The Chip Tariff That Hasn't Hit Capex

Chip stocks opened higher on Nvidia’s print. The same morning’s other headline was a Politico report that the Trump administration is considering a new round of semiconductor tariffs that could reach products made with chips — laptops, consoles, and the servers that fill data halls. Commerce Secretary Howard Lutnick, the report said, wants relief tied to investment in U.S. fabrication. The framework is unannounced, still revisable, and may include a phase-in.

That is the dominant pairing: AI hardware investment versus White House duties. It treats three different instruments as one tax.

What is already law, and what is still a leak

Proclamation 11002, signed January 14, imposed a 25 percent Section 232 duty on a “very narrow category” of advanced computing chips — the White House named Nvidia H200 and AMD MI325X as examples — effective the next day. The same clause exempts imports for U.S. data centers, repairs, research, startups, consumer and industrial uses, and public-sector applications. The 25 percent is the legal form of a China-bound transit tax: BIS rules require certain chips destined for the People’s Republic to be tested in the United States first, which makes them imports. Nvidia’s $108 billion guide already assumes zero data-center compute from China. The duty that is actually collecting is not the one sitting on hyperscaler racks in Virginia or Oregon.

August 6 added a second signed instrument: minimum import prices on polysilicon and a 15 percent tariff on downstream derivatives, effective December 4. Semiconductor-grade polysilicon is 2.4 percent of global poly output; the proclamation is mostly a solar-onshoring tool with a national-security caption. Allies (Japan, Korea, Taiwan, the EU) see the 15 percent capped against Column 1 rates. That is not a GPU-hall tax either.

The investment that can be delayed is the unsigned one: a duty on the server, not the China-bound H200.

The residual is why a leak about phase two is being written as if it had already slowed capex when Wednesday’s books show the opposite. Revenue rose 18 percent sequentially. Gross margin held at 75 percent. Guidance stepped to $108 billion. Whether those GPUs are consumption or crates is a separate ledger. Tariffs are not yet on it.

The mismatch is years, not basis points

January’s proclamation already sketched phase two: after trade talks, “broader tariffs on semiconductors, at a rate of duty that is significant,” plus an offset for firms investing in U.S. production. Thursday’s leak is that sketch with servers attached. People close to the administration, in the same reporting, put meaningful U.S. capacity expansion at least five years out. Grace periods on other 232 regimes have often been shorter. If Lutnick’s bargain is tariff-free import quotas against pledged domestic output, the cost hits imported accelerators and assembled racks first. The fab shell comes later. Custom silicon at Broadcom is one hyperscaler hedge against that queue. Beijing’s five-year plan is the other geography’s answer to the same compute race.

What prices still mis-specify is a single “tariff risk” overlay on Nvidia. A 25 percent that does not apply to U.S. data-center imports cannot be the reason a $89 billion quarter exists. A 15 percent poly duty in December cannot be the reason TSMC wafers move. An unsigned expansion onto servers could be — if it is proclaimed, if data-center carve-outs die, and if phase-in is shorter than a fab. The capex curves already rhyme with an older boom; adding a draft duty does not make them a bust.

The testable claim is the next Federal Register notice, not the next print. If a signed 232 lists HTS lines for data-center servers and removes the U.S.-buildout exemption, then Thursday’s leak becomes the tax the wire already named. If the proclamation keeps the January carve-out and only widens the China-transit net, Nvidia’s $108 billion guide and the tariff story can coexist because they were never the same object.