The accelerators at the heart of the AI buildout spent most of 2026 crossing the border close to duty-free. Memory, not customs, raised your server quote.
If you priced an eight-GPU server in January and again in September, the quote went up. Most procurement teams have a one-word explanation ready for the CFO: tariffs. It is the wrong word for the compute line, and getting it wrong leads to the wrong hedges.
Here is what actually happened in 2026. The headline 25 percent chip tariff exempted chips bound for US data centers. The Supreme Court struck down the broad emergency tariffs in February. The metal tariffs were rewritten in April to explicitly exclude GPU boards, accelerator cards, and server racks. The accelerators at the heart of the AI buildout spent most of the year crossing the border close to duty-free, while memory prices, not customs, did the damage to server quotes.
Tariffs did change data center procurement, just not where most buyers are looking. The duty landed on transformers, switchgear, cooling, and structural steel. The bigger change is contractual: vendors now reprice until shipment, quotes expire in days, and a pending policy decision could put assembled servers in scope for the first time. This article maps where the tariff exposure really sits, models what a server tariff would cost, and lays out the procurement moves that hold up under every scenario.
The legal scaffolding changed four times in seven months. Each change moved the risk around the bill of materials rather than simply raising it.
On January 14, 2026, a Section 232 proclamation imposed a 25 percent tariff on certain advanced computing chips, naming the NVIDIA H200 and AMD MI325X, effective January 15 (Thomson Reuters, 2026). The same proclamation exempted chips imported for US data centers, along with repairs, research and development, startups, consumer and industrial uses, and the public sector. In practice it worked as a tax on chips leaving the country more than on chips staying in it. Taiwan signed a companion trade agreement letting companies building US fabs import up to 2.5 times their planned capacity tariff-free during construction, falling to 1.5 times once production starts (GEODIS, 2026).
On February 20, the Supreme Court ruled 6 to 3 that the International Emergency Economic Powers Act (IEEPA) does not authorize tariffs (Perkins Coie, 2026). The administration replaced them within hours with a 10 percent global surcharge under Section 122 of the Trade Act of 1974, a tool capped at 150 days. When that clock ran out on July 24, Section 301 duties of 10 to 12.5 percent on 60 economies took over the same night, with no expiry date (Winston Taylor, 2026). Section 232 goods sit outside that regime.
Steel and aluminum duties sit at 50 percent, and copper has carried a 50 percent tariff since August 2025. The April 2026 amendment explicitly excluded GPU boards, accelerator cards, and server racks from the derivative product lists, according to a May 2026 analysis by the Center for Strategic and International Studies (CSIS, 2026). Structural steel and cooling units remained exposed.
The January proclamation required Commerce to report on the data center chip market by July 1. As of late August, that report had not been published (The Next Web, 2026). On August 27, Politico reported, citing eight people familiar with the talks, that a second round could cover finished products containing chips, including data center servers (Anadolu Agency, 2026). On September 2, Commerce Secretary Howard Lutnick confirmed the direction: companies producing in the US would get tariff relief, and those that do not would pay (The Korea Times, 2026). No rate, scope, or start date has been published.
| Date | Action | Rate | Data center impact |
|---|---|---|---|
| Jan 15 | Section 232 chip tariff | 25% | Exempt |
| Feb 24 | IEEPA struck down, Section 122 surcharge | 10% | Temporary, 150 days |
| Apr 2026 | Metal tariff amendment | 50% | GPU boards, racks excluded |
| Jul 24 | Section 301 replaces Section 122 | 10-12.5% | 60 economies, no expiry |
| Sep 2 | Phase Two chip tariff confirmed | Not set | Servers may be covered |
CSIS estimates that roughly 54 cents of every data center dollar goes to semiconductors, with servers alone making up about 52 percent of total capital expenditure (CSIS, 2026). That is exactly the slice 2026 policy shielded. Bloomberg made the same point in its August 8 New Economy newsletter: the AI buildout was largely routed around the tariff wall, with most GPUs entering duty-free and the remaining exposure concentrated in power and cooling equipment.
The facility side is a different story. Transformers, switchgear, uninterruptible power supplies (UPS), and cooling plants carry steel, aluminum, and copper, and even domestically assembled units lean on imported subcomponents. CSIS cites industry reporting that "Made in USA" cooling systems typically contain 30 to 60 percent imported content. Tariffs are not even the worst of it. A September 2026 research update from NextMSC on 100 MW data hall costs concluded that Section 232 is not the dominant cost driver, pointing to transformer lead times of about 36 months as the bigger constraint (NextMSC, 2026).
The procurement lesson: if your tariff budget assumes duty on GPUs, you are probably over-reserving on compute and under-reserving on electrical gear. Electrical gear also has the longest lead times, so an order placed late costs both the duty and the delay.
| Component | Duty exposure | Main cost driver | Procurement move |
|---|---|---|---|
| GPU servers | Low | Memory prices, Phase Two risk | Repricing caps, offset question |
| Storage (NVMe, DRAM) | Low | DRAM and NAND inflation | Lock volumes early |
| Transformers, switchgear | High | Lead times about 36 months | Order first |
| Cooling, structural steel | High | Imported subcomponents | Quote landed cost |
Memory did what tariffs did not. TrendForce projected on September 30 that conventional DRAM contract prices would rise another 10 to 15 percent quarter over quarter in Q4 2026, with NAND Flash up 15 to 20 percent and server DRAM still undersupplied (TrendForce, 2026). HPE CEO Antonio Neri told analysts that elevated memory prices would persist well into 2027 (CRN, 2026). This is the same supply squeeze behind the NVMe price shock our upcoming NVMe shortage report covers in depth.
The result is the procurement change that matters most in 2026, and it is contractual rather than tariff-driven. In February, HPE introduced terms allowing it to adjust server and GreenLake prices up to the date of shipment (TechRadar, 2026). Gartner analyst Rangit Atwal told Channel Dive that the entire supply chain has shifted to short-term contracts because of memory prices and uncertain CPU availability (CIO Dive, 2026). A signed purchase order is no longer a fixed price unless you negotiate it into one, which is one more hidden cost buyers now have to read for.
This is where the tariff and memory stories converge. Vendors built repricing rights for memory, but the same clause lets them pass through any Phase Two duty that lands between order and shipment. If your contract allows repricing until shipment and a server tariff takes effect while your order is in the queue, you pay it.
The most widely quoted estimate comes from the Computer and Communications Industry Association (CCIA). Its June 2026 analysis multiplies the compute share of data center spending (78 percent) by the imported share of that compute (80 percent) by a 25 percent tariff, producing an effective 15.6 percent tax on US data center construction and roughly $450 billion of 2026 to 2030 capacity at risk of cancellation, delay, or relocation (The Next Web, 2026). Treat it with care. CCIA counts the companies that would pay the tariff among its members, and the 80 percent import share is the assumption carrying the most weight. CSIS, which received CCIA support for its brief, reaches a similar order of magnitude: an additional $1.4 trillion burden on the buildout under a 100 percent tariff on all semiconductors and chip-containing products (CSIS, 2026).
For a single buyer the arithmetic is simpler. Take a $10 million order of imported GPU servers and apply the rates that have been floated, assuming full pass-through and no tariff offset allowance. At today's Phase One treatment the duty is effectively zero. At 10 percent it is $1 million. At 25 percent it is $2.5 million. At 50 percent, the rate already applied to steel, aluminum, and copper, it is $5 million, half the original hardware budget. The offset program changes who pays: an OEM or hyperscaler with a qualifying US manufacturing commitment could import under an allowance, while a mid-tier cloud or enterprise buying through a non-qualifying channel would not.
That asymmetry is the real 2026 story. Phase Two, as described, favors buyers whose suppliers have US fab commitments and penalizes everyone else. Which channel your servers come through may soon matter as much as which GPU is inside them.
Split every bill of materials into compute, storage, networking, and facility lines, and assign each its own duty assumption. Compute is currently near zero with a binary Phase Two risk. Facility electrical gear carries steady duty plus extreme lead times. Storage carries memory inflation rather than tariff exposure. One blended tariff percentage across a project hides all three.
The data center exemption depends on end-use certification at entry. If the exemption is tightened rather than removed, the burden of proof rises. Keep site, facility, and deployment records ready for every accelerator shipment, and confirm your broker files under the correct HTSUS lines. Duty treatment follows the entry date, not the order date, so a late shipment can land under a new rate.
Ask for a cap on post-order repricing, a carve-out that excludes new duties from pass-through on orders placed before an announcement, or a right to cancel without penalty if landed cost rises beyond a threshold. Vendors that will not move on any of these are telling you who carries the risk.
Add one question to every hardware RFP: does your manufacturing footprint qualify for tariff relief under the announced offset framework? It belongs alongside the first question we ask every GPU cloud: do you own the GPU? Both determine whether the price you are quoted is the price you will pay.
Renting GPU capacity from a neocloud such as CoreWeave, Nebius, or Lambda Labs moves the import decision onto the provider's balance sheet. It does not make the cost disappear: providers buying new fleets will pass Phase Two duty into hourly rates over time, and fleets already deployed are the ones shielded. Short contracts and multi-provider strategies, which already make sense as GPU supply loosens, gain another argument. Remember the data cost of switching: every provider move drags data with it, and egress fees on that data can wipe out the savings that prompted the move. Zero-egress storage keeps that option cheap, which is why storage is the anchor of the AI stack.
Three signals will decide whether 2027 budgets need a tariff line for compute. First, publication of the Commerce data center chip report, which determines whether the exemption survives, narrows, or disappears. Second, the Phase Two proclamation itself: its rate, whether servers are covered as derivative products, and the length of any phase-in. Third, the offset rules, specifically whether allowances attach to the chipmaker, the server OEM, or the importer of record. Until all three are known, the defensible position is to keep compute orders flexible, lock electrical gear early, and push repricing risk back onto vendors wherever you can.
In 2026 the GPU was the cheapest thing to import and the transformer was the most expensive. Budget for the tariff you have, not the one in the headlines, and write the one that is coming out of your contracts.
Compare AWS, Google Cloud, Azure, and alternatives like Backblaze B2 Discover how much you could save in seconds
