The Section 232 Chip Tariff's 100-Megawatt Loophole: Why a 25% Levy on H200 and MI325X GPUs Exempts Hyperscalers by Design

Nobody had to lobby for a hyperscaler carve-out in the Section 232 chip tariff. It was written directly into the tariff schedule as a number: 100 megawatts. Clear that threshold and every H200 or MI325X you import is duty-free. Fall short of it — which every self-hosted PaaS operator, every SMB buying its own GPU fleet, and nearly every company on Earth does by definition — and you pay the full 25% on every chip, every time.

That's not a loophole anyone gamed. It's the exemption's design point, and it's worth understanding precisely, because the internet's shorthand for this tariff ("hyperscalers got a sweetheart deal") is only half right, and the half that's wrong matters for anyone doing real cost math on owned GPU hardware in 2026.

What the tariff actually taxes​

On January 14, 2026, the White House issued Proclamation 11002 under Section 232 of the Trade Expansion Act, imposing a 25% ad valorem duty on a narrow category of advanced computing chips, effective for goods entered for consumption starting January 15, 2026.

The tariff isn't brand-specific — it's defined by performance thresholds. Logic integrated circuits under HTS subheadings 8471.50, 8471.80, or 8473.30 are covered if they meet one of two parameter sets:

• Set A: Total Processing Performance (TPP) 14,000–17,500 combined with DRAM bandwidth of 4,500–5,000 GB/s
• Set B: TPP 20,800–21,100 combined with DRAM bandwidth of 5,800–6,200 GB/s

The White House fact sheet names Nvidia's H200 and AMD's MI325X as the chips that land in these bands. A chip outside both ranges — an older H100, or a next-gen part engineered to sit just above or below the thresholds — isn't covered at all, which is its own story for a different post.

For the chips that are covered, the math is straightforward. H200 GPUs run roughly $30,000–$45,000 depending on configuration and supplier, call it $31,000 as a representative single-unit price. A 25% tariff adds $7,500–$11,250 per GPU — about $7,750 on the representative price, landing around $38,750. Multiply by eight for a standard HGX-class server and the tariff alone adds roughly $60,000–$90,000 to a system that already ran $250,000–$350,000 before duty.

The exemption that actually matters, in HTS-code specificity​

The proclamation doesn't just tax; it carves out categories, each with its own Chapter 99 HTS code:

Code 9903.79.03 is the one doing the real work, and it's worth translating "100 megawatts" into hardware. An H200 draws up to 700W; a fully loaded 8-GPU HGX server, GPUs plus networking, storage, and cooling overhead, pulls somewhere around 10kW under load. A 100MW facility is therefore built around roughly 10,000 servers — on the order of 80,000 H200-class GPUs of continuous, simultaneous draw. That's not a mid-size colo footprint or a self-hosted PaaS's GPU node pool. It's a hyperscale campus — the kind of facility Amazon, Google, Microsoft, and a handful of AI-infrastructure specialists like CoreWeave operate, and essentially nobody else does.

So the exemption isn't a favor granted to specific companies. It's a bright-line physical threshold that only entities already operating at hyperscale can clear, by construction. A self-hosted PaaS importing a rack of GPUs into a US facility doesn't come close, and there's no version of "grow a little" that gets it there — 100MW is roughly the draw of a small city's data center district, not a scaling milestone.

The other carve-out — reported, not signed​

Separately, there's a second story that's been widely reported but is not yet policy: the Commerce Department has reportedly been preparing tariff carve-outs specifically for Amazon, Google, and Microsoft, tied to TSMC's $165 billion commitment to build chip fabrication capacity in Arizona. Under the reported mechanism, TSMC would be able to extend tariff relief to its US hyperscaler customers, scaled to the fab capacity TSMC brings online domestically.

As of the most recent reporting available, an administration official described the plan as "in flux" and not yet signed by the President — this is a proposed mechanism, not an operative exemption. It's easy to conflate this story with the 100MW carve-out above, because both end with the same three companies paying less, but they're different things: one is a specific, still-pending political deal; the other is a generic, already-in-effect threshold in the proclamation's own text that happens to only be reachable at hyperscale. If the TSMC-linked deal is eventually signed, it stacks on top of an exemption those same companies mostly already qualify for on data-center scale alone.

Three buyers, one tariff, three different bills​

Put concrete numbers next to three ways of getting GPU compute in 2026:

A hyperscaler renting to you. AWS, Google Cloud, and Microsoft Azure import chips into 100MW-class facilities that clear the exemption outright (and may eventually layer the TSMC-linked carve-out on top). The tariff doesn't show up in your invoice because it was never charged on their import. H200 on-demand pricing at the major clouds runs roughly $5–$11/GPU-hour in 2026 — that price reflects scarcity and margin, not tariff pass-through.

An SMB or self-hosted PaaS importing GPUs into a US facility. If you need GPU capacity physically located in the US — for latency, data residency, or a customer contract that requires it — and your fleet is nowhere near 100MW (it isn't; almost nobody's is), you pay the full 25% on landed cost. That's the ~$60,000–$90,000 tariff bill on an 8-GPU server from the math above, due at the border, before you've served a single request.

A self-hosted PaaS running non-US-import hardware. Section 232 is a US import duty — it only bites when covered chips cross into the United States. A fleet built on European bare metal (Hetzner and similar providers run GPU-capable and CPU-only node pools out of German, Finnish, and other EU facilities) never triggers a US import event. Company size is irrelevant here; the tariff simply doesn't apply, for the same reason a VAT rate in Singapore doesn't apply to a purchase made in Ohio.

What the tariff does to the "buy beats rent" argument​

The self-hosted pitch for owned hardware has always been a breakeven argument: the upfront capital cost of a GPU server pays for itself against cloud rental after some number of months, and every month after that is pure margin versus renting. The tariff doesn't break that argument — it's a flat 25% surcharge on the capital side, so it stretches the breakeven timeline proportionally, by exactly 25%, regardless of which rental price you're comparing against.

Take a single H200 at 70% utilization — a reasonable stand-in for "sustained but not maxed-out" production load, giving about 6,130 effective hours a year:

(This is a simplified capital-recoup comparison against list rental prices — it excludes power, hosting, and staff costs, which apply on both sides of the ledger roughly equally and don't change the delta the tariff adds.)

The pattern holds across every rental benchmark: owning still beats renting eventually in every column, and the tariff never flips that conclusion — it just costs you an extra 20–30% of however long the breakeven already was. Against a cheap specialist cloud, that's six extra months of paying rental rates before the owned box turns a corner. Against premium hyperscaler pricing, it's closer to six extra weeks. Either way, the tariff is a real, quantifiable tax on the "buy instead of rent" argument, not a rounding error and not the argument's death sentence.

Where the advantage is untouched​

Two situations sidestep this tariff entirely, and they're both common enough to matter:

CPU-only fleets. A self-hosted PaaS running application workloads, background jobs, and most production traffic on CPU-only Hetzner-class hardware never touches a chip that meets the TPP/DRAM thresholds in Proclamation 11002. Zero exposure, zero math required.

Non-US GPU capacity. If your GPU workload doesn't have a hard US-residency requirement — training jobs, batch inference, agent sandboxes that don't serve latency-sensitive end users — sourcing GPU-capable node pools from EU or other non-US bare-metal providers keeps the entire fleet outside Section 232's jurisdiction, at any company size.

The tariff only bites the specific, narrower case: US-region GPU capacity, owned rather than rented, at a scale nowhere near 100MW. That's a real cost, and it's worth pricing into any node pool that needs to check that box. It's also a shrinking share of what a self-hosted PaaS actually needs to run.

Bex.co is the open-source, AI-native Render alternative — push a git repo, get a running HTTPS service on machines you own. When those machines are CPU-only or sit outside US jurisdiction, tariffs like this one are somebody else's problem. Star the repo on GitHub or deploy your first app today.

Sources​

• Presidential Proclamation 11002 — Adjusting Imports of Semiconductors
• Section 232 Semiconductor Tariff: Rate, HTS Codes (9903.79.01–.06) & Exemptions — GingerControl
• President Trump orders narrowly targeted 25% Section 232 tariff on certain advanced semiconductor articles — White & Case
• US plans Big Tech carve-out from next chip tariffs, FT reports — Yahoo Finance
• Trump considering tariff carve-out for hyperscalers linked to Taiwanese chip investments — Data Center Dynamics
• NVIDIA H200 Price Guide 2026 — Jarvislabs
• NVIDIA AI GPU Prices: H100 ($27K-$40K) & H200 ($315K/8-GPU) Cost Guide — IntuitionLabs
• H200 Power Consumption: A Complete Guide — TRG Datacenters
• H200 Cloud Pricing: Compare 33+ Providers (2026) — GetDeploying