The prompt:

I’d like to construct a speculative and fictional narrative grounded in factual history. The core of the thesis is that the AI stock market bubble will arrive simply because the technology to run LLM effectively on a small scale will proliferate, making the enormous data centres excessive for the actual demand. This has precedence in the recent history of computers themselves, where upon introduction, computers were large mainframes for several decades, but as their utility grew, the technology improved and their size and power requirements became such that we can all carry around computers more powerful then the original mainframes in our pockets as iPhones. Leverage this metaphorical alignment, and any others you can derive from history that will show the future to be a decentralization of LLM technology, and a focus on distributed, specialized nodes. Much in the way the internet enabled computers to join at scale and accomplish incredibly complex tasks, and internet-of-llm communication technology will unlock the same decentralized and distributed model. Please construct a narrative of the future that reflects this perspective.

ChatGPT:

What makes this thesis particularly interesting is that it doesn’t require assuming that AI becomes less important. Quite the opposite: the bubble can burst precisely because AI becomes too useful, too cheap, and too ubiquitous to remain primarily a data-center product.

The historical analogy is stronger than simply “computers got smaller.” The deeper pattern is:

A technology begins as scarce centralized infrastructure, then becomes standardized, miniaturized, specialized, networked, and eventually becomes an ambient capability embedded everywhere.

That pattern has happened repeatedly in computing.

The IBM System/360 is a particularly useful historical anchor. Introduced in 1964, it unified previously incompatible machines into a scalable architecture, allowing customers to move between different sizes of computer without rewriting their software. IBM describes it as a transition toward treating computing as a platform rather than as individually engineered machines.

And the next stages were not simply “smaller mainframes.” They were minicomputers → workstations → PCs → laptops → smartphones → distributed cloud systems → edge computing. A history of the industry explicitly describes this progression as a long-term trend toward miniaturization.

That gives us a rather compelling fictional future.

The Great Unbundling

An imagined history of the AI industry, 2026–2045

In retrospect, the great mistake of the early AI era was not believing too much in artificial intelligence.

It was believing too much in the data center.

The assumption seemed perfectly reasonable.

Large language models were enormous. Training them required enormous amounts of computation. Running the most capable models required expensive accelerators, enormous quantities of memory, sophisticated networking, immense amounts of electricity, and buildings constructed specifically to house them.

The economics appeared to point in one direction.

Bigger models.

More GPUs.

More data centers.

More power.

More capital.

And for a while, the strategy worked.

The companies building the largest models became some of the most valuable corporations in history. Semiconductor companies supplying the accelerators became infrastructure monopolies. Utilities began planning new generation capacity around AI demand. Entire regions competed to attract data centers.

The analogy of the time was that artificial intelligence was becoming a new form of electricity.

But there was another analogy hiding in plain sight.

It was the computer.

I. The Mainframe Mistake

The first computers were not personal possessions.

They were institutions.

In the 1950s, a computer could occupy a room, consume enormous amounts of electricity, require specialized operators, and cost more than almost any ordinary organization could contemplate spending.

The computer was therefore not something people owned.

It was something institutions used.

A corporation bought computing.

A government bought computing.

A university bought computing.

A scientist submitted a problem to computing.

The distinction seems obvious now, but it was profound.

The physical machine and the computational capability were essentially the same thing.

Then came standardization.

Then transistors.

Then integrated circuits.

Then increasingly sophisticated operating systems.

Then smaller machines.

The IBM System/360 was an important milestone because it separated the architecture and software from the particular physical machine. A family of machines could provide radically different levels of performance while remaining compatible. IBM’s six initial System/360 models spanned roughly a fiftyfold performance range.

That was more important than making one machine faster.

It made computing portable across machines.

The industry had begun to abstract computation from the hardware that performed it.

The minicomputer followed.

Then the microcomputer.

Then the personal computer.

And eventually something happened that would have seemed absurd to the engineers of the mainframe era.

The computer disappeared.

Not because computers became unimportant.

Because they became too cheap and too small to notice.

A computer became a watch.

A computer became a telephone.

A computer became a car.

A computer became a thermostat.

A computer became a camera.

A computer became a refrigerator.

By the beginning of the twenty-first century, a person could carry several computers more powerful than the machines that had once required entire rooms and teams of operators.

The mainframe did not vanish.

That distinction matters.

It became specialized infrastructure.

The mistake was assuming that because mainframes remained extremely useful, the world would continue to organize all computing around them.

It didn’t.

II. The First AI Mainframes

The first generation of artificial intelligence followed exactly the same pattern.

Except nobody recognized it.

The earliest frontier models were enormous centralized systems.

The model was the machine.

The machine was the service.

And the service lived in a data center.

A user typed a prompt.

The prompt travelled across a network.

A gigantic cluster performed inference.

The answer travelled back.

This architecture made perfect sense.

Frontier models were expensive.

Accelerators were scarce.

Memory was expensive.

Inference was computationally intensive.

And the companies capable of building the infrastructure were few.

So the economic structure naturally became centralized.

The world began constructing enormous AI factories.

The factories were spectacular.

Rows upon rows of accelerators.

High-speed interconnects.

Liquid cooling.

Dedicated power infrastructure.

Gigawatts of electrical capacity.

Entire buildings whose purpose was essentially to transform electricity into tokens.

The capital markets loved the story.

The logic was compelling:

AI demand will increase → models will become more capable → models will become larger → inference will increase → more accelerators will be required → more data centers will be required.

The conclusion seemed inevitable.

The AI economy would therefore be dominated by whoever controlled the largest machines.

And for a period, it was.

III. The Curve Nobody Was Watching

But underneath the spectacular growth of centralized AI, another curve was developing.

It was not the model-size curve.

It was the efficiency curve.

Every generation of hardware became better.

Quantization improved.

Attention mechanisms improved.

Architectures improved.

Distillation improved.

Speculative decoding improved.

Memory management improved.

Inference engines improved.

Model architectures became increasingly specialized.

And engineers discovered something that should have been obvious from the history of computing:

You don’t need the largest possible computer to perform most useful computations.

You need the smallest computer capable of performing this particular computation.

That distinction changed everything.

A model that required a cluster to run in 2025 might eventually run on a workstation.

Then a desktop.

Then a laptop.

Then a phone.

Not because the phone suddenly became a data center.

Because the model became efficient enough to no longer require one.

The first generation of AI had asked:

“How large a model can we run?”

The second generation began asking:

“How little computation can accomplish this task?”

That was the beginning of the end of the AI data-center monopoly.

IV. The Smartphone Moment

The decisive moment did not arrive with a single revolutionary model.

It arrived when people stopped being impressed by the size of the machine.

A small local model running at hundreds of tokens per second on a consumer device was not as intellectually impressive as a frontier model running on thousands of accelerators.

But economically, it was revolutionary.

Because the local model had three properties the data-center model could never possess:

It was already paid for.

It was already connected to the user.

And it could operate without asking permission from anyone else.

The economics of intelligence changed.

A company no longer needed to purchase every token from an AI provider.

It could purchase a model once.

It could run that model indefinitely.

And increasingly, it could fine-tune or customize it for its own purposes.

This was the equivalent of the personal-computer revolution.

The PC did not destroy the mainframe by becoming a better mainframe.

It destroyed the assumption that every computing problem should be sent to a mainframe.

The same thing happened with AI.

The question became:

Why send this inference to a $10 billion data center?

If a $2,000 workstation can do it locally, why?

V. The AI Bubble Bursts

This was the moment the financial markets misunderstood.

The AI bubble did not burst because artificial intelligence stopped working.

It burst because artificial intelligence worked too well.

Investors had valued AI infrastructure using an assumption of continuously expanding centralized demand.

They had extrapolated the first phase of the computing revolution indefinitely.

They assumed that more intelligence would always mean more centralized computation.

Instead, the cost of intelligence began collapsing.

The same phenomenon had happened repeatedly in computing.

Storage became cheaper.

Bandwidth became cheaper.

Processors became cheaper.

Memory became cheaper.

Computing became cheaper.

And every time computing became cheaper, the market did not consume the same quantity of computing.

It discovered entirely new things to do with it.

That was the paradox.

Efficiency did not kill computing demand.

It redistributed it.

But redistribution was disastrous for businesses whose valuations depended upon scarcity.

The market had priced AI as though intelligence were going to remain a scarce commodity.

Instead, intelligence was becoming a commodity.

VI. The Great Unbundling

The next decade became known as the Great Unbundling.

The monolithic AI system began to break apart.

A single enormous model had originally been expected to do everything:

reason,

write,

see,

hear,

speak,

code,

search,

plan,

remember,

control tools,

and interact with the physical world.

But specialized models proved dramatically more efficient.

A vision model did not need to be the world’s best mathematician.

A speech-recognition model did not need to know how to write software.

A translation model did not need to understand molecular biology.

A robotics model did not need to compose poetry.

And suddenly the ideal AI architecture began to resemble something surprisingly familiar.

Not a brain.

A computer network.

VII. The Internet of Minds

The internet had already demonstrated the fundamental principle.

You don’t need one computer to contain the entire world’s knowledge.

You need computers that can communicate.

The modern Internet is powerful precisely because computation is distributed among specialized machines.

A browser does not contain Google.

Google does not contain your banking system.

Your banking system does not contain the weather service.

The weather service does not contain your email.

They cooperate through standardized protocols.

AI eventually adopted the same architecture.

The first protocols were crude.

Models exchanged text.

Then structured messages.

Then compressed representations.

Eventually, systems developed mechanisms for exchanging richer intermediate representations.

An AI system could ask another AI system:

Analyze this image.

Another could answer:

Here is the semantic representation.

A reasoning model could then ask:

Determine whether the object violates the safety constraints.

A specialized mathematical model could calculate the answer.

A planning model could construct the sequence of actions.

A speech model could turn the result into audio.

A robotic controller could execute it.

None of these models needed to be capable of doing everything.

They only needed to be capable of doing their part.

The Internet had connected computers.

The emerging AI protocols connected capabilities.

VIII. The Model Became an Operating System

The most important AI systems of the 2030s therefore weren’t necessarily the largest models.

They were orchestrators.

A powerful general model remained important, but increasingly it behaved like an operating system.

It understood the user’s intent.

It decomposed problems.

It selected specialists.

It delegated work.

It verified results.

It maintained context.

And it assembled the results into coherent actions.

Underneath it existed an ecosystem of specialized intelligence.

There were vision nodes.

Speech nodes.

Reasoning nodes.

Coding nodes.

Mathematical nodes.

Scientific nodes.

Legal nodes.

Medical nodes.

Robotic nodes.

Memory nodes.

Planning nodes.

Simulation nodes.

Translation nodes.

Search nodes.

And millions of tiny models trained for tasks so specific that nobody would have bothered building them when inference cost dollars per interaction.

Now they cost fractions of a cent.

And because they were small, they could exist almost anywhere.

IX. The New Computer

The result was not the death of the data center.

It was something more interesting.

The data center became the equivalent of the mainframe.

There were workloads for which centralized infrastructure remained vastly superior.

Training frontier foundation models still required extraordinary resources.

Huge simulations still required enormous clusters.

Global services still needed centralized infrastructure.

Massive knowledge repositories still benefited from centralized storage.

But everyday intelligence migrated outward.

Into laptops.

Phones.

Cars.

Factories.

Homes.

Cameras.

Robots.

Servers.

Industrial controllers.

Scientific instruments.

And eventually into almost anything containing a processor.

The data center became one node in a much larger computational organism.

X. Intelligence Became a Network Resource

By the middle of the 2030s, nobody spoke about “using an AI model” quite the way people had in the previous decade.

They spoke about using the intelligence network.

A household might have several local models.

One handled speech.

One handled vision.

One managed household automation.

One specialized in children’s education.

One maintained private personal knowledge.

One controlled robotics.

A workstation might expose dozens of specialized inference services.

Companies operated internal model networks.

Factories had thousands of embedded inference nodes.

Cars communicated with roadside infrastructure and other vehicles.

Research institutions shared specialized models across institutional boundaries.

The important unit was no longer the model.

It was the capability.

Just as the Internet transformed computers from isolated machines into a global computational network, AI networking transformed models from isolated predictors into a distributed intelligence network.

The intelligence of the system was greater than the intelligence of any individual node.

XI. The Second Internet

Historians eventually called this period the Second Internet.

The first Internet connected computers.

The second connected computation itself.

The distinction was subtle but enormous.

The first Internet allowed one computer to request a service from another.

The second allowed one intelligent system to request a capability from another.

A machine didn’t need to know where the capability lived.

It simply asked the network.

The network located an appropriate specialist.

The specialist performed the computation.

The result returned.

And the original system continued reasoning.

The physical location of intelligence became almost irrelevant.

A model could run in your pocket.

It could run in your house.

It could run on a nearby server.

It could run in a data center on another continent.

The protocol determined the destination according to latency, cost, privacy, capability, and availability.

Computing had finally become what engineers had been promising for decades:

a resource rather than a place.

XII. The Data Center Crash

The great irony of the AI bubble was that the companies building the largest infrastructure were not necessarily wrong.

They were simply early.

They had built the equivalent of the great mainframes.

And, just like the mainframe companies, they discovered that enormous machines could remain enormously useful while simultaneously becoming less central to the computing economy.

The crash therefore looked catastrophic.

Data-center expansion slowed.

Some facilities became uneconomic.

GPU utilization forecasts were revised downward.

Investors questioned enormous infrastructure commitments.

Power contracts were renegotiated.

Hardware prices fell.

Companies that had been valued primarily on expected AI infrastructure growth collapsed.

The financial press called it the AI Winter.

But it wasn’t really an AI winter.

It was an AI spring.

The cost of intelligence was falling faster than anyone had anticipated.

And falling costs created demand.

XIII. The Cambrian Explosion

Once intelligence became cheap enough, entirely new applications appeared.

A model that would never have been economically viable at $1 per interaction became viable at one-thousandth of that price.

Then at one-millionth.

Models began appearing in places nobody had considered worth automating.

Every sensor could interpret.

Every camera could understand.

Every machine could diagnose itself.

Every software application could reason about its own state.

Every robot could perceive its environment.

Every document could become queryable.

Every device could become an interface.

The number of AI instances exploded.

Not because humanity built more giant models.

Because humanity built billions of small ones.

This was the Cambrian explosion of machine intelligence.

XIV. The New Economics of AI

The economic structure changed accordingly.

The scarce resource was no longer raw inference.

It was specialization.

A small model trained specifically for a particular industrial process could be enormously valuable even if it had only a fraction of the parameters of a frontier model.

The winners were companies that could build exceptional specialists.

The equivalent of software libraries emerged for intelligence.

Developers stopped asking:

Which model should I use?

They asked:

Which capabilities should my system compose?

Model marketplaces evolved into capability markets.

AI systems negotiated computational resources.

Inference became routable.

Specialists advertised their capabilities.

Nodes could join and leave networks.

Reputation systems measured reliability.

Verification systems compared outputs.

Redundant models cross-checked one another.

The architecture began to look remarkably like the Internet itself.

Except the packets were no longer merely carrying information.

They were carrying requests for thought.

XV. The Phone Call Between Machines

The first generation of computers communicated by exchanging data.

The second generation of networked AI communicated by exchanging conclusions.

A vision node might send:

Object detected: damaged hydraulic valve.

A diagnostic model might respond:

Probability of seal failure: 83%.

A maintenance planner might respond:

Replace seal within 47 operating hours.

A logistics model might respond:

Replacement part available 11 km away.

A robotic system might respond:

Maintenance can be completed autonomously.

The human might receive only:

Your hydraulic system is likely to fail within two days. I have scheduled preventive maintenance.

The individual models could be extraordinarily small.

The system could nevertheless appear extraordinarily intelligent.

This was the great architectural insight of the era:

Intelligence did not have to reside inside a model.

It could reside between models.

XVI. The Lesson of the Mainframe

Historians eventually recognized that the early AI boom had repeated a pattern that computing had already experienced.

The mainframe era had taught humanity that computers were powerful.

The personal-computer era taught humanity that computers could be ubiquitous.

The Internet era taught humanity that computers could cooperate.

The AI era initially taught humanity that machines could reason.

The post-AI era taught humanity something deeper:

Reasoning itself could be distributed.

The first computers centralized computation because computation was expensive.

The first AI systems centralized intelligence for the same reason.

But technology has a peculiar habit.

It takes something that was once scarce and makes it abundant.

And when that happens, the architecture built around scarcity eventually becomes the wrong architecture.

XVII. The Forgotten Assumption

Looking backward from 2045, the most surprising thing about the AI boom was not that investors overestimated artificial intelligence.

They underestimated it.

They assumed intelligence would remain concentrated because they assumed computation would remain concentrated.

They confused the location of intelligence with intelligence itself.

They looked at the first generation of AI and saw enormous machines.

They should have looked at the history of computing.

The mainframe was never the final form of the computer.

It was merely the first economically viable form of the computer.

The data center was not the final form of artificial intelligence.

It was merely the first economically viable form of machine intelligence at scale.

XVIII. The End of the Beginning

By 2045, the great data centers still existed.

They trained frontier models.

They performed enormous simulations.

They hosted global services.

They remained among the most powerful machines humanity had ever constructed.

But they were no longer synonymous with artificial intelligence.

A child could carry an intelligent system in her pocket.

A factory could contain ten thousand specialized models.

A home could operate its own private intelligence network.

A scientific laboratory could combine hundreds of specialized reasoning systems.

And a single general-purpose AI could orchestrate them all.

The world had not replaced the giant machine with the small machine.

It had replaced the giant machine as the unit of intelligence.

That distinction changed everything.

The computer revolution had begun with the question:

How can we build a machine powerful enough to compute?

It ended by asking:

How many machines can we connect?

The AI revolution began with:

How can we build a model intelligent enough to reason?

Its next revolution began with a different question:

How many intelligences can we connect?

And that was when artificial intelligence stopped being a product.

It became an infrastructure.

Not infrastructure in the sense of buildings, power stations, and data centers.

Infrastructure in the older and more profound sense:

a layer of civilization upon which everything else could be built.

The mainframe had not been defeated by the personal computer.

It had been surrounded by it.

The same fate awaited the AI data center.

The future did not belong to the largest model.

It belonged to the network.