What We Are Watching
Artificial Intelligence · Technology
The People Asserting That Compute Will Grow Forever Are the People Who Sell Compute
The dominant narrative around AI infrastructure investment is being advanced loudest by the companies whose revenues depend on it being true. The demand signal is real — but the relationship between AI demand and compute demand is not fixed, and the investment theses that assume it is are exposed to a technical reality moving in the opposite direction.
2/10/2026
What We Are Watching is Vaxa's signal intelligence column. We identify markets, technologies, and structural shifts already in motion but may not yet reached the corporate strategy conversation.
highlights
01
The 'compute will grow forever' narrative is being driven by the companies that sell compute — not neutral observers
02
AI infrastructure investment is real and large — but the demand curve for inference is structurally different from training
03
Efficiency gains in model architecture are compressing the compute required per unit of AI output
04
The companies building data center capacity at scale are making a bet on demand that is not yet proven at the utilization levels the economics require
05
The strategic question is not whether AI matters — it is which part of the AI value chain captures durable value
Signal
The dominant narrative around AI infrastructure investment is that compute demand will grow exponentially and indefinitely — that the scale of model training required to advance AI capabilities, combined with the inference demand from billions of users, means that data center capacity, power generation, and networking infrastructure are all in structural undersupply for the foreseeable future.
That narrative is being advanced loudly and consistently. It is also being advanced primarily by the companies whose revenues depend on it being true.
Every major technology build-out in the past thirty years has been accompanied by a version of this narrative. The infrastructure is always described as undersupply. The demand is always described as insatiable. The economics eventually find the clearing price.
Pattern
The compute-will-grow-forever narrative has a structural problem: the people asserting it most forcefully have the most to gain from it being believed. Nvidia's ability to sustain GPU pricing depends on compute demand continuing to grow faster than supply. Hyperscalers justify their capex to investors by pointing to AI demand that requires the infrastructure they are building.
The demand signal for AI compute is real. The question is whether the specific version of the demand curve being used to justify current investment levels reflects the actual trajectory of the market.
The people telling you compute demand is insatiable are the same people selling the compute.
Mechanism
The most underexamined pressure on the compute-forever narrative is what is happening to the compute required per unit of AI output. Model architectures are becoming more efficient. The same quality of output that required a large model twelve months ago can now be achieved with a smaller, faster, cheaper model.
If inference efficiency continues to improve at the rate it has been improving, the compute required to serve a given volume of AI queries declines over time. That dynamic puts meaningful pressure on the utilization assumptions behind the current wave of data center investment.
Capital flowing into a market is not proof the market is as large as the capital assumes.
Implication
The capital flows into AI infrastructure are genuinely large and accelerating. That is a real signal. But capital flows are not demand data. They are bets on demand, made by investors whose incentives are to be in the market during what appears to be a structural buildout.
The more useful signal is utilization rates — what percentage of deployed AI compute capacity is actually being used productively. That data is largely not public.
Nobody selling compute is going to be the one who tells you when demand peaks.
Question
The companies most exposed when utilization data eventually surfaces are the ones that built long-term capital commitments on the assumption that demand would simply validate itself.
The more defensible position is to treat the compute-forever narrative as one plausible scenario among several — not the base case current investment plans are being built around.
The narrative that a market will grow forever has never once been the one that turned out to be true.
The question to ask.
Is your AI infrastructure strategy built around demand data — or around the demand narrative being advanced by the companies that benefit most from you believing it?