The next cost shock for AI infrastructure may not come from Nvidia, cloud demand, or another shortage of high-bandwidth memory. It may come from trade policy. US Trade Representative Jamieson Greer said Friday that the administration is still weighing tariffs on imported semiconductors, even as he signaled there are no immediate plans to impose new levies.

That distinction matters. For chipmakers, it says tariffs are still part of the policy toolkit. For AI companies, it says hardware budgets are now tied even more closely to Washington's industrial strategy. If the US moves from subsidies and export controls toward broader import duties, the economics of building data centers, buying servers, and deploying edge AI systems could change quickly.

According to Reuters, Greer made the comments at a Micron Technology event in Manassas, Virginia, where the company marked the start of 1-alpha DRAM manufacturing, described by Micron as the most advanced memory produced in the United States. Greer said semiconductor tariffs are important, but timing and scale matter even more. That is not a retreat. It is a warning that the policy is being sequenced.

The obvious target is imported chip supply. The practical impact would be broader. AI startups rarely buy bare semiconductors directly from foundries. They buy finished servers, accelerator systems, networking gear, storage, and edge devices that carry chips through several layers of the supply chain before they reach a deployment site.

That is where tariff exposure gets complicated. A startup leasing capacity from a hyperscaler may feel the effect later, through higher cloud pricing or tighter access to premium GPU clusters. A company building its own inference stack may feel it sooner, especially if it is sourcing appliances or specialized hardware that depends on chips fabricated or packaged in Taiwan, South Korea, or other major Asian supply hubs.

Legacy enterprise hardware buyers have a different problem. They usually have more procurement leverage, longer contracts, and established vendor relationships. But they also buy at scale and plan refresh cycles years ahead. A tariff structure that raises the landed cost of AI servers or networking equipment could force CIOs to delay upgrades, renegotiate supplier terms, or shift workloads back toward cloud providers with deeper purchasing power.

None of this means tariffs would automatically stop AI investment. Demand is too strong. Companies are still chasing faster training, cheaper inference, and more control over proprietary data. But it could separate firms with flexible capital from those already stretching their balance sheets to secure compute. In AI, the cost of hardware is not a side issue. It defines who can experiment, who can scale, and who has to wait.

Domestic capacity is moving, but not fast enough

The political logic behind tariffs is easy to understand. The US wants more chip production at home, and the CHIPS Act gave companies subsidies to help make that happen. Tariffs would add pressure from the other side by making some foreign supply less attractive. The problem is that semiconductor capacity does not appear just because policy wants it to.

Micron's Virginia announcement gives the administration a useful example. Memory is essential to AI systems, and domestic production of advanced DRAM supports industries from defense to networking and medical devices. But Micron has said only about 2% of the world's memory chips are manufactured in the US, all at its Manassas facility. That shows both progress and the size of the gap.

TSMC's Arizona project is another important piece. The company has been ramping US production for major American customers, including Nvidia, Apple, AMD, and Qualcomm, while its second Arizona fab is expected to move into high-volume production later this decade. Nvidia has already celebrated US-made Blackwell wafer production at TSMC Arizona, but advanced AI chips still depend on a global chain of fabrication, packaging, testing, and equipment suppliers.

That is why tariffs are a blunt instrument. They can make domestic investment more attractive, but they cannot instantly replace the advanced packaging capacity, supplier density, and manufacturing experience concentrated in Taiwan and South Korea. Samsung, TSMC, Intel Foundry, Micron, and others are all part of the answer, yet each operates on timelines measured in years, not quarters.

The administration has already used a narrower tariff approach this year, with a 25% Section 232 duty applying to certain advanced semiconductor articles under specific conditions. The current discussion points to something potentially wider, or at least more strategically timed. Greer's message was not that tariffs are off the table. It was that Washington wants them to support domestic output instead of simply raising costs before alternatives are ready.

For AI builders, the practical takeaway is straightforward. Hardware sourcing is becoming a policy risk as much as a supply risk. Teams planning data center buildouts should model tariff scenarios alongside power availability, GPU supply, and financing costs. Edge AI companies should watch component origin more closely, because a small device can carry a large exposure if its highest-value parts are caught by a new duty structure.

The next phase of the chip fight will not be only about whether the US can design the best AI processors. It will be about where those chips are made, how quickly domestic capacity can scale, and who pays the premium while the supply chain is rebuilt.

Source: SF

• USTR Makes Findings and Proposes Action in 60 Section 301 Investigations Relating to Failures to Take Action on Trade in Forced Labor Goods
• [VCCI] The US officially initiates a Section 301 Investigation into Vietnam on Intellectual property
• General Secretary and President To Lam's working visit to three ASEAN countries: Demonstrating the stature of proactive and constructive diplomacy.
• Viet Nam accelerates innovation drive as skilled workforce becomes key to science and technology growth
• The US is increasing controls on AI chip exports