The capital stock framing is right, but the mechanism goes one layer deeper. We didn't just move production offshore — we moved the process knowledge that ran it. Capital can fund new structures and machines. It can't compress the institutional memory that accumulates in tooling, process engineers, and supplier relationships over decades. TSMC Arizona is the proof: unlimited capital, world-class engineering, still three-plus years to high-volume production. That ceiling is the constraint the investment charts can't show.
The machinery line is the one to sit with: construction spend more than doubled since 2022 and still lands at 3.9 billion, which shows how thin the base is. The deeper gap is that the machines that make the machines sit abroad. China now produces roughly a third of the world's machine tools and just passed Germany as the largest exporter. A funded reshoring push still routes its deepest tool layer through foreign suppliers, which is the part a CHIPS-style subsidy doesn't reach.
MIIT ran an MLCC domestic substitution symposium this week — exact same category as your "stuff that makes the stuff." AI server demand is the stated rationale for accelerating local sourcing of capacitors, inductors, and magnets. China is building the capital stock for the stuff America is importing.
Put two of these numbers next to each other. Hyperscalers spend $600B+ on data centres this year. Construction spending on machinery factories, the actual stuff-that-makes-the-stuff category, is $3.9B. Roughly 150 to one in favour of buildings for thinking machines over buildings for machine-making machines.
And the two aren't separate stories. They compete for the same inputs, turbines, transformers, electricians, grid connections, all sold out for years, and the marginal unit of each goes to whoever pays most, which is always the data centre. The AI buildout is bidding those inputs away from the broader industrial one. The factory that would make machine tools loses its electricians to the server hall next door.
The 2013 reversal makes it compound, too. Import parts and you're still sovereign. Import the machines that make the machines, and expanding your own capacity runs on someone else's export licence. I'd say maybe 30% that ex-AI manufacturing investment meaningfully accelerates by 2028. The boom-let is real. It might also be eating the boom it was supposed to start.
The capital stock framing is right, but the mechanism goes one layer deeper. We didn't just move production offshore — we moved the process knowledge that ran it. Capital can fund new structures and machines. It can't compress the institutional memory that accumulates in tooling, process engineers, and supplier relationships over decades. TSMC Arizona is the proof: unlimited capital, world-class engineering, still three-plus years to high-volume production. That ceiling is the constraint the investment charts can't show.
https://levelset.substack.com/p/waymo-tesla-robotaxi-manufacturing-architecture
The machinery line is the one to sit with: construction spend more than doubled since 2022 and still lands at 3.9 billion, which shows how thin the base is. The deeper gap is that the machines that make the machines sit abroad. China now produces roughly a third of the world's machine tools and just passed Germany as the largest exporter. A funded reshoring push still routes its deepest tool layer through foreign suppliers, which is the part a CHIPS-style subsidy doesn't reach.
MIIT ran an MLCC domestic substitution symposium this week — exact same category as your "stuff that makes the stuff." AI server demand is the stated rationale for accelerating local sourcing of capacitors, inductors, and magnets. China is building the capital stock for the stuff America is importing.
Love your “Big Software Move” chart. Very informative. Also, I give “zero-clicks” is an interesting phenomenon.
Put two of these numbers next to each other. Hyperscalers spend $600B+ on data centres this year. Construction spending on machinery factories, the actual stuff-that-makes-the-stuff category, is $3.9B. Roughly 150 to one in favour of buildings for thinking machines over buildings for machine-making machines.
And the two aren't separate stories. They compete for the same inputs, turbines, transformers, electricians, grid connections, all sold out for years, and the marginal unit of each goes to whoever pays most, which is always the data centre. The AI buildout is bidding those inputs away from the broader industrial one. The factory that would make machine tools loses its electricians to the server hall next door.
The 2013 reversal makes it compound, too. Import parts and you're still sovereign. Import the machines that make the machines, and expanding your own capacity runs on someone else's export licence. I'd say maybe 30% that ex-AI manufacturing investment meaningfully accelerates by 2028. The boom-let is real. It might also be eating the boom it was supposed to start.