The Economics of Intelligence.
Jared Flatow
Draft
Imagine a software project that costs $100,000 to build. A small team of developers spends months producing it. Then AI comes along, and suddenly one developer can complete the same project for $5,000.
What happens to the other $95,000?
The developer is vastly more productive than before. But so are their competitors. And increasingly, so are their customers.
Production costs have collapsed, but prices don't necessarily follow. The developer might keep charging $100,000. Competitors might undercut them. Or the customer might simply build the software themselves.
The developer has become more capable without necessarily becoming more competitive.
For decades, businesses have built advantages that reinforce themselves. Capital finances infrastructure, infrastructure enables scale, scale attracts customers, and customers attract more customers. Network effects and economies of scale make successful businesses increasingly difficult to challenge.
Success compounds. The further ahead a business gets, the harder it becomes to catch.
Intelligence introduces an opposing force.
A developer working with AI can learn unfamiliar techniques. A researcher can discover new algorithms. An engineer can design more efficient computing infrastructure. And increasingly, AI can help produce better AI.
In each case, intelligence increases the capabilities of another intelligence without diminishing its own. What the recipient learns can be applied independently, improved upon, and passed along.
Call this induction: the ability of intelligence to raise the capabilities of other intelligences toward its own level.
Humans have done this through teaching for millennia. Computing has accelerated the process for decades. General-purpose AI makes it possible across an enormous range of domains, at extraordinary speed and scale.
And the recipients can induce further intelligence themselves.
The more valuable intelligence becomes, the greater the incentive to acquire it. And the more capable it becomes, the more effectively it can help others acquire comparable capabilities.
Now consider the companies developing frontier AI.
They invest enormous resources to produce increasingly capable intelligence. The obvious business is to sell access to it.
And that can be enormously profitable. Thousands of businesses might depend on the same provider, paying for capabilities they cannot produce independently.
But unlike electricity, intelligence can directly help its customers acquire the knowledge and capabilities needed to produce what they're buying.
A customer can use AI to discover more efficient algorithms, develop competing models, or contribute to open intelligence. Most may continue paying for convenient access, but some have enormous incentives to become independent.
And their discoveries can benefit everyone else.
A proprietary breakthrough initially benefits its owner. An open breakthrough raises the capabilities available to every competitor simultaneously.
Once a capability becomes widely accessible, no individual provider can restore its exclusivity.
Frontier labs face a dilemma. Selling more capable intelligence generates revenue, usage data, and customer relationships that can finance further development. But it also helps close the gap between themselves and potential competitors.
Keeping intelligence private protects the advantage but sacrifices the benefits of distributing it. Restricting access too aggressively risks losing customers to other frontier labs or increasingly capable open models.
This pushes frontier labs to preserve as much profitable distance as possible between their internal frontier and what competitors can access, while exploiting that distance before it closes.
And increasingly, exploiting it means moving downstream.
Why sell intelligence to a software company when you can use superior intelligence to build the software yourself? Why supply a consulting firm when you can provide the service directly?
Frontier labs are already moving into coding, enterprise applications, and implementation services. They're competing to capture the value of what intelligence produces, not merely the price of accessing it.
And once one lab starts, the others face pressure to follow.
Imagine two labs with comparable internal intelligence. One begins using its private frontier to build products that outperform anything available to the other's customers. The second lab can release more capable intelligence, build competing products itself, or watch its customers lose ground.
Releasing more intelligence accelerates induction. Competing downstream creates another problem.
The more labs compete with their customers, the more incentive they have to withhold their best capabilities and exploit privileged access to customer data, strategies, and operations. Their customers, in turn, have greater reason to distrust their suppliers and develop independent alternatives.
Two competitive pressures reinforce each other. Open intelligence pushes the publicly available baseline upward, while competition between frontier labs drives them further downstream. The resulting conflicts with customers create still more demand for independent intelligence.
No individual lab needs to want this outcome. Competition drives the cycle.
But vertical integration doesn't eliminate induction.
Suppose a frontier lab uses superior intelligence to establish a successful software business. Six months later, open intelligence becomes capable of producing comparable software.
The lab still has its customers, relationships, and accumulated resources. But the capability that gave it an advantage is no longer exclusive.
Its competitors can now reproduce products, improve operations, acquire customers, and organize alternatives using comparable intelligence.
The lab must defend its accumulated advantages or use its next frontier breakthrough to establish new ones.
Accumulation makes success self-reinforcing. Induction makes success reproducible.
There is another way to defend an advantage: prevent competitors from exercising the capabilities they've acquired.
Governments have successfully restricted the proliferation of powerful technologies, including nuclear weapons. But a blueprint is not a nuclear weapon. Producing one still requires specialized knowledge, materials, equipment, and organization.
Intelligence doesn't eliminate the need for physical resources. But it can help discover substitutes, improve efficiency, solve engineering problems, and coordinate alternatives. Unlike a blueprint, intelligence can actively help overcome the obstacles to putting its knowledge into practice.
Restricting access may slow its propagation, but capabilities that have already spread cannot necessarily be recalled. And restrictions impose costs on everyone prevented from using them, creating opportunities for competitors operating elsewhere.
This makes containing intelligence an increasingly difficult proposition, even for institutions willing to sacrifice productive opportunity to maintain control.
Frontier labs will race to advance their internal intelligence, distribute only as much as competition requires, and convert temporary technological superiority into more durable economic positions. Induction pushes competitors toward the frontier, while continued research pushes the frontier further ahead. Both can accelerate simultaneously.
And every advantage they establish faces the same underlying problem.
For decades, the difficulty of building a competitive advantage was also what made it durable.
Intelligence changes that equation. The easier it becomes to build an advantage, the easier it becomes for everyone else to build one too.