Too Big to Tariff

The US trade deficit in goods hit $1.2 trillion in 2025. A record. And the single largest contributor wasn’t oil, or cars, or consumer electronics in the traditional sense. It was AI hardware.

Imports of computers, accessories, and semiconductors surged 60% in twelve months — over $450 billion worth of chips and servers flowing in from Taiwan, South Korea, and Japan. The infrastructure buildout that Jensen Huang spent three hours presenting at GTC this week? Most of it gets manufactured in Asia.

Here’s the catch-22: the AI boom is one of the few things keeping the US economy growing. JPMorgan estimated that AI-related capital spending outpaced consumer spending as a growth factor in 2025. Total AI spending is projected to hit $700 billion in 2026. Kill that, and the economy loses one of its few remaining engines.

But every dollar of that spending makes the trade deficit worse. The very thing keeping the economy alive is destroying the president’s signature economic metric.

Trump wants tariffs to shrink the deficit. But tariffing AI hardware would be like taxing your own oxygen supply. You can’t build domestic AI capacity — the thing that might eventually reduce the deficit — while making the imports needed to build that capacity more expensive. The CHIPS Act, which Trump wants Congress to kill, was specifically designed to solve this problem by bringing semiconductor manufacturing onshore. But it was Biden’s law, and apparently that matters more than the economics.

So we’re left with a policy paradox that reveals something deeper about the AI moment: AI infrastructure has become too systemically important to subject to normal trade policy.

“Too big to fail” was about banks in 2008. “Too big to tariff” is about compute in 2026. The dependency is different but the structure is identical — an industry so central to the economic system that the government can’t apply its own rules to it without risking collapse.

Brad Setser at the Council on Foreign Relations put it simply: the construction of AI data centers is “very import intensive,” and “it’s hard to see how this doesn’t result in a larger U.S. trade deficit in 2026.”

The irony compounds when you consider what’s being imported. These aren’t finished consumer products. They’re capital goods — machines that build the capability to eventually produce value domestically. In theory, today’s imports become tomorrow’s exports (AI services, models, inference). But the transition period creates a window where the deficit balloons before it could ever shrink.

That window is measured in years. TSMC’s Arizona fab won’t reach volume production until 2028 at the earliest. Intel’s foundry ambitions are on life support. Samsung’s Texas fab is delayed. The Foxconn investments Trump announced are mostly assembly, not leading-edge fabrication.

So the administration is stuck in a temporal mismatch: it needs the benefits of AI investment now, but the import substitution won’t happen for years. And in the meantime, every new data center announcement — every Stargate, every xAI Memphis facility, every Meta $65B capex commitment — adds to the very number the president built his economic identity around reducing.

The deepest irony: manufacturing construction spending has fallen every month since Trump took office. Factory activity is in retreat. Jobs are disappearing, not appearing. The tariffs that were supposed to bring manufacturing back have instead created enough uncertainty to suppress the very investment they were meant to encourage.

AI is the exception. AI investment keeps flowing because the opportunity cost of not investing is perceived as existential. No CEO wants to be the one who paused their AI buildout because of tariff uncertainty. So the money moves, the imports surge, and the deficit grows.

Too big to tariff. The three most expensive words in trade policy.