Two Leaders, One Long Table
Donald Trump and Xi Jinping sat down in Busan, South Korea, on 30 October 2025, on the margins of the APEC (Asia-Pacific Economic Cooperation) leaders' meeting. It was their first face-to-face encounter since Trump returned to the White House in January. Cameras caught the handshake. The two men talked for roughly 100 minutes, and tariffs dominated the public readouts. Artificial intelligence dominated everything else.
The deal they struck was narrow and temporary. Washington trimmed the fentanyl-related tariff on Chinese imports from 20 percent to 10 percent. Beijing suspended the rare earth export licensing regime it had announced on 9 October, promised to buy 12 million metric tons of American soybeans in 2025 and 25 million tons a year through 2028, and resumed cooperation on fentanyl precursors. The truce runs for one year. Neither side surrendered a single rule governing advanced computing.
That omission matters more than the tariff cuts. Semiconductors and the machines that train large language models (LLMs) sit outside the bargain. The Bureau of Industry and Security (BIS), a branch of the US Department of Commerce, still decides what Nvidia may sell to Chinese buyers, and the Entity List still names hundreds of Chinese firms barred from receiving American technology without a license.
The Chip as Chokepoint
Compute is the binding constraint on frontier AI. Training a model such as GPT-4 or DeepSeek-V3 requires thousands of graphics processing units (GPUs) wired into clusters, each chip running for weeks at high utilization. Nvidia holds roughly 80 to 90 percent of the market for the accelerators that do this work. Control the flow of Nvidia's silicon, and you control the speed at which Chinese labs can scale.
Washington has pulled that lever repeatedly. In October 2022, BIS banned exports of advanced chips and the tools used to make them. October 2023 lowered the performance thresholds and closed loopholes around the A800 and H800, cut-down parts Nvidia had designed for China. December 2024 added high-bandwidth memory (HBM) to the control list. July 2025 reversed course on the H20, a deliberately weakened chip, allowing sales again in exchange for a 15 percent share of the revenue.
CoCom's Ghost
There is precedent for this kind of economic containment. From 1949 to 1994, the Coordinating Committee for Multilateral Export Controls (CoCom) kept strategic goods out of the Soviet bloc, with the United States, Japan, and Western Europe enforcing a common list. Today's controls are unilateral in form but multilateral in effect, because the choke points sit in allied hands. ASML of the Netherlands is the only supplier of extreme ultraviolet (EUV) lithography machines. Japan's Tokyo Electron and Nikon supply the deposition, coating, and metrology gear. Without them, nobody builds a leading-edge fab.
What Beijing Built Without EUV
China cannot buy EUV, so it has improvised. Semiconductor Manufacturing International Corporation (SMIC) reached a 7-nanometer node using deep ultraviolet (DUV) immersion lithography with multi-patterning, a slower and costlier route that yields fewer usable dies per wafer. Huawei's Ascend 910C accelerator is fabricated on that line and stacked with domestic HBM. Neither part matches Nvidia's top-end hardware on throughput per watt.
Software has compensated. DeepSeek, a Hangzhou lab spun out of the quant fund High-Flyer, trained its V3 model on about 2,048 Nvidia H800 chips for a reported $5.6 million in rental cost. The architecture leans on mixture-of-experts routing, which activates only a fraction of parameters per token, plus heavy use of lower-precision arithmetic. The result narrowed the gap with American frontier models without requiring a matching cluster.
Rare Earths as the Counter-Lever
Beijing holds a chokepoint of its own. China refines roughly 90 percent of the world's rare earth elements and dominates processing of gallium, germanium, antimony, and graphite. In December 2024 it banned exports of gallium, germanium, and antimony to the United States. The October 2025 rules went further, requiring licenses for any product containing more than 0.1 percent Chinese-origin rare earth content, wherever that product is made. That extraterritorial reach mirrors the US Foreign Direct Product Rule.
Rare earths matter for AI because they go into the permanent magnets inside data center cooling pumps, disk drive actuators, and the motors of the robots that factories are starting to install. A one-year suspension of those controls buys time, not certainty.
Compute as the Unit of Power
Both governments now measure national strength in floating point operations (FLOPs) and gigawatts. The United States has committed to Stargate, a joint venture of OpenAI, Oracle, and SoftBank carrying a headline figure of $500 billion in planned data center investment. China's answer is a state-directed build-out of hyperscale campuses in Guizhou, Inner Mongolia, and Ningxia, where electricity is cheap and coal is plentiful. Power, not silicon, is becoming the harder limit to move.
Scaling laws gave the transformer architecture its leverage: more parameters and more tokens have reliably produced better models. If that relationship holds, the country with the most usable compute wins the next few years. If it breaks, as some researchers argue it must, then algorithmic efficiency decides, and export controls lose much of their bite.
An Entangled Contest
The two economies remain tied together in ways neither leader can cut cleanly. Nvidia still counts China as a major market. TSMC in Taiwan fabricates the chips both sides depend on. Apple assembles iPhones in Zhengzhou. Trump and Xi shook hands in Busan and postponed a fight they cannot end, because the race they are running has no finish line in sight.
