The 25% Tax on AI: How Chip Tariffs Are Sorting the Industry

The US just put a 25% tariff on the chips that power AI. It carved out an exemption for data centers drawing 100MW or more for AI, training, or simulation workloads. Every major hyperscaler clears that line. Most startups don't. The tariff lands three years before domestic chip production can replace the imports it taxes. That gap is where the sorting happens.

The Structure: Who Pays and Who Doesn't

On January 15, 2026, Section 232 tariffs went into effect on advanced AI chips. NVIDIA H200s, AMD MI325X, and anything meeting specific performance thresholds now carry a 25% ad valorem duty at the border.

The exemption list looks generous on paper: US data centers over 100MW, R&D use, startup use, public sector, repairs. But the 100MW data center exemption is self-certifying for companies that already operate at scale. The startup and R&D exemptions require end-use certifications through Commerce and CBP. The process for those certifications? Undefined. Commerce hasn't published Federal Register guidance.

So the companies with 500MW campuses import chips tariff-free today. The companies trying to stand up their first 5MW cluster wait for a bureaucratic process that doesn't exist yet.

The Math: 50-75% Cost Increases Aren't Hypothetical

Semiconductors represent over 50% of total AI server cost, according to SemiAnalysis. A 25% tariff on chips is already painful. But chips aren't the only imported component in a data center.

Power infrastructure, cooling systems, and networking equipment face their own tariff regimes, covering an additional 25-30% of total facility cost. When you stack the chip tariff on top of tariffs on transformers, chillers, and switches, non-exempt buyers face a compounded 50-75% increase in total infrastructure cost.

Here's the cost anatomy for a typical AI data center:

Component	% of Total Cost	Tariff Exposure
GPUs/Accelerators	40-50%	25% Section 232
Other semiconductors	10-15%	Phase 2 target
Power infrastructure	15-20%	Separate tariff regimes
Cooling systems	15-20%	Tariff-affected imports
Networking equipment	5-10%	Tariff-affected
Land/construction	5-10%	Minimal exposure

A single AI server rack with multiple GPUs already exceeds $500K. A 10MW GPU cluster needs $5-20M in cooling infrastructure alone. These aren't rounding errors.

The 100MW Line: A Class Divide in Kilowatts

The 100MW exemption threshold creates three tiers:

Hyperscalers (insulated). Amazon, Alphabet, Meta, Microsoft, and Oracle are on track for $660-690 billion in combined AI infrastructure capex in 2026, nearly double 2025 levels. (I wrote about the disconnect between that spending and actual GDP returns last week.) They self-certify into the 100MW exemption. They negotiate volume GPU pricing. They locked in long-term supply contracts before the tariff took effect. And when costs rise anyway, they pass them through to cloud customers.

Mid-tier AI companies (squeezed). Companies like OpenAI and Anthropic operate their own clusters or lease dedicated capacity. They face an estimated 15-25% compute cost increase flowing through cloud pricing. They can negotiate, but they don't have hyperscaler leverage. A $200M training run that becomes $250M is survivable. The margin for error just got thinner.

Startups under $10M (priced out). Price-takers on hardware. Too small for the 100MW exemption. The startup exemption exists in theory but lacks a certification path in practice. Self-hosted infrastructure costs 50-75% more. The alternative is renting cloud compute from the same hyperscalers who got the exemption, paying both the tariff premium baked into cloud pricing and the cloud margin on top.

Andrew Ng put it directly: when regulations change overnight by tweet, other geographies with more stable structures become more attractive. Data center expansion in Malaysia, Singapore, and Europe is accelerating.

The Double Tax: Tariff Premium Plus Cloud Margin

This is the second-order effect worth watching. Startups priced out of buying their own hardware get pushed onto cloud compute. That compute runs on hardware the hyperscaler imported tariff-free. The hyperscaler charges you for compute at rates that reflect their scale advantages, plus their margin.

So the startup pays a cloud price that includes the hyperscaler's cost basis (low, because exempted) plus the hyperscaler's margin (high, because where else are you going?). The tariff didn't raise costs equally. It raised costs for the buyers who already had the least leverage, and it funneled their spending to the buyers who already had the most.

Roughly 1,000 AI startups with budgets under $10M are navigating this right now. Their strategic options are narrowing to three: go cloud-dependent, go overseas, or go smaller.

BIS: The Agency That Can't Keep Up with Its Own Policy

The Bureau of Industry and Security administers both export controls and tariff exemptions. It has lost 101 employees since 2024, a 19% reduction. Twenty percent turnover among rulemaking and licensing staff specifically. Ten of twelve senior Export Administration leadership roles have turned over since early 2025.

License processing is down roughly 25%. NVIDIA hasn't shipped a single H200 to China, months after the White House approved the deal. The agency got a 23% budget increase for FY2026, but most of that funds enforcement operations, not the licensing staff clearing backlogs.

This matters for the tariff story because tariff exemption certifications go through the same depleted pool of staff. A startup trying to qualify for the R&D or startup exemption is filing paperwork with an agency that can barely process its existing workload. The exemptions exist on paper. The administrative capacity to grant them doesn't match the demand.

AI Trade by the Numbers

Michael Waugh's NBER Working Paper 35053 measured trade in AI-related products by using LLM classification of HS10 customs codes. The findings:

• AI-related products now represent 23% of all US imports (2025)
• AI import growth since 2023: 73%
• Non-AI import growth over the same period: 3%
• Mexico and Taiwan account for roughly 50% of all US AI-related trade

The divergence started in early 2024, coinciding with the GPU buildout surge. Without the AI import boom, the US goods trade deficit would have been nearly $200 billion smaller in 2025. The tariff regime targets the single fastest-growing import category in the US economy. AI-related products grew 24x faster than everything else, and the policy response is to tax them.

The Domestic Production Timeline Doesn't Match

TSMC has committed $165 billion to Arizona expansion: six fabs, two packaging facilities, and an R&D center. The $250B Taiwan trade agreement locks in investment commitments. This is the CHIPS Act's carrot. The Section 232 tariff is the stick.

The problem is timing. TSMC's first Arizona fab is already producing 4nm chips for Apple and NVIDIA. The second fab targets 2027 production, ahead of schedule. But full buildout across all six planned fabs, plus packaging and testing capacity, stretches to 2029 and beyond. And there are still no dominant US players in downstream assembly, testing, or packaging. The supply chain being taxed today won't have a complete domestic replacement for three more years.

So the tariff protects a domestic industry that doesn't exist at scale yet. The cost falls on companies buying chips today. The benefit accrues to fabs that aren't operational until late decade.

Phase 2: The Report on the President's Desk

On April 14, 2026, Commerce delivered its Phase 2 report recommending broader tariffs on all semiconductors and manufacturing equipment at rates described as significant. This could expand well beyond advanced AI chips to cover memory, analog, and commodity semiconductors.

Phase 2 also proposes a tariff offset program for companies investing in US semiconductor production. If you're building fabs in America, your import costs go down. If you're just buying chips to train models, they don't.

For context: the US-EU deal caps semiconductor tariffs at 15% for EU imports. Taiwan, South Korea, and Japan are still negotiating. The July 1 review will determine whether Phase 1 rates get modified based on those outcomes. Memory prices are already surging 30-50% quarterly through the first half of 2026, independent of tariff effects. Phase 2 arriving on top of that would compound an already expensive year for anyone buying compute.

Training Economics: The Narrowing Window

Frontier model training costs have been on an exponential curve since 2017. From roughly $1K for early transformer work to $200M for GPT-4 class models. DeepSeek V3 at $5.6M represents the floor for meaningful frontier research. A 25% hardware cost increase on that curve doesn't stop the largest labs. It stops the next ten that might have tried.

The annual growth rate in training costs runs 2-3x. Every year, the minimum viable budget for frontier work climbs. The tariff adds a percentage on top of a number that's already doubling. For a well-funded lab spending $200M, a 25% hardware premium means $50M more. For a startup that scraped together $8M, it means the project doesn't happen.

What This Looks Like in 18 Months

The consolidation pattern is already visible. Startups migrate to cloud, giving hyperscalers more revenue and more data about what workloads matter. Mid-tier companies negotiate harder on cloud pricing and explore overseas capacity. Hyperscalers absorb the tariff through their exemptions, pass cost through to cloud customers, and collect both the infrastructure advantage and the margin.

By the time domestic production arrives at scale in 2029, the market structure will reflect three years of tariff-driven consolidation. The companies that survived will be the ones with either enough capital to absorb the premium or enough cloud dependency that the hyperscaler absorbed it for them.

The tariff didn't create this dynamic. The $660-690B capex gap between Big Tech and everyone else was already staggering. But trade policy is supposed to strengthen the domestic industry, not accelerate its consolidation. A 100MW exemption floor in an industry where 95% of participants operate below that threshold does one thing well: it sorts.

And the sorting has already started.