There Is No Single ‘AI Trade’ and Demand Is Only Growing
Winners in GPUs, memory, cloud margins, power and AI agents will be determined at different speeds.
You've reached your free article limit
You've read 0 of 1 free Pro articles.
Something bothered me during the July 2026 AI-stock rout. Chipmakers and data center stocks were getting crushed because investors had decided the industry was building too much, too fast.
Fine.
Then why was it becoming more expensive to rent the machines everyone supposedly had too many of? Wall Street was saying, “Enough GPUs.” The people using them were asking, “How soon can I get one?”
The selloff wasn’t purely a referendum on AI demand. Late in the month, steep losses at Leopold Aschenbrenner’s highly leveraged Situational Awareness fund led it to unwind its public-equity portfolio, adding forced selling to an already falling market. Some of the rout reflected leverage and liquidation—not a sudden collapse in demand for computing power.
So, the question wasn’t whether AI stocks were falling. It was whether AI demand was falling with them.
Gavin Baker, founder and chief investment officer of Atreides Management, called July “2022 in a month” during his recent “Invest Like the Best” conversation with Patrick O’Shaughnessy.
The interview is important because Baker wasn’t visiting Silicon Valley to collect evidence for a bullish story. He was trying to destroy his own thesis. His mission, as he put it, was to find “something negative.”
O’Shaughnessy asked whether Baker had found a single quantitative sign that AI demand was decelerating.
Nothing.
Instead, Baker said GPU availability, rental prices, memory prices and token growth were all accelerating. That doesn’t prove the AI boom will last forever. It does mean anyone declaring a glut must explain why the supposed surplus still looks like a shortage.
Consider rental prices. Nvidia B200 GPUs went from roughly $2 per GPU-hour to about $4 in seven months.
That doesn’t mean cloud providers automatically collect twice as much. Their fleets include long-term contracts, discounted capacity, internal workloads and several generations of chips. But spot prices still tell us what customers will pay for the next available machine.
When developers build too many apartments, rents fall. When farmers grow too much corn, corn gets cheaper. When the spot rental price of a machine doubles, the market is sending a simple message:
We need more machines.
Wall Street, meanwhile, is staring at the capital spending of Alphabet (GOOGL), Amazon (AMZN), Microsoft (MSFT) and Meta (META) and breaking into a cold sweat. The numbers are enormous. But spending is only one side of the equation. The other side is what those assets earn.
Much of today’s GPU capacity was contracted before rental prices climbed. As those agreements reset, cloud providers may collect more revenue from the same equipment. Baker argues that rising operating cash flow could finance more of the buildout.
The machine may help buy the next machine.
This is also where comparisons with the telecom bubble start to weaken. That boom funded miles of (dark) fiber customers weren’t ready to use. Today’s largest AI spenders entered the cycle with profitable businesses, existing customers and enormous cash flows.
Can they overspend? Of course. Intelligence doesn’t vaccinate a corporation against stupidity. But these aren’t pre-revenue companies laying cable through the desert because an investment banker made an exciting spreadsheet.
Then there are AI agents.
An agent doesn’t merely answer a question. Give it a goal, and it can plan the work, gather information, operate other software and revise its approach. For consumers, that might mean planning and booking a vacation. For businesses, it could mean researching prospects, analyzing contracts, or resolving customer problems.
Chatbots provide answers. Agents perform work — and could turn every user into the manager of a small digital workforce.
Baker estimates that only 250,000 to 500,000 people seriously use agentic AI today. That sounds large until you compare it with billions of potential users. Then it looks like five people discovering email in 1992 and declaring the inbox overcrowded.
Agents also consume far more compute than a conventional search. One assignment may require dozens or hundreds of steps.
Open-source AI could add fuel. Free models may compress software margins, but cheaper intelligence attracts more users. More users mean more chips running, more memory filling, and more electricity being consumed.
Cheaper intelligence could produce more total spending on intelligence.
There are risks.
GPU rental premiums could collapse. Algorithms could accomplish the same work with less computation. Regulators could block data centers over electricity, water, or land. Scarcity is powerful, but it isn’t immortal.
The mistake is treating AI as one trade with one expiration date. Software prices, GPU supply, cloud margins, memory, power and agent adoption will move at different speeds.
The opportunity, then, isn’t to buy “AI” as if it were one stock. It is to follow the bottleneck.
A shortage can boost the earnings of the company controlling it while delaying revenue and squeezing margins for everyone behind it. Scarce memory may benefit memory suppliers while slowing data-center deployments. Limited power may increase the value of electricity while leaving expensive GPUs sitting idle.
That can interrupt an AI rally without ending the AI buildout. Demand may remain enormous even while individual companies miss a quarter — or lose their place in the chain.
AI is not one trade. The real question is who owns the next constraint, and who will be forced to pay for it.
At the time of publication, Byrne had no positions in any securities mentioned.
