— On “The AI Bubble: We Know Where to Look First”
Over the past two years, whenever AI stocks rise or data-center investment accelerates, the same question returns: is AI a bubble?
The question looks simple, but it is difficult to answer because both sides of the debate are looking at real facts. That leaves us with a strange picture:
Investment looks bubbly, yet demand does not look fake.
This contradiction is hard to explain using only traditional investment frameworks and the user metrics inherited from the internet era. The purpose of the article is to take the problem apart more carefully.
The first thing we found is that today’s AI infrastructure is no longer quite the same thing as the data centers of the past.
Traditional data centers mainly stored, transmitted and processed information. What is now called an “AI factory” combines chips, electricity, networks, data, models and software into the computational capability required to perform intelligent tasks. More importantly, once goals and permissions are set, AI can continue searching, reasoning, coding, testing and executing without requiring a person to remain involved in every step.
That is why we are willing to keep a deliberately strong formulation: AI factories are becoming productive capacity in their own right.
A conventional tool usually requires continuous human participation before it produces labor output. AI is different. Human work can stop while machine cognitive work continues. That difference changes how AI demand should be understood.
In the internet era, we learned to ask: how many users are there, how many are active each day, and how many pay?
But a person who asks AI two questions a day and a person who has several agents continuously conducting research, coding and analysis may both appear as “one user” in conventional statistics.
Clearly, they do not represent the same amount of demand.
What we really need to measure, therefore, is not only the number of AI users, but how much cognitive work humans are handing over to machines.
If AI were merely another internet application, today’s enormous capital expenditure would be easy to question.
But if AI is creating a new form of labor supply—allowing one person to mobilize increasing amounts of machine cognitive work—then the future market cannot be estimated simply from population or user counts. AI does not add people, but it can increase the amount of equivalent cognitive labor each person can command. One person is still one person, yet an expanding number of machines may be working behind that person.
That leads to a different question:
Is today’s enormous AI investment running ahead of future demand, or have we simply not learned how to measure that demand correctly?
This is the real dividing line of the article.
Our answer is not that “AI has no bubble.” Quite the opposite: bubbles probably exist, but they are more likely to be concentrated in specific companies, assets, financing structures and valuations than to mean that the entire AI system fails at once.
Much of the infrastructure needed for AI expansion already exists. The internet, cloud computing, smartphones, digital payments and global software distribution are mature. AI does not have to wait for society to digitize before it can spread. This is one of the biggest differences from the early internet era.
Long-term AI demand can be real while some companies, financing structures, data centers and valuations are still clearly excessive. The question can therefore be narrowed further:
If local bubbles do exist, where are they most likely to surface first?
At this point in the research, we did not begin with NVIDIA, nor with Microsoft, Google or Amazon.
The large technology companies are spending enormous sums, but they also have substantial internal absorption capacity. If external demand disappoints, they can redirect compute toward their own models, cloud businesses, search, advertising, office software and internal R&D. Their investments can still be wrong and returns on capital can still fall, but they are less likely to face survival pressure first.
Model companies have another form of flexibility. If a model path fails, they can change algorithms; if price competition intensifies, they can reprice; if foundation models do not make money, they can move toward agents, enterprise applications or open ecosystems. Their greatest pressure is whether they can keep innovating.
Pure compute lessors are different.
They borrow and raise capital, buy GPUs, secure power, build facilities, and then wait for customers to use the capacity.
Models can be modified and algorithms retrained. A GPU data hall cannot easily become another kind of asset. Even when the facility is underused, depreciation continues, power contracts remain in force and interest still has to be paid. What pure compute lessors fear most is not necessarily that “AI disappears.” They fear something subtler:
AI may remain highly successful while compute simply stops being as scarce as it once was.
A modest decline in pricing, slightly shorter customer contracts, wider discounts or lower utilization can flow into cash flow and the balance sheet much faster than they would at a diversified technology giant. That is why we eventually arrived at a simple observation chain.
First, watch a market-facing compute provider such as Nebius: is compute still expensive?
Then watch an aggressive builder such as Nscale: can huge contracts actually become financing, construction, energized capacity and revenue?
Finally, watch a mature neocloud such as CoreWeave: have these pressures begun to reach cash flow, debt and the balance sheet?
These three companies are three pressure gauges.
If the sequence gradually becomes persistent softness in compute pricing, greater difficulty signing long-term contracts, tighter financing, project delays, and finally simultaneous deterioration in cash flow and debt metrics, then “the AI bubble” will begin to move from a market opinion to a verifiable chain of facts.
So far, we have not seen that chain establish itself.
That is what separates this article from much of the AI-bubble debate. We are not trying to predict a grand date when “the bubble bursts.” We first identify the most fragile point in the system, and then wait for evidence there.
There is, of course, another possibility that cannot be ruled out: we may still be underestimating AI.
Chat, coding, multimodality, enterprise workflows and agents may not be different stages of a single S-curve. They may be several S-curves arriving one after another: one curve is still developing while the next has already started. Lower costs may also create entirely new tasks that were previously too expensive or too slow to attempt at all.
If one person can eventually orchestrate multiple machine agents, humans will no longer be the only bottleneck in the supply of labor. Chips, electricity, networks and data centers may then become the hardest constraints again.
That helps explain why large companies continue investing even when they know overbuilding is possible.
They may not simply be betting on “more users.” They may be betting on something larger:
An increasing share of cognitive work that used to be performed by humans—or was never worth doing at all—will be handed over to machines.
If that is correct, the risk is not only that capital expenditure may be too large. The opposite error also matters: we may still be using an old ruler and therefore measuring new demand as smaller than it really is.
The article therefore makes the problem clearer:
Do not begin by asking whether AI is a bubble. First ask whether machine cognitive labor is still expanding. Then ask whether compute is becoming harder to sell.
Next, watch whether contracts, financing and cash flow weaken in sequence. Only if these signals persist and confirm one another do we have reason to say that an inflection point is forming.
If they do not appear, we should not judge the future simply because today’s capital expenditure looks enormous.
A great technological revolution and enormous local bubbles can coexist. AI may be similar.
We already know where to look first if the bubble begins to break.
