, Salesforce, and Bloomberg combine switching costs with proprietary data, network effects, workflow embedding, and institutional standing; AI attacks the labor component of leaving, not the whole moat.11
Ben Thompson is an illustration. He worked out aggregation theory by watching Google commoditize publishers — then refused to be one. Stratechery runs on his own site, for subscribers he serves directly, rather than renting both operation and audience from the Times or Substack. The move generalizes: keep the operating assets and customer relationship that an aggregator would otherwise control.
The AI labs: the hardest case
The AI labs are the hardest case for this defense: they hold unusually concentrated leverage and have limited ability to commit to buyer sovereignty. Sovereignty still shields you from a lab, but it cannot make a lab structurally trustworthy. A customer can own its operating memory and still depend on a provider for the live model, continuing improvements, serving capacity, and access policy.
Open-weight models — increasingly including strong models from Chinese labs — soften this. Weights you host cannot be revoked by the publisher, which makes them a useful outside option.12 But the escape is partial: your copy stops improving when the publisher stops publishing, and a regulated U.S. buyer adopting a Chinese model takes on provenance, security, supply-chain, and regulatory risks. Lab promises therefore remain among the least credible in the stack, where the temptation to reprice or restrict access is largest.
The problem follows the position, not the pedigree. An incumbent that adds a model can inherit the same tension: a franchise at risk is a reputational incentive, not a structural commitment, and the point of this defense is to stop accepting incentives as substitutes for control.
This isn’t hypothetical. In June 2026, Anthropic’s Fable 5 release showed the risks posed by the labs in three ways at once:13
- access to covered models required a retention-enabled workspace or environment with thirty-day retention, including for organizations that otherwise used zero-data-retention environments;
- broad “safeguards” that could block or reroute legitimate work; and
- within days, a suspension of access while Anthropic implemented a U.S. government directive restricting foreign nationals.
Access was later restored, but two parties — Anthropic and the U.S. government — each held an off-switch the customer did not. Two prominent observers framed the issue similarly: Palantir’s CEO went on CNBC describing enterprise customers “livid” at the frontier labs — demanding to know who controls the weights, data, and alpha of their business — while Ben Thompson called the retention change the most underrated aspect of the release.14
So labs cannot be disciplined from inside alone. They are checked from outside: by customers that keep their own data and stay able to switch models and harnesses, by firms that keep the customer relationship out of a lab-owned interface, by real-world inputs the labs cannot manufacture (compute above all), and by government. The same principle that protects every other layer helps here too: preserve control over the parts of the operation that are not free to copy.
What this predicts
Pricing power tends to collect at the two scarce ends — judgment and tacit knowledge on one end, ownership of demand on the other — and around constrained real-world inputs. It drains from the reproducible middle. Individuals appear most exposed. Companies split: those anchored in scarce physical assets, regulated standing, or embedded relationships hold more leverage, while purely digital firms — where value is reproducible and demand is mediated — face greater hollowing risk. Software vendors that hoard data are more exposed to absorption;15 those that hand it back and charge for work are better positioned. And labs and incumbents keep fighting for the agent interface, so their conflicts with customers and governments recur rather than settle.
Three honest limits. Timing is uncertain — the direction appears clearer than the speed, which depends on adoption, regulation, and compute. The safe ground shifts — when robots flood physical labor, the protection real-world work enjoys today erodes. And the defense is a floor, not a fortress: it protects margins and continuity, not immunity. A consumer-facing firm can do all three and still watch demand migrate to conversations it is not in — its customers are people, the most exposed layer. What the defense buys there is being the name the customer carries in their own context.
The bottom line: sell what stays hardest to reproduce — judgment, taste, and the tacit remainder; the reproducible middle loses pricing power. Then defend what you sell:
- Own your operating memory — your data and your agents’ accumulated judgment, in a form you can pick up and move.
- Make models contestable — replaceable models and harnesses pointed at your operating memory; use any lab or runtime, depend on none.
- Keep your customer — whoever owns the demand owns the game.
And charge for service and production rather than captivity or access to the customer’s own data. Pair customer-owned memory with production-priced services, and leverage the new rather than automate the old.16
This is the macro map. AI Agents Are the New Employees shows how fragmented company context becomes a recurring operating cost; Own Your Operating Memory defines the first defense and its ownership test. Seller- and buyer-side playbooks follow. Disclosure: I build IndustryVault, whose architecture reflects the ownership model argued for here.
Definitions
- Customer: a relative term — every layer is a customer of its suppliers and a supplier to its own customers. Unqualified, it means the buyer in the pair under discussion; the “customer relationship” the aggregators race for is the far end of the whole chain, where demand aggregates.
- Aggregator: a player that owns demand — the direct customer relationship — and makes suppliers compete for access to it, on its terms. Google to publishers, Amazon to merchants; the model-plus-aggregator is the AI-era form.
- Substrate: the customer-controlled systems, formats, schemas, code, tool contracts, infrastructure, and rights that store, govern, and make operating memory usable and recoverable.
- Operating memory: the governed, actor-specific live state and codified judgment people and agents need to act correctly: records, rules, exceptions, workflows, examples, evals, approvals, history, provenance, and recorded decisions.
- Custody: control over that substrate in a way that lets a supplier deny, degrade, or reprice access.
- Access: reading or using the customer’s existing records.
- Production: creating, correcting, validating, or maintaining authoritative state.
- Sovereignty: the customer’s practical ability to keep operating, or appoint a successor, without the supplier’s permission.
- Model contestability: the ability to replace the model or harness and test a successor against the same owned operating specification, records, evals, and accumulated judgment — without requiring identical behavior or losing operating memory.
- Off-switch: the power to interrupt the customer’s use of an asset they depend on.
Footnotes
Geoffrey Hinton predicted in 2016 that AI would make radiologists obsolete within five years; radiologist employment rose instead. Part of the job (reading the scan) is a repeatable procedure; the rest — judgment, accountability, the production system around it — is not. The reproducibility line runs through jobs, not just between them.↩︎
Cheaper generation is not the same as cheaper delivered work. Verification, integration, security, accountability, context assembly, and distribution remain real costs — and they are where scarce complements and margin relocate. The reproducible portion loses pricing power; the parts that decide, validate, and stand behind the result keep it.↩︎
“Tacit knowledge” is Michael Polanyi’s “we know more than we can tell.” Once know-how can be reliably articulated, validated, and governed, it becomes operating memory; the embodied remainder stays tacit. TSMC’s own annual reporting treats process technology, manufacturing know-how, and trade secrets as proprietary assets. Its advantage combines those assets with yield learning, supplier coordination, and accumulated process skill — not public blueprints alone.↩︎
Economists call this the “smiling curve,” after Stan Shih of Acer (1990s): value high at design and at the customer end, low in the manufacturing between.↩︎
Physical work is also safe for now — but only because this flood is cognitive. When robots flood the market for hands, the same logic arrives, so treat that safety as borrowed.↩︎
Ben Thompson, “Aggregation Theory” (Stratechery, 2015) and “Defining Aggregators” (2017): once the internet drove distribution costs to zero, power shifted from controlling supply to owning demand — and aggregators capture demand by owning discovery over an abundance of supply.↩︎
Not everything collapses into one interface. The grab bites hardest on standardized tasks a buyer will delegate, where demand flows through the interface and nothing about the supplier’s brand, regulation, or operational standing pulls the customer back. Regulated, relationship-heavy, operationally embedded buying stays multi-homed. The claim isn’t that every customer relationship moves to the agent interface — only that enough task demand does that vendors whose value is reproducible and whose customers are mediated lose leverage.↩︎
The “hold-up problem” (Klein, Crawford & Alchian 1978; Williamson). Ownership as the fix: Grossman, Hart & Moore; Oliver Hart shared the 2016 Nobel for the idea.↩︎
More precisely, three switching costs. Labor (migration, schema translation, integration rewrites) is what AI attacks first. Custody (the vendor denies, throttles, or reprices access) is what ownership attacks. Operational costs (risk, downtime, compliance, retraining) remain legitimate friction. Exit never becomes free; vendors just can’t hide custody rent inside labor friction and call the whole bundle inevitable.↩︎
Farrell & Gallini (1988): a supplier can raise its own profit by deliberately creating the customer’s escape route. Open-source software runs on the same logic (Lerner & Tirole 2002).↩︎
“Economic moat” is Warren Buffett’s coinage; switching costs were canonized as one of the seven durable powers in Hamilton Helmer’s 7 Powers (2016) and became standard venture diligence.↩︎
DeepSeek, Qwen, and their successors have kept open weights within months of the frontier — the engine of the commoditization the labs are fleeing. For a regulated U.S. buyer they discipline more than they deploy: provenance, security review, and possible restrictions on Chinese models complicate adoption, and continued publication is a publisher’s choice, not a commitment. The structural fact survives either way: weights you hold cannot be taken back.↩︎
Anthropic announced Claude Fable 5 and Claude Mythos 5 on June 9, 2026. Its Covered Models retention policy required thirty-day retention for prompts and outputs in environments using covered models; an organization using zero data retention had to use a retention-enabled workspace, subscription, or cloud environment, and could isolate that use in a separate sandbox. Anthropic’s June 12 statement said it had received a U.S. government directive to suspend access to Fable 5 and Mythos 5 for foreign nationals, leading it to suspend access while implementing controls; its redeployment statement said the export controls were lifted on June 30 and access restored July 1.↩︎
Alex Karp on CNBC, discussing Palantir’s partnership with Nvidia to bring the open-source Nemotron model to the U.S. government: enterprises want “control over their compute, their models, their data stack and their alpha — they want to know they own the means of production.” Ben Thompson (Stratechery) on the Fable retention change: Anthropic “didn’t put in any sort of safeguards to guarantee they wouldn’t” train on the retained data — and the labs have “an economic imperative to move up the stack and own the customer directly.”↩︎
How absorption works, spelled out. The aggregator sits at the interface, so it sees demand for what you do; software production is getting cheaper, so it can rebuild more of your function natively on the surface the customer already uses. A vendor whose revenue rests on custody is especially exposed when its function is reproducible. The only thing still holding the customer may be their data in the silo, a barrier hostile to the customer, who therefore helps break it by demanding portability at renewal or letting agents reconstruct the state. Revenue priced on captivity can break quickly when the wall does. The hand-it-back vendor is better positioned because there is less to seize: its price was production, its defense is work interface traffic does not teach, and agents can compose with it instead of tunneling through it.↩︎
The distinction is Alan Kay’s, by way of Arthur Koestler: a new technology first automates the old form — the horseless carriage, the word processor as a faster typewriter — before someone uses it to reach a plane that wasn’t possible before. Hamilton Ulmer frames the working choice crisply as “automate the old, or leverage the new.”↩︎