Compute leaves the fab

The binding constraint of the AI buildout has moved. For two years the limiting input was silicon — wafer starts, packaging capacity, the most advanced node. Now it is physics. Call it the handoff from silicon to industry: compute leaves the fab.

The pattern underneath is industrial. A data center is an industrial heat exchanger with a server problem stapled to it, and the only firms on earth that already engineer large-scale heat exchangers, gas turbines, and grid-scale power systems are heavy-industrial conglomerates. Mitsubishi Heavy Industries' July 9 cooling-optimization demonstration — a standalone demonstration of cooling and power efficiency in an operational data center — was the dress rehearsal, Nikkei reported, for the partnership the outlet described as still in the considering stage. The next AI supplier is not a chip designer. It is a thermal engineer.

The mechanism is repeatable. Each new generation of accelerators pulls more power and sheds more heat per rack. Liquid cooling, direct-to-chip loops, and grid-scale power conditioning become prerequisites, not options. Whoever owns that physics owns the next margin. The chips got cheap; the kilowatts will not.

Mitsubishi Heavy joins the AI supplier stack the way gas-turbine makers joined the LNG boom — not because they were tech, but because the new bottleneck was their old business.