AI Is Hiring

For two years the consensus was clean: AI would flatten headcount and expand margins. This week, from operators, macro desks, and payments data, the story quietly turned. The companies leaning hardest into AI are hiring faster, not slower, and the bull case now depends on AI consuming more of the physical economy, not less.

Start with the payroll data. Work from Ramp and Rellio Labs shows companies with high AI adoption grew headcount roughly 10% on average over the past two years, while low-adoption peers were basically flat. That is the inverse of the pitch deck. Anthropic has disclosed that 65% of its product team’s code is now initiated from Slack via Claude Code, which is not a story about fewer engineers but about engineers becoming the interface for many more parallel jobs.

The org chart is being redesigned around that reality. Adam Mosseri has said Instagram is moving to “pods” of four to six generalist engineers and a new “product staff” role that folds PM, design, data, and research into one person leveraging AI tools. Investor Mike Mignano put the operating principle bluntly: “Don’t automate, obliterate.” The unit of production is shrinking; the number of units is going up.

The founder side of the labor market shows the same signal from the other end. Per Stripe Atlas, solo founders accounted for 63% of C-corp formations in Q2 2026, an all-time high. Small teams are shipping what mid-sized ones used to. Operator Chris Hladczuk has described Hanover Park as “no product managers, no designers, just engineers shipping code sitting alongside fund accountants.” None of this looks like a labor glut. It looks like leverage that pulls in more hands, not fewer.

Then there is the bill. Uber capping AI spend at $1,500 per person per month is a small tell that individual seat economics are running hot. The bigger tell is on the input side. Jeff Currie has argued the thing Wall Street is missing is that “AI is no longer infinitely scalable at zero marginal cost,” because training and inference now consume raw materials whose supply curves slope up like everything else. Currie’s corollary: energy is roughly 3% of the S&P and the super-cycle case requires it closer to 10 to 15%, taken out of the AI sector’s own multiple.

That reframes the capex boom. If Currie is right, the AI trade and the commodity trade are the same trade with opposite P&L signs, and the marginal AI dollar increasingly buys electrons and copper rather than software margin. Kunle Olukotun has made the mirror-image point from the silicon side: inference is now bottlenecked on moving weights and KV cache, not on compute, with GPUs running at 10 to 20% of peak. The physical constraints are binding at both the datacenter and the grid.

The turn to notice is not that AI failed to deliver productivity. It is that the productivity is showing up as more headcount, more founders, more spend, and more raw-material draw, all at once. The 2024 story was AI as a substitute. The 2026 story, as it is actually being told by the people running the P&Ls, is AI as a complement with a very large power bill.