Azhar counts nearly every hand up when he asks 250 IT executives about AI results

Azhar calls his own evidence what it is. "It's a qualitative signal," he wrote ^[5]. The count was made by eye from a stage, in a room of people who had travelled to an IT conference in Las Vegas ^[1]. At 95 percent, about 238 of the 250 hands went up ^[2][22]. The comparison is not the same room a year earlier but his estimate of what a room like it would have shown then, roughly a quarter ^[3]. He pairs the show of hands with Exponential View's revenue numbers, which he says show AI revenue growing faster in August than in July, and faster in July than in June ^[6]. The newsletter does not say which companies those figures cover.

The board-deck version runs to two lines: results are near-universal, spending goes up everywhere, so the only open question is sizing ^[2][4]. It is incomplete: a hand count is one datum from a self-selected room, and a finance chief comparing his own firm against 95 percent does not know what he is comparing against. In that same room sat a century-old American institution that had moved entirely to open-weight models, and a hospital running a mix of OpenAI and Anthropic models ^[7]. Each of those is one hand.

The other number in the week is measured in gigawatts. Bloomberg reported that Microsoft plans to take AI-serving capacity from about 2 GW today to nearly 13 GW by 2032, inside a fleet growing from 12 GW to 38 GW ^[8]. Eleven of the 26 GW added is AI-serving, about 42 percent of the additions ^[11], and AI's share of the fleet goes from roughly a sixth to roughly a third ^[10]. Azhar wrote that the 26 GW implies Microsoft expects demand it currently cannot serve ^[9].

Once spending is baseline, the harder governance problem is checking the output. The Navier-Stokes result shows that problem at full size. OpenAI put 10,000 agents on an unreleased model; across 2.7 million messages and 130 billion tokens, the run took 88 hours to reach a solution to a decades-old problem ^[17][23]. Azhar estimates it cost a few million dollars at today's prices, and a few tens of thousands in two years ^[18]. The proof runs past 500 pages and, he wrote, will not be intelligible to any human ^[19]. Terence Tao said that "[t]echnically, one of the most prominent open problems in mathematics would now be solved; but there would be almost no value added to mathematics as a consequence" ^[20]. Tao and twenty-four other Fields Medalists signed a public declaration warning about the way AI is used in mathematics ^[21]. An enterprise meets a smaller version of the same thing every time a system produces more output than the people accountable for it can read.

The macro case underneath the spending is the middle one. Azhar's models land near Anthropic's "substantial" scenario, in which AI adds about 8.3 percent to US GDP by 2030, with growth shifting from labor to capital and unemployment rising, mostly among knowledge workers ^[12][13]. Anthropic's extreme end has GDP up an additional 32.4 percent and unemployment doubling ^[14]. Azhar discounts that end on two grounds he states plainly: changes inside a firm take time, and political pressure builds as unemployment grows, slowing the pace ^[15][16]. Both scenarios run to 2030, and the increases the room committed to are for next year ^[4].