The reported second phase of US semiconductor tariffs contains one idea that is structurally different from everything in the first, and it has been almost entirely lost in coverage focused on which gadgets get more expensive. Chip tariff exemptions would no longer be a category you fall into. They would be a volume you earn, in proportion to what you have committed to building in the United States. This piece covers what the January proclamation already does, what is being weighed now, how the proposed exemption mechanism works, why that makes it industrial policy rather than trade policy, the capital spending it would land on, and how firm any of it actually is.
What Phase 1 already did
The existing measure is narrow and specific, which is worth establishing before discussing an expansion of it.
Proclamation 11002, signed on January 14, 2026, imposed a 25% duty on a narrow set of advanced accelerators. Nvidia’s H200 and AMD’s MI325X were named explicitly. Tom’s Hardware reports this was labeled Phase 1 at the time, which implied more was coming.
Critically, that proclamation came with exemptions covering data centers, research and development, startups, and consumer devices. Those carve-outs are why the January measure has had limited practical effect on most buyers. A 25% duty that exempts data centers is not a data center cost.
That is the baseline against which everything below should be read. The current regime is a high rate on a small number of parts, with the largest buyers written out of it. Almost any change to the exemption structure matters more than a change to the rate.
What is now being weighed
The reported expansion changes the shape of the thing in two ways at once.
| Phase 1 (in force) | Reported Phase 2 | |
|---|---|---|
| Covers | A narrow set of advanced accelerators | Finished goods containing chips: data center servers, laptops, gaming consoles |
| Rate | 25% | Not settled in reporting |
| Exemptions | Data centers, R&D, startups, consumer devices | May not carry over; replaced by a volume cap |
The first change is scope. Duties on bare semiconductors are relatively easy to design around, because you can import the finished machine instead. Extending to servers, laptops and consoles closes that, and it is the difference between a policy that touches chip vendors and one that touches everyone who buys computers.
The second change is the more consequential one. Commerce officials have reportedly indicated in private talks that the January exemptions may not carry over. If the data center carve-out disappears while the scope expands to servers, the practical exposure goes from near zero to substantial in one step.
Chip tariff exemptions pegged to committed US production
Here is the mechanism, and it is genuinely novel.
Commerce Secretary Howard Lutnick is reported to favor a structure that would cap duty-free chip imports at a volume pegged to each company’s committed United States production.
Read that carefully, because it is not how exemptions normally work. A conventional exemption is categorical: you are a data center, data centers are exempt, you are exempt. This is quantitative and firm-specific. Your duty-free allowance is a number, and that number is a function of what you have promised to build domestically.
Two firms buying identical servers for identical purposes would face different costs, determined by their capital commitments rather than by anything about the transaction. A company with announced US fab investment imports a certain volume duty free. A company without imports the same servers and pays.
Why that is industrial policy rather than trade policy
The distinction matters for how you plan around it.
Trade policy adjusts the price of a category. You respond by substituting, resourcing, or absorbing the cost, and every firm in the category faces roughly the same calculation. It is a market-wide input change.
What is described here is a subsidy for domestic capital commitment, collected as a tariff on firms that have not made one. The lever is not what you buy or where you buy it from. It is what you have agreed to build. That converts tariff exposure into a function of corporate strategy decisions made years earlier, and it advantages incumbents with the balance sheet to commit to fabs over everyone else.
It also creates an unusual planning problem. You cannot hedge a tariff whose rate depends on your own capital allocation by changing suppliers. The only lever is the one that takes three years and several billion dollars to pull. For most buyers, including everyone reading this, there is no lever at all, only exposure to whichever side of the line your vendors sit on.
There is a second-order effect worth anticipating. If chip tariff exemptions become a firm-specific allowance rather than a category, the allowance itself acquires value, and anything with value tends to get traded, allocated internally, or priced into vendor contracts. Nothing in the reporting addresses whether allowances would be transferable, but it is the first question a finance team will ask, and the answer would determine whether this behaves like a quota system or simply a differentiated tariff. Industry groups have so far argued about the headline rate rather than the mechanism.
The number this would land on
The scale is what makes this more than a procurement footnote.
CNBC reported in February that big tech AI spending approaches $700 billion in 2026. Published estimates vary considerably, from around $630 billion at the low end to $760 billion at the high end depending on what is counted and who is counting, so treat the figure as an order of magnitude rather than a number.
A duty on imported servers lands directly on the hardware portion of that. Even a modest rate applied across that base is a large absolute sum, and it arrives on top of the component cost pressure we covered in our piece on server memory prices.
Jonathan McHale of the Computer and Communications Industry Association put the industry position plainly, saying the added costs put infrastructure investment at risk. That is a lobbying position and should be read as one, but the arithmetic behind it is not in dispute.
The downstream effect reaches ordinary hardware too. The accelerator market we mapped in the AI accelerator landscape is directly in scope, and consumer machines such as the Nvidia RTX Spark superchip and the M6 Mac mini sit in the finished-goods categories the expansion would cover.
How firm is any of this
Not very, and that should be stated clearly rather than buried at the bottom.
The underlying reporting is from POLITICO, based on eight unnamed sources familiar with internal talks. Everything downstream, including CNBC and Tom’s Hardware, is reporting on that reporting. No rule has been published, no rate has been set, and no proclamation has been signed.
A phase-in period is under discussion and the framework could still change substantially in the coming weeks. Sources indicated recent talks moved against the industry, which tells you the direction of travel but not the destination.
So this is a description of a proposal at the deliberation stage. It is worth knowing about because the exemption mechanism is unusual enough to change planning assumptions if it survives. It is not worth reorganizing a procurement cycle around yet.
What to do while it is unsettled
Three things, none of which are regrets if the proposal dies.
Find out whether your hardware vendors have committed US production. Under the reported mechanism that becomes a cost input rather than a press release. It is a reasonable question to ask a vendor now.
Ask what happens to quoted prices if duties change. Multi-year hardware contracts written before January 2026 may not address this at all. Knowing who absorbs a new duty is worth establishing before you need the answer.
Do not pre-buy on speculation. Nothing has been signed, a phase-in is under discussion, and buying a year of servers early against a rule that may not exist is a worse outcome than paying a duty that does.
The thing to watch for is a published rule rather than more sourcing on deliberations. Chip tariff exemptions under the current proclamation are categorical and written down; whatever replaces them, if anything does, will also have to be written down before it binds anyone. Until that document exists there is no rate to plan against, and the more definitive headlines are running ahead of what the sourcing supports.
Frequently Asked Questions
What are the current chip tariffs?
Proclamation 11002, signed January 14, 2026, imposed a 25% duty on a narrow set of advanced accelerators including Nvidia’s H200 and AMD’s MI325X. It came with exemptions covering data centers, research and development, startups and consumer devices.
What would the second phase cover?
Reporting indicates duties would extend from bare chips to finished goods containing them, specifically naming data center servers, laptops and gaming consoles. The rate has not been settled in any published reporting.
How would the new exemption mechanism work?
Commerce Secretary Howard Lutnick is reported to favor capping duty-free chip imports at a volume tied to each company’s committed US production. Rather than a categorical carve-out, each firm would have a quantitative allowance based on what it has agreed to build domestically.
Would the data center exemption survive?
Reportedly it may not. Commerce officials have indicated in private talks that the exemptions attached to January’s measure may not carry over, which is the single change that would matter most to infrastructure buyers.
Has any of this been enacted?
No. The original reporting is from POLITICO citing eight unnamed sources on internal deliberations. No rule has been published and no rate set. A phase-in period is under discussion and the framework could change substantially.
How much spending would this affect?
Big tech AI capital spending in 2026 approaches $700 billion by CNBC’s reporting, though published estimates range from roughly $630 billion to $760 billion. Duties on imported servers would land on the hardware share of that.
Can a buyer do anything to reduce exposure?
Very little directly, because the proposed lever is domestic fab investment rather than sourcing. What is available is asking vendors about their US production commitments and confirming who absorbs new duties under existing contracts.
Should we accelerate hardware purchases?
Not on this reporting. Nothing is signed, a phase-in is under discussion, and pre-buying against a rule that may never take effect carries its own costs in depreciation and stranded capacity.
