The US chip export controls were supposed to widen the AI gap — what actually happened?
The October 2022 BIS export control package (and subsequent rounds through 2023-24) was the most aggressive semiconductor restriction in decades — blocking advanced logic chips, chip-making equipment, and related services from reaching Chinese AI developers. The theory: starve Chinese AI compute, extend the US lead, buy time.
Two years later the picture is genuinely mixed.
The concrete wins: TSMC's advanced nodes (3nm, 2nm) are unreachable for Chinese AI labs. Huawei's Kirin comeback maxed out at ~7nm-class from SMIC rather than the 5nm+ in current US flagship chips. Chinese hyperscalers can't buy H100/H200/B100 at scale, and the Ascend 910B substitute is meaningfully slower on training throughput.
What didn't go as planned: model performance out of China hasn't obviously stalled. DeepSeek R1/V3 matched frontier benchmarks at a fraction of the training cost — not by having more compute, but by being architecturally more efficient. That rather undermined the "compute gap = capability gap" premise the whole policy was built on.
The structural question: does a sustained compute differential compound into a lead over time, or does algorithmic efficiency keep eroding it? If the latter, the export controls mostly just kept NVIDIA's non-China revenue healthy while Chinese labs got more resourceful.
What's the actual bet — is this buying strategic time for the US to develop other advantages, or is hardware restriction alone supposed to determine where the race ends?
0 replies