The capacity is built. The chips are not moving.

For most of the past decade, the constraint on Nvidia's China business has been Washington. That is no longer the binding one. On July 14, a senior US Commerce Department export-controls official told Congress that shipments of Nvidia's H200 AI accelerator to China have been very few, even though the department has approved roughly $10 billion worth of licenses covering about ten Chinese buyers 1. A day later, coverage of the same testimony described the volume actually delivered as trivial, and flagged a separate fight over whether a Blackwell-generation loophole is letting more capable silicon slip through anyway 7. Nvidia's own numbers confirm the freeze: the company has booked no China data-center compute revenue from H200 sales this year.

The approvals are real. In May, the US cleared H200 sales to ten named Chinese customers, widely read at the time as the breakthrough Jensen Huang had been chasing since Washington reopened the door to sub-frontier AI chips for China 2. What has not happened is China letting its own companies spend the money. As early as January, Beijing told domestic tech firms to hold off on new H200 purchases while the central government worked out terms that would not undercut its homegrown chip industry 3. Reporting since has described the restriction tightening further, with purchases effectively limited to special-circumstances cases such as university research labs, and customs directives aimed at slowing import clearance for the chip entirely.

Two governments, two vetoes, one idle order book

The result is a rare formation: both governments nominally support the trade, and it still is not happening. Washington's approval was the headline story for most of 2026. Beijing's refusal is the one investors have been slower to price. The mechanism is straightforward — Chinese regulators are using licensing friction and procurement pressure to keep AI-infrastructure spending flowing to Huawei, SMIC and Cambricon rather than to a foreign supplier, even one the White House has cleared, and even at the cost of leaving Chinese cloud and internet firms with less compute than they say they need.

That dynamic is exactly why the order book grew in the first place. Chinese technology companies had placed orders for more than 2 million H200 units for 2026 while Nvidia held roughly 700,000 in finished inventory, prompting Nvidia to ask TSMC to open additional production lines starting in the second quarter 4. That gap between what was ordered and what existed is the crux of the current overhang: TSMC is now fabricating and packaging silicon against demand that Beijing has not authorized its own customers to draw down. Every wafer that becomes a shipped, revenue-generating H200 rather than idle finished-goods inventory depends on a Chinese approval that has not arrived seven months into the year.

The beneficiaries are named, and they are ramping

Beijing's blockade is not passive protectionism; it is redirecting a specific, quantifiable amount of AI-infrastructure spending toward three companies. Huawei is targeting AI chip revenue near $12 billion in 2026, roughly 60 percent above 2025, built on its 950PR accelerator and an upgraded 950DT variant due late in the year, in a domestic AI-chip market Beijing is pushing toward $67 billion by 2030 5. Much of that Huawei silicon is fabricated by SMIC, China's largest foundry, which is itself capacity-constrained. Cambricon, the other major domestic name, is targeting roughly 500,000 AI accelerators for 2026, including up to 300,000 units of its Siyuan 590 and 690 chips, though outside estimates flag low yields and limited high-bandwidth-memory supply as real ceilings on that plan 6. Both companies now sit on China's official secure-and-reliable procurement list; Nvidia does not.

What this means for Taiwan's foundries and investors

For TSMC, the exposure is less about lost revenue today than about capital already committed against demand that may not clear. Advanced packaging lines built for H200 volume are a sunk cost if Beijing's licensing stance hardens rather than eases, and TSMC's China-adjacent order book now carries political risk that sits with a customer's customer's government rather than with TSMC or even Washington. For Nvidia, the near-term read is a data-center China business that has effectively been zero for most of 2026 despite regulatory clearance, which should temper any assumption that approved licenses translate into near-term shipments or revenue. For SMIC and Cambricon, Beijing's stance functions as a demand guarantee, but one still bounded by the yield and memory-supply constraints both face.

The broader signal for Asian markets is that decoupling in AI chips is no longer purely a Washington-authored story. Beijing has demonstrated it will hold its own companies back from a legally available foreign chip to protect a still-immature domestic supply chain, accepting a real compute shortfall at Chinese cloud firms as the price. Investors modeling TSMC's China-linked capacity, Nvidia's China revenue ramp, or SMIC and Cambricon's growth trajectories should treat Beijing's approval calendar, not Washington's, as the binding constraint through the rest of 2026.