A weekly briefing on AI’s macro impact.
Competitive shifts that once took decades now happen in weeks, and a shift this large almost always plays out in counterintuitive ways. Second Order looks for what AI is doing to capital and work.
Published weekly since June 2026
Separating prompt processing from answer generation can give older GPUs another job later in life. NVIDIA and Sarvam AI showed a 2x throughput lift on one H100 system, but the reviewed loans still repay from the original customer's contract, not a proven residual value.
Residential electricity prices jumped about a third in five years, but the data center is not why. The decade's increase is delivery cost on an aging grid; the real data-center effect is a recent capacity charge in the PJM belt and a modest forward charge as the net-new build is recovered.
A companion to Issue 06. The net load that will connect is a fraction of the gross queue, the binding near-term constraint is the gas turbine, and the stranding risk sorts by ownership: ratepayers under regulation, equity and credit in the merchant world.
Hyperscaler capital spending is guided toward $635 billion in 2026, and a rising share of that number is memory price, which buys no added capacity. The three firms that supply high-bandwidth memory sit at the center of the AI trade, and two of them are Korean.
A firm-level study of 21,559 firms looks at first like a refutation of Issue 03. On the contested point it is the opposite: the heaviest AI adopters grew entry-level headcount and tilted toward juniors.
In one June fortnight Washington restricted its best models and Beijing released its own as open weights. The binding constraint on the global AI standard moved from capability to diffusion.