Gyeongju, October 30
Donald Trump and Xi Jinping met in Gyeongju, South Korea, on the last day of October 2025, on the fringe of the APEC summit. The session ran about ninety minutes. Both men left calling it a success, and for once the paperwork matched the mood. Washington cut the fentanyl-related tariff on Chinese goods from 20 percent to 10 percent, pulling the average rate on Chinese imports down from roughly 57 percent to about 47 percent. Beijing suspended, for one year, the rare-earth export licensing regime it had announced on October 9. It promised to buy 12 million metric tons of American soybeans before the end of 2025 and 25 million tons a year through 2028, and it paused its antitrust investigation into Nvidia. Port fees on each other's ships went on hold. The US suspended the so-called 50 percent rule that had stretched entity-list restrictions to majority-owned subsidiaries.
None of that reaches the thing both governments actually care about. Semiconductors and the models that run on them sat outside the truce. The export controls stayed. So did the entity list, the investment screening, and the licensing regimes that decide which Chinese firm gets to train on which accelerator. Trump and Xi shook hands over soybeans and left the compute question exactly where they found it.
The Choke Point Is Compute
US controls on advanced chips to China began in October 2022 and tightened in October 2023, when the Commerce Department replaced a performance threshold with one based on total processing speed and interconnect bandwidth. That closed the loophole that had let Nvidia sell the A800 and H800. December 2024 added high-bandwidth memory to the list, which mattered because HBM, not logic, is the harder bottleneck. A rule published in January 2025 tried to cap compute exports by country tier; it was rescinded in May after allies objected and Nvidia lobbied.
What survived was licensing. Nvidia's H20, a cut-down Hopper part built for the Chinese market, required an export licence from April 2025, and the company took a $4.5 billion charge on unsellable inventory. A revenue-share arrangement announced in August let it resume sales with 15 percent of the proceeds going to the US government. Beijing's response came in September, when regulators told major Chinese technology firms to stop buying H20 for state-funded projects. Nvidia's data-centre revenue from China, once around a fifth of the total, fell to near zero.
What China Built Anyway
Huawei's Ascend 910C and its successor parts are the core of the domestic answer. Cambricon, Biren, and Moore Threads fill in behind. SMIC fabricates at its N+2 node, nominally 7-nanometre, without access to ASML's extreme ultraviolet lithography, which means multi-patterning, lower yields, and higher cost per wafer. CXMT is scaling HBM. None of this matches a Blackwell rack. All of it is enough to train and serve models that are competitive on many tasks.
The proof arrived in January 2025, when DeepSeek released R1, a reasoning model trained for a fraction of what US labs spent. Alibaba's Qwen, Moonshot's Kimi, Zhipu's GLM, and Baidu's Ernie followed, most of them with open weights. Cheap inference is a distribution strategy. A model that anyone can download in Lagos or Jakarta becomes the default, and defaults are hard to dislodge.
Power is China's quiet advantage. Its grid added more generation in a single year than several European countries run in total, and the East Data West Computing programme sites data centres in Guizhou, Gansu, and Ningxia where electricity is cheap. US labs face interconnection queues that stretch past 2030 in some regions.
Rare Earths Run the Other Way
China refines around 90 percent of the world's rare-earth elements and dominates gallium, germanium, antimony, and graphite processing. The October 2025 rules required licences for any product containing more than 0.1 percent Chinese-origin rare earths by value, which reaches samarium-cobalt magnets in F-35 actuators, neodymium magnets in MRI machines and wind turbines, and the motors in electric vehicles. That is a chokepoint pointed at the US defence and energy industries rather than at its data centres. Stockpiles in Japan and Australia buy time, not independence.
Selling the Stack Abroad
Washington's AI Action Plan, published in July 2025, folded export promotion into national policy. Commerce now packages American chips, models, and cloud capacity into single deals, and the American AI Exports Program funds the financing. The template deals with G42 in the United Arab Emirates and Humain in Saudi Arabia put US accelerators and US model weights into Gulf data centres under US legal terms.
Beijing answers with governance rather than hardware. The Global AI Governance Initiative, and a proposed World AI Cooperation Organization headquartered in Shanghai, aim to set norms for countries that will never buy an Nvidia part. Huawei Cloud sells capacity across Africa and Southeast Asia. Open-weight releases do the rest of the work.
What the Truce Leaves Standing
The physical supply chain is unchanged. TSMC makes nearly all leading-edge logic, and its CoWoS advanced packaging is booked out for years. SK Hynix, Samsung, and Micron supply HBM. ASML in Veldhoven and Tokyo Electron in Japan sit upstream of everything, and both governments in The Hague and Tokyo have aligned with US controls, though not without grumbling about lost sales.
Taiwan remains the single point of failure. Neither leader mentioned it in the readouts from Gyeongju, and neither has an alternative to it before 2028 at the earliest, when TSMC's Arizona fabs reach volume.
What to Watch
Three things will show whether the truce means anything. First, whether Nvidia gets licences for anything above the H20 tier, and whether Beijing lets its firms buy. Second, China's 15th Five-Year Plan, covering 2026 to 2030, which will set domestic compute targets and probably name semiconductor self-sufficiency as a binding goal. Third, HBM. Memory bandwidth constrains training more than logic does, and China's progress there is the metric that matters most.
The truce expires in November 2026. It covers tariffs, soybeans, and port fees. It does not cover the race, and neither man pretended otherwise.
