Two armies, one bottleneck

The military-AI competition between the United States and China is usually framed as a contest of algorithms and ambition. In practice the decisive variable is more prosaic: semiconductors. Whoever can design, manufacture and deploy the most capable AI chips at scale sets the ceiling on everyone else's autonomous weapons.

The capability race

Both sides are moving fast. The US military aimed to deploy thousands of autonomous drones by the end of 2025 — an explicit effort to counter China's numerical advantage in unmanned systems (National Defense Magazine). China, for its part, has integrated AI into a growing fleet of autonomous platforms, including 'robotic wolves' shown during an amphibious-landing exercise, and is reportedly procuring AI systems at speed — some requests for proposal specifying acquisition timelines of just three to six months, with DeepSeek-related military procurement notices appearing through 2025.

The chip chokepoint

China's constraint is hardware. Analysts assess that its inability to produce advanced semiconductors at scale is the central bottleneck preventing it from overtaking the US in AI — though China is pushing hard for self-sufficiency and may meet domestic inference-chip needs around 2028, while still trailing on high-end training chips (CSIS; FDD).

This is why export controls are the real battleground. A contested decision in late 2025 to let Nvidia export its H200 chips to China sparked alarm: Business Standard reported that PLA-linked universities — part of China's 'Seven Sons of National Defence' — were seeking access to exactly those chips. Critics argued that licensing the H200 would help modernise Chinese capabilities across autonomous drones, electronic warfare, precision strike and command-and-control.

The dilemma

The US faces a genuine bind. Tight export controls slow China's military AI but cost American chipmakers a vast market and push China toward indigenous production. Loose controls keep US firms dominant commercially but risk handing the PLA the compute it needs. There is no clean answer — only a moving line that each new chip generation forces policymakers to redraw. The race for military AI supremacy, in the end, is being run in fabs as much as in labs.