Discussion about this post

User's avatar
Crosscurrents's avatar

Xu's framing on risk is candid in a way worth taking seriously. If concern about AI danger only shows up once a lab can feel it firsthand, the compute gap may also push back the point where Chinese developers treat safety as their own problem. That timing could matter a lot for anyone hoping for a US-China floor on dangerous capabilities. It also sits awkwardly next to his 2029 estimate for the hardware crunch, since a few more years of accelerating first is a long window. Thanks for translating the full transcript, the Guancha version would have been easy to miss outside Chinese-language circles.

Dmitry Mintz's avatar

The rule published on 15 January 2026 makes case-by-case review available below two numbers: total processing performance under 21,000 and total DRAM bandwidth under 6,500 GB/s. Both have to be met, not one of them. BIS names the H200 as an example of hardware falling below them, together with its equivalents and less advanced chips.

One side of that comparison is counted downstream of a licence decision rather than a purchase decision. Ascend's constraint, in his account, is supply against domestic demand. What reaches the other side goes through that route, case by case.

https://dmitrymintz.substack.com/p/the-rule-was-written-for-the-chip

Ready for more?