business 6 min read

US Chip Curbs Hit Samsung and China’s AI Build-Out

Expanded US semiconductor export controls are forcing Samsung and TSMC to halt shipments to Chinese AI firms, triggering a pivot toward indigenous chip production that could reshape the global supply chain within three years.

  • Semiconductors
  • China Tech
  • AI Hardware
  • US Trade Policy

The November 19 turning point

The US Bureau of Industry and Security’s new regulations took effect on November 19 with consequences that were immediate and visible. Samsung Electronics stopped shipments to the Kunlun AI division of Baidu and to ByteDance, two of China’s largest artificial-intelligence consumers of foreign silicon. TSMC was pulled into the same pattern. What made this particular round different from earlier restrictions was not the destination list but the breadth of coverage — consumer smartphone chips from Xiaomi, OPPO, and vivo were added to the permit regime.

The shift from data-center GPU blocks to mobile-application processor controls signals a deliberate widening of the chokepoint. In earlier iterations, Washington left room for volume through older nodes or end-user certifications. The new framework removed that margin for devices shipping into China with chips containing even minimal US-origin design tooling or fabrication inputs.

Samsung’s exposure is significant. Industry estimates place China revenue at roughly 20 percent of the company’s total sales, a figure concentrated in memory and foundry segments where Chinese purchasing power has grown fastest over the past five years. Halted shipments do not disappear from the balance sheet overnight, but the order book for the next two quarters already reflects the disruption.

TSMC faces the same calculus, except the exposure is broader because its client roster in China spans both cloud providers and device makers. A 2017–2023 period saw Chinese clients climb from single-digit to mid-single-digit percentages of TSMC’s foundry revenue. That trajectory now stalls unless permits clear — and the permitting process, by all accounts, is slower than anyone inside the fabs anticipated when the rule was announced.

Who moves first out of the hole

Alibaba made the boldest public commitment. Starting in the first quarter of 2025, it said its largest language models would be trained exclusively on its own TingTing processor, a chip developed through its semiconductor subsidiary. The announcement was not purely symbolic. Tongyi Lab engineers have spent the better part of four years iterating on the architecture, and the third-generation variant already matches roughly 80 percent of the performance envelope of an entry-level NVIDIA A100 on standard transformer benchmarks. It does not yet match the A100 on memory bandwidth or multi-GPU scaling, but the gap closes every six months under accelerated funding.

Alibaba’s Chief Executive, Wu Yongming, projected that domestic chip production would begin ramping substantially from 2028 onward. That timeline matters because it coincides with when another generation of hyperscalers enters maturity — the data centers currently being built around Hangzhou, Beijing, and Shanghai will need replacement capacity by then regardless of what Washington decides next.

Tencent offered a less dramatic but equally pointed assessment. Its chief strategy officer stated publicly that all Chinese AI companies would eventually rely on domestic chips, and notably cited token throughput numbers — the actual cost-per-inference metric that determines whether a model remains economically viable — as exceeding one-year-old forecasts. The implication is not optimism about policy reversal; it is a practical acknowledgment that the alternative was already too expensive.

Moonshot AI, a relatively young player funded partly by Chinese state capital, cancelled a $150 million contract to lease NVIDIA Blackwell GPUs housed in an Indonesian data center. The cancellation came after the Commerce Department expanded its permit requirements to cover cloud-computing compute services, effectively closing the offshore-lease loophole that many Chinese firms had relied on since the original rounds of restrictions took hold in late 2022.

The Indonesia channel closure is worth watching closely. It was never meant to be permanent; it was always a compliance workaround until regulators adjusted the text. Other routes — Singapore, UAE, Turkey — face similar scrutiny now. Firms that assumed they could maintain access through regional intermediaries were operating on information that predates the latest rulemaking.

The supply-chain geometry

The immediate effect of the new regulations is a reordering of purchasing decisions across three layers: foundry capacity, equipment, and end-use software stacks.

On the foundry side, Samsung and TSMC are absorbing the shock on both sides of the Pacific. Chinese fab operators, primarily SMIC and Hua Hong, are quietly expanding capacity to serve the very customers losing foreign supply. Reports indicate SMIC is moving 7-nanometer-class wafers through N+2 processes at volumes that would have been invisible twelve months ago. The yield numbers remain undisclosed and likely variable, but the direction is unambiguous.

Equipment suppliers occupy a more complicated position. ASML, Lam Research, Applied Materials, and Tokyo Electron benefit from the demand surge tied to domestic Chinese fab expansion, even though they cannot sell their most advanced tools into China directly. The secondary-market refurbishment pipeline — older EUV and DUV machines maintained and resold through third countries — is growing faster than official trade statistics capture. Sanctions enforcement has historically struggled with documentation gaps that exist precisely because the equipment moves through multiple corporate shells before reaching final assembly lines in Suzhou and Beijing.

Software stacks are where the slowest but deepest adaptation occurs. Chinese cloud providers have rebuilt their ML frameworks to prioritize domestic accelerators. Alibaba’s TingTing support is now embedded in its primary training platform. Tencent has aligned its internal model-building pipelines to run cleanly on Huawei’s Ascend processors. This migration required months of engineering work and introduced latency into existing product roadmaps, but the alternative — relying on permits that may never arrive — carried unacceptable risk.

What happens next

The most concrete short-term indicator is the permit-clearance rate at BIS. If the agency processes applications within thirty to sixty days, some volume returns to Samsung and TSMC’s Chinese customer lists, though likely at reduced quantities. If processing drags beyond ninety days, the structural shift accelerates faster than anyone outside Beijing’s planning committees expected.

China’s domestic ecosystem has enough momentum to absorb a prolonged restriction without collapse. SMIC’s output will not match TSMC’s at the leading edge, but the mid-range segment — 7 nm to 14 nm equivalents — is where most commercial AI inference runs today. That is the contested space, and the contest is still forming.

For Samsung and TSMC, the question is duration. Both companies are large enough to survive a two-year revenue dip in any single market. They are not large enough to ignore a structural decline that persists beyond that window, particularly if American competitors such as Intel Foundry Services gain ground while Asian fabs lose share in China simultaneously.

NVIDIA’s China exposure deserves separate tracking. Even with restricted exports, prior-generation chips continue to move through indirect channels, and Chinese firms pay premiums of 30 to 50 percent above official list prices for used or refurbished units. That premium compresses margins for sellers but signals sustained demand regardless of policy texture. NVIDIA shares have absorbed the narrative headwinds, but the underlying revenue question remains open-ended.

The clearest outcome line emerges from the timing of Alibaba’s 2028 ramp forecast. If domestic production hits scale by that date, Chinese hyperscalers will have exited dependency on foreign supply for routine workloads. Edge cases — cutting-edge research models, extreme-scale reinforcement learning — will still seek the best available silicon wherever it originates. But those represent a small fraction of aggregate compute spend.

For investors, the signal is not directional but compositional. Restriction winners cluster around US design-tool vendors, domestic Chinese fab-equipment proxies, and companies with diversified geographic revenue bases. Losers concentrate among pure-playfoundry operators with outsized Chinese customer weight and cloud providers whose training pipelines depend on single-source accelerators. The transition period between those outcomes is measured in quarters, not years.