For the past several years, the dominant assumption in AI has been remarkably simple.
It has been a general consensus that the intelligence lives in large foundation models and everyone subscribes to their services or falls behind.
Users send prompts to increasingly sophisticated foundation models, receive answers, and pay for the privilege of doing so. Every interaction begins and ends with another inference request.
The business model is equally straightforward: intelligence is rented.
The more I think about where AI is heading, the less convinced I am that this architecture represents the end state of the foundation model industry.
In fact, it may instead eventually resemble the early mainframe era of computing instead.
The assumption embedded in most discussions about AI is that the valuable asset is the model itself. OpenAI, Anthropic, Google, xAI, and others are racing to build increasingly capable frontier systems under the assumption that this centralized model is the basis for their future.
That assumption deserves scrutiny.
The more interesting possibility is that foundation models eventually become teachers rather than workers.
The distinction matters.
Today, when a developer asks a frontier model to solve a problem, the interaction is largely ephemeral. A question is asked. An answer is returned. The transaction ends. Yes, the agents that sit on top of the LLM’s have evolved into magnificent beasts making a huge difference in outcomes and productivity but for now let's focus on the LLM’s themselves. Bear with me on this for now.
But the answer the LLM provides contains considerably more value than just the immediate response itself.
It contains reasoning patterns.
Alternative approaches.
Evaluations.
Critiques.
Design tradeoffs.
Domain knowledge.
In other words, it contains training material.
Once organizations begin systematically capturing, validating, and distilling those interactions, the economics change.
A frontier model interaction is no longer just an inference event.
It becomes an asset creation event.
The Difference Between Renting Intelligence and Owning It
Historically, expertise accumulated inside people.
Organizations attempted to preserve that expertise through documentation, training programs, architecture reviews, design records, and institutional processes.
The problem was that much of the knowledge remained trapped inside individuals.
AI changes this equation.
For the first time, expertise can be transformed into a durable machine-readable asset.
The important shift is not that AI can answer questions.
The important shift is that AI can convert answers into future capability.
Consider what happens when a software architect uses a frontier model to solve a difficult design problem.
Under the current model:
Question → Answer → Done
Under a distillation model:
Question → Answer → Validation → Training Asset → Improved Local Model
The interaction continues creating value long after the original question has been answered.
The resulting model becomes slightly better aligned with the organization, the repository, the architecture, and the developer’s way of working.
Repeated thousands or millions of times, the effect compounds.
Foundation Models Optimize for Breadth. Organizations Optimize for Depth.
The AI industry often treats model capability as a single dimension.
It isn’t.
Frontier models are attempting to approximate the entire world.
They must understand programming languages, biology, finance, law, mathematics, literature, history, and countless other domains simultaneously.
Their objective is breadth.
Organizations rarely need breadth.
They need depth.
A pharmaceutical company needs expertise in pharmaceutical development.
A law firm needs expertise in legal reasoning.
An engineering organization needs expertise in its codebase, architecture, standards, workflows, and institutional knowledge.
The optimization problem is fundamentally different.
A specialized model does not need to compete with GPT-6 across every possible domain.
It only needs to outperform general-purpose models within the domain that matters.
That is a much smaller problem.
And smaller problems tend to require smaller models.
The Economics Start to Break
This is where the conversation becomes interesting.
The current AI boom assumes that increasingly capable models will naturally translate into increasingly valuable inference businesses.
History suggests otherwise.
Infrastructure businesses rarely become more valuable as the underlying technology becomes cheaper.
They usually experience margin compression.
Compute becomes cheaper.
Inference becomes cheaper.
Open models improve.
Hardware improves.
Specialized accelerators proliferate.
Local inference becomes practical.
At the same time, organizations accumulate proprietary training assets that become increasingly difficult for competitors to replicate.
The result is a subtle but important shift.
General intelligence becomes abundant.
Specialized intelligence becomes scarce.
The economic value migrates accordingly.
The winners may not be the companies serving the most tokens.
The winners may be the companies that most effectively transform frontier intelligence into specialized capability.
What About Agents?
At this point, some readers are probably thinking:
“If local and specialized models become increasingly important, why are OpenAI, Anthropic, Google, xAI, and others spending tens of billions of dollars building larger foundation models?”
It is a reasonable question.
In fact, I think the answer may actually support the broader thesis.
Much of the recent progress in AI has not come from larger models alone. It has come from the increasingly sophisticated systems being built around them.
The industry has quietly moved beyond simple prompt-and-response interactions.
Modern AI systems increasingly consist of planning agents, coding agents, retrieval systems, memory systems, evaluation frameworks, search capabilities, tool integrations, and workflow orchestration layers. The model remains important, but it is no longer the entire product.
This creates an interesting dynamic.
Organizations that embed these capabilities deeply into business processes create switching costs that extend beyond the foundation model itself. Development workflows, architecture reviews, operational procedures, customer support systems, and internal knowledge platforms become dependent on the surrounding ecosystem.
The moat is no longer just the model.
The moat increasingly becomes the system.
This matters because many of these agentic architectures are beginning to resemble the very pattern described earlier.
What appears to users as a single intelligent system is often a collection of specialized systems working together behind the scenes.
Requests may be routed through planning agents, coding agents, retrieval systems, search services, domain-specific models, evaluation frameworks, and external tools before a final response is produced.
In many respects, foundation model providers have already discovered the limits of monolithic architectures.
As systems become more sophisticated, specialization naturally emerges.
Different components become optimized for different tasks.
Different models become optimized for different forms of reasoning.
Different agents become responsible for different stages of problem solving.
The architecture evolves from a single intelligence into a distributed collection of specialized intelligences.
That observation is important because it raises an obvious question.
If specialization improves performance inside frontier AI systems, why would we assume the same forces stop at the boundary of the provider?
Why wouldn’t those same forces eventually extend into enterprises, teams, repositories, workflows, and even individuals?
The same economic pressures that encourage foundation model providers to create specialized internal architectures may eventually encourage organizations to build specialized models aligned to their own domains and expertise.
Viewed through that lens, agentic AI is not a counterargument to the distributed intelligence thesis.
It may be one of the earliest indicators that the transition is already underway.
The Future May Look More Like Education Than Search
This possibility suggests a very different role for foundation model providers.
Today we think of them as answer engines.
Tomorrow they may function more like universities.
Their primary value may not come from answering every question forever.
It may come from teaching millions of specialized models.
In this world, frontier systems remain critically important.
They remain the source of frontier reasoning, discovery, synthesis, and evaluation.
But increasingly routine work migrates elsewhere.
The specialized model handles daily execution.
The frontier model handles novel problems.
This creates a distributed intelligence architecture rather than a centralized one.
Instead of millions of users depending on a handful of centralized models, a handful of frontier models continuously train millions of specialized systems.
The Second-Order Effects Are Larger Than They Appear
If this model proves correct (pun intended), the implications extend far beyond AI architecture.
Organizations begin accumulating intelligence as a balance-sheet asset.
Personal models become repositories of individual expertise.
Codebases become trainable knowledge systems.
Institutional knowledge becomes durable rather than transient.
The value of AI shifts from answering questions to creating capability.
And perhaps most importantly, the center of gravity shifts from consumption to ownership.
The question is no longer:
“Which model should I use?”
The question becomes:
“How quickly can I convert external intelligence into specialized capability that I own?”
That is a very different economic model.
And it may ultimately prove more durable than selling inference one token at a time, particularly as the foundation model providers continue to raise their prices and using their subscription models becomes more expensive.