Return on Intelligence, Part 2: Prototypes

Part 2: Prototypes

Part 1 argued that markets routinely misprice real paradigm shifts by pricing the mature economy too early. This chapter asks the harder follow-up question: if AI utility is real, where does the value actually settle, and who captures it before the story breaks again in Part 3?

The early companies and interfaces that make a new paradigm legible are prototypes of futures they may not own.

A prototype can be directionally right and economically doomed.

That is the central claim of this chapter. Pets.com understood something real about the internet. AOL made the web accessible to millions. The first AI chat applications are teaching ordinary people to think in prompts. Each of those is a prototype: evidence that the direction is correct, not proof that the current owners are the final ones.

The easiest way to misunderstand the AI bubble is to assume there are only two possible positions.

Either AI is revolutionary, therefore the valuations make sense, or AI is overvalued, therefore the technology is fake.

I think both positions are wrong. The technology is real. The valuations can still be wrong.

That was the first lesson of the dotcom bubble. A technology can be world-changing and still destroy enormous amounts of capital in its first speculative wave. The internet did not fail in 2000. The market story failed. Investors had correctly sensed a paradigm shift, then funded almost every early expression of that shift as if it had already reached its mature form.

That is why the current argument around AI often feels so shallow. People keep asking whether AI is real, whether it is useful, whether it is better than search, whether it can code, whether it can reason, whether it can replace workers, whether companies are using it.

Those questions matter, but they are not enough.

The more important question is this:

If AI is real, who actually captures the value?

That is where the bubble sits.

Not in the claim that AI matters.

In the claim that the companies currently priced as if they own the future will, in fact, own it.

Real utility creates more convincing bubbles

The most dangerous bubbles are not built on nothing.

They are built on enough truth to make the overreach feel rational.

AI has that truth.

Stanford’s AI Index is one of the clearest high-level reminders that AI is no longer a speculative abstraction. And Microsoft now markets Microsoft 365 Copilot as something woven directly into the everyday work surface of documents, meetings, files and chats. Those facts do not prove the current valuations are correct. They do make the bubble more convincing.

A developer using a coding assistant can feel it. A student using it to understand a topic can feel it. A writer using it to structure a draft can feel it. A support team using it to summarize tickets can feel it. A manager using it to synthesize documents can feel it. A designer using it to explore visual directions can feel it. A small business owner using it to write copy, build a spreadsheet or understand a contract can feel it.

This matters.

AI is not a pitch deck technology. It is already in people’s hands. The usage is not theoretical. The productivity gain is uneven, but it is not imaginary.

That is exactly why the bubble can run so far.

When people can feel a technology working, they become much more willing to believe that the current winners are inevitable. They confuse personal proof of utility with market proof of durable economics.

I can use AI every day and still believe OpenAI is overvalued.

I can believe AI will transform software and still believe many SaaS companies will be repriced downward.

I can believe AI will become part of almost every business and still believe that much of the current infrastructure spending will be mistimed, overbuilt or monetized by someone other than the people funding it.

This is the central paradox of the series:

AI is real enough to destroy the current AI trade.

The market is pricing too many assumptions at once

The current AI market story rests on several assumptions.

First, that frontier models remain scarce and expensive enough for a few model labs to capture large rents.

Second, that hyperscaler capital expenditure converts into durable, highly utilized, high-margin AI revenue.

Third, that SaaS incumbents can add AI features and preserve their existing application and workflow rents.

Fourth, that enterprise adoption moves from experimentation to deep operational transformation quickly enough to justify the spend.

Fifth, that AI agents become reliable enough to automate meaningful amounts of work, while still depending on centralized cloud model infrastructure.

Sixth, that customers tolerate the pricing required to support the infrastructure.

Seventh, that the value of AI does not rapidly commoditize as models get smaller, cheaper, open, local and embedded.

That is a lot of belief stacked together.

A bubble does not need every assumption to be false. It only needs enough of them to soften at the same time.

The market may be broadly correct that AI is a paradigm shift and still be wrong on timing, winners, margins, architecture, pricing power and capital intensity.

That is how real-technology bubbles work.

This is not another 2008

It is useful to separate this from 2008.

The financial crisis was not society over-extrapolating a new general-purpose technology. It was a crisis of leverage, opacity, securitization, ratings, incentives and fraud. The system pretended that risk had been transformed when much of it had simply been hidden, sliced, resold and misunderstood.

The dotcom bubble had bad behaviour too. Every bubble does. But its emotional centre was different. It was not primarily a hidden balance-sheet rot. It was a visible overbelief in a new technology.

AI feels closer to dotcom than to 2008.

Not because the assets are identical. They are not.

Not because the same companies are involved. They are not.

Not because the crash will look the same. It will not.

It feels similar because the psychological structure is similar.

People can see the future, but they cannot yet see the final business model.

So the market invents one.

In the late 1990s, the invented story was that every business needed to become a dotcom and that first movers with traffic, brand and capital would dominate huge new markets.

Today, the invented story is that intelligence will flow through a small number of model and cloud platforms, that every SaaS incumbent can defend itself by adding AI, and that enterprise adoption will generate enough cash flow to justify the infrastructure buildout.

Some of that will be true.

Not enough of it has to be true for every valuation to survive.

The wrong lesson from Pets.com

People use Pets.com as if it proves that everyone in the dotcom bubble was delusional.

I think it proves something more interesting.

The idea was early, not permanently absurd.

A pet supply business on the internet is normal now. Subscription pet food, online pharmacy, bulk delivery, automatic repeat orders, logistics optimization, customer reviews, card payments, warehouse networks, mobile commerce, all of that now exists around the category.

The early company failed because the world around the idea had not matured.

The same lesson applies to AI.

Some ideas that look absurd now will become normal later. Some current AI companies that look central now will vanish. Some use cases that feel clumsy today will become invisible infrastructure tomorrow. Some infrastructure built too early may become valuable later under different ownership.

The market will not be wrong to believe in AI.

It will be wrong to believe that the first visible structure is the final structure.

Utility does not equal capture

This is the most important distinction.

Utility is not capture.

A technology can create enormous value for society while the companies that funded the first wave capture less of it than expected.

The internet created enormous consumer surplus. It made communication, search, publishing, commerce and software distribution radically cheaper. But many of the first companies that created or popularized those behaviours did not survive.

The value migrated.

It migrated to search. It migrated to cloud. It migrated to social networks. It migrated to app stores. It migrated to mobile operating systems. It migrated to advertising platforms. It migrated to payment systems. It migrated to logistics networks.

AI value will migrate too.

It may migrate away from pure model APIs toward chips, devices, local inference, operating systems, browsers, enterprise data, identity, workflow orchestration, systems of record and AI-operated businesses.

It may migrate away from generic SaaS applications toward custom generated software and internal control planes.

It may migrate away from human-heavy services companies toward tiny AI-operated competitors.

The AI economy can grow while many AI-era valuations collapse.

That is not a contradiction.

It is the most likely path for a real paradigm shift.

A paradigm shift can feel tangible on the desk long before the ownership map is clear.

Why enterprise adoption will be slower and deeper than the demos suggest

AI demos are seductive because they remove the mess.

A clean prompt. A clean answer. A generated app. A summarized document. A working prototype. A sales email. A graph. A plan. A piece of code.

But businesses are not demos.

Real businesses have permissions, policies, compliance requirements, legacy systems, bad data, partial data, missing data, political constraints, audit requirements, human exceptions, edge cases, incentives, vendor contracts, customer commitments and regulators.

This does not make AI less important.

It makes the transformation slower, deeper and more structural than the market wants to price.

The real value of AI in an enterprise is not a chatbot sitting next to a process. It is AI embedded inside the process, with access to the right data, governed by the right permissions, orchestrated through durable workflows, evaluated against real outcomes and constrained by human approval where needed.

That is hard.

It is not impossible. It is precisely where the long-term value sits. But it means a lot of early enterprise AI spending will disappoint before the real operating model emerges.

Again, this is dotcom logic.

A website was not a business model.

A chatbot is not a business model.

The first AI story will fail

I think the first AI story is already visible.

It says that the model labs own intelligence. The hyperscalers own the compute. The SaaS companies own the workflows. Enterprises will pay for AI features inside the software they already use. Agents will increase usage. Infrastructure spending will keep compounding. The companies closest to today’s AI excitement will become tomorrow’s dominant economic layer.

That story is too neat.

The real story is likely messier.

Models will get smaller. Inference will get cheaper. More capability will move locally. Open models will keep improving. Devices and operating systems will absorb more intelligence. Enterprises will build private and hybrid AI stacks. SaaS will be pressured by custom software. AI-operated companies will compete with human-heavy incumbents. Some infrastructure will be overbuilt. Some capital will be burnt. Some of today’s winners will remain winners, but for different reasons than the market currently believes.

The bubble will burst when the first story can no longer carry the valuation.

That does not require AI to fail.

It requires AI to mature.

That is why I do not think the AI bubble will burst because AI is fake.

I think it will burst because AI is real, and because reality will be more disruptive to today’s market story than scepticism ever could be.

In Part 3, I take that question directly into the current ownership bet: the idea that intelligence as an API is itself the durable moat.