The Best-Informed AI Bull Just Lost Money. Microsoft Had Its Best Day Since 2008.
Both happened on July 30. What separated them was not who understood AI better, it was who held a claim on capacity. That is what the market has quietly started paying for.
On July 30, a $24 billion hedge fund founded by a former OpenAI researcher was unwinding many of its trades after heavy losses on AI stocks. On the same day, Microsoft shares jumped 15.5 percent, the biggest one-day gain since 2008, on the back of Azure cloud growth.
One of those firms is staffed by people with an unusually early and unusually specific view of where AI is going. The other one rents servers.
Guess which one got paid.
The point is not that Leopold Aschenbrenner was wrong about AI. He may turn out to be very right. Situational Awareness LP lost money on AI stocks and on a short against software names, which is a statement about positioning and timing rather than about whether capability curves bend the way he thinks. But it does confirm something the market has been saying all quarter: being early to the correct AI thesis has stopped paying. Holding the capacity does.
## What the money actually bought in July
Look at where committed dollars went rather than where opinions went.
Anthropic signed a roughly $10 billion, six-year compute deal with Volta, capacity in a 133 megawatt Norwegian data center running Nvidia hardware. The European Union put 10 billion euros behind seven AI gigafactories, each meant to house at least 100,000 advanced chips, explicitly designed to pull in another 20 billion euros of private money. Neither commitment is a bet on which model wins. Both are purchases of a place to compute.
Our own deal tracking shows the same rotation. Compute contracts were 3.9 percent of the AI deals we track in Q4 2025 and 14.6 percent in Q3 2026. Acquisitions went the other way over the same stretch, from 16.5 percent down to 3.1 percent. Those are shares of deal count rather than deal value, so read them as a change in what people sign, not in how much they spend. The industry has moved from buying companies to booking capacity.
## The constraint is physical, and it is being rationed
Capacity prices like an asset because somebody is now standing at the door.
On August 4, Texas ordered its utility commission and grid operator ERCOT to audit every new data center project before it can connect to the grid. ERCOT's interconnection queue has reached about 474 gigawatts, roughly 90 percent of it data centers.
A queue that size is a waiting list, not a pipeline. Once a state starts auditing who gets to plug in, a signed interconnection agreement is worth more than a conviction about AGI timelines. Analysis of where AI is heading is abundant and close to free. Substations are neither.
That is the decoupling. For three years the edge was understanding what was coming, and the people who understood it earliest built funds on that edge. The supply of that understanding has now caught up with demand, and the scarce input has moved into the physical world, where it is held by incumbents with balance sheets, land and permits.
## What to do with this
Two concrete moves.
First, when you read an AI forecast, ask what the author's position actually pays off on. A correct call on capability and a profitable call on equities have become different trades, and July showed you can win the first while losing the second badly. Judge the forecast on its own terms, and judge the portfolio separately.
Second, if you want a leading indicator for AI supply, watch interconnection queues and grid regulators rather than benchmark releases. ERCOT's 474 gigawatts, and the audit now sitting in front of it, will tell you more about how much AI actually gets served in 2027 than any model launch this autumn.
The market has quietly stopped paying for being right. It is paying for being plugged in.