For years, the dominant narrative around artificial intelligence was essentially linear: more compute would make it possible to train better models; better models would generate more applications; more applications would justify new investment; and that investment would make it possible to expand compute once again.
An apparently virtuous circle.
A contradiction is now emerging. In September 2026, calls to moderate progress at the frontier intensified on the grounds of safety, control and independent evaluation. Those reasons must be assessed on their own merits. There is no evidence that safety warnings conceal a financial motive, and claiming otherwise would be intellectually irresponsible.
But the absence of a proven causal link does not require us to ignore a relevant economic coincidence: slowing the pace of technological obsolescence is becoming objectively advantageous to a significant part of the industry.
Safety, a possible infrastructure bubble and cash-flow pressure are not mutually exclusive explanations. They can coexist without any one of them turning the others into a pretext.
The seller and the buyer do not face the same problem
NVIDIA occupies a singular position. Its business is precisely to sell ever more compute. In the second quarter of fiscal 2027, it reported $96.2 billion in revenue, of which $89 billion came from Data Center. Data Center revenue grew 117% year over year.
From an accelerator manufacturer's perspective, a new and more efficient generation can create another sales opportunity. Jensen Huang's incentive is understandable: maintaining the technology race expands the market to which NVIDIA supplies capital equipment.
For the buyer of that equipment, the equation is different.
Microsoft reported $41 billion of CapEx in its fourth fiscal quarter of 2026 and said that roughly two thirds went to short-lived assets, primarily CPUs and GPUs. The remainder funded infrastructure assets expected to be monetized over much longer periods.
The distinction is fundamental. Land, grid connections, substations, fiber and even buildings may retain economic usefulness for decades. A GPU may not.
Physical, accounting and economic life
A processor does not need to stop working to become economically obsolete. It may continue running its workloads perfectly while facing a new generation capable of delivering far more inference with the same energy, occupying less space or materially reducing the cost per token.
Three different clocks then appear:
- Physical life: how long the equipment can operate.
- Accounting life: the period over which it is depreciated.
- Economic life: how long it can compete reasonably with the next generation.
The third may be considerably shorter than the first two. This is the financial obsolescence of compute: the hardware still works, but a sufficiently rapid fall in the cost per unit of intelligence impairs its ability to deliver the expected return.
Artificial intelligence is developing inside an unusual industry: infrastructure investment associated with capital-intensive sectors, technology cycles associated with advanced electronics and margin expectations associated with software.
We build like a utility. We innovate like a semiconductor company. We expect returns like a software business.
Those three speeds need not be compatible.
Labs do not need to own the chips to bear the risk
OpenAI, Anthropic and other model developers do not necessarily own all the infrastructure they use. That does not remove the problem; it transforms it.
OpenAI entered a $38 billion, seven-year commitment with AWS to access hundreds of thousands of NVIDIA GPUs. Anthropic announced an expansion of its Google Cloud use to as many as one million TPUs, a commitment worth tens of billions of dollars and expected to bring more than one gigawatt of capacity online.
In such cases, the lab may not have to depreciate a physical GPU. But it must monetize the compute capacity it has purchased, contracted or financed.
The economic risk is distributed throughout the chain:
- Financial capital lends or invests.
- Cloud providers and data centers build facilities and buy equipment.
- Chipmakers sell compute.
- Labs contract and consume that capacity.
- Businesses and consumers ultimately pay for the services produced on top of it.
At the end of that chain, something remarkably simple must appear: cash flow.
OpenAI's own economic flywheel acknowledges it
OpenAI has publicly described its strategy through a particularly clear relationship: more compute enables more intelligent models; better models enable better products; better products should accelerate adoption; adoption should produce more revenue and cash flow; and that cash flow makes reinvestment possible.
It is a perfectly reasonable business formulation. It also contains the stability condition for the entire system.
Growth in AI-generated cash flow ≥ growth in the economic cost of sustaining its infrastructure.
As long as that relationship holds, the flywheel can continue. If it fails for long enough, the nature of the problem changes. The question would no longer be whether artificial intelligence works, but whether the economy built around it works.
A revolutionary technology can also create a bubble
The two ideas are not contradictory. Railways transformed the world and produced financial bubbles. The internet transformed the economy and coexisted with the dot-com bubble. A technology can change civilization while simultaneously attracting too much capital, too quickly and under mistaken return assumptions.
That is why asking whether there is an “AI bubble” is too imprecise. There may be no technology bubble and still be an infrastructure bubble.
Models could keep improving. Utilization could keep growing. AI could transform entire industries. Yet some data centers, capacity contracts or associated financing could still fail to earn the expected return.
That would not be the failure of artificial intelligence. It would be a misallocation of capital.
The speed paradox
If AI progresses too slowly, it becomes difficult to justify the enormous volume of investment directed toward it. But if it progresses too quickly, it can accelerate the economic obsolescence of the hardware we have just installed.
Too slow → the CapEx is not justified.
Too fast → the CapEx is not amortized.
Between those extremes lies a pace of development that is technologically competitive and financially sustainable. Finding it will not be easy.
A technological slowdown, even if defended for entirely different reasons, can have a concrete economic effect: more time to use installed capacity, develop products, win customers and turn investment into cash flow.
This proves no intention. It does create an incentive. Economic incentives deserve examination.
Financial capital may be the real inflection point
NVIDIA has announced platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR intended to mobilize more than $500 billion of third-party capital for AI infrastructure. The scale shows that the race no longer depends only on chips and energy. It depends on capital continuing to accept risk, duration and technological obsolescence.
The limit of an investment cycle of this size may not first appear because demand for artificial intelligence disappears. It may appear when financiers begin asking: when do I recover the money from the infrastructure I have already funded?
At that point, the critical variable ceases to be only the power of the next model. It becomes the return on capital.
Three speeds we can no longer confuse
The real debate may not be whether artificial intelligence should accelerate or slow down. We may be entering a stage in which we must distinguish three different speeds:
- The technically possible speed.
- The socially acceptable speed.
- The economically sustainable speed.
Until now, we have implicitly assumed that all three would advance together. There is no reason why they should.
The uncomfortable question is this: can artificial intelligence advance fast enough to justify the investments we are making and, at the same time, slowly enough to allow them to be amortized?
If the answer is ultimately no, the next AI crisis may not begin because the technology stops working. It may begin precisely because it works too quickly for the financial structure we have built around it.
Sources and interpretation note
- Reuters: calls to pace the frontier and strengthen independent evaluation.
- NVIDIA: second-quarter fiscal 2027 results.
- Microsoft: fourth-quarter fiscal 2026 results.
- OpenAI and AWS: $38 billion compute-capacity commitment.
- Anthropic: expansion to as many as one million Google Cloud TPUs.
- OpenAI: infrastructure, adoption, revenue and cash flow.
The financial facts and capacity commitments come from the linked sources. The thesis concerning financial obsolescence, compatible speeds and economic incentives is the author's interpretation. An objective incentive does not demonstrate that safety warnings are financially motivated.
© 2026 Javier F. Pérez Bernabé. All rights reserved. Published by Quórum 12.
ONLINE SUBSCRIPTION
Receive every new analysis and invitation to open events.
This is a separate public category: it does not confer membership, admission or access to the private portal.