Good timing to revisit this. Since April 8, three data points update the demand-side picture.
Kimi K2.6 open-sourced yesterday on Huawei Ascend infrastructure, benchmarking 58.6% on SWE-Bench Pro while running 300 parallel agents and completing a 13-hour autonomous financial exchange refactor. Domestic chips now sit at ~41% of AI servers in China (from Chinese media sources tracking the Ascend deployment wave).
The pricing signal is the most interesting update. Moonshot raised K2.6 API prices 58% the same day it open-sourced the weights. Yang Zhilin's argument is that agents don't buy tokens, they buy compute runs. Autonomous workloads are orders of magnitude more compute-intensive per task than single-turn inference.
If that holds, the demand-side calculus shifts. The relevant metric isn't total tokens processed but simultaneous autonomous workflows supported. DeepSeek V4's native Ascend integration (expected this week) will be the first real test of whether Chinese infrastructure can meet that curve.
This pair of articles on Chinese compute have been really excellent
What is the top green sliver on the pie chart representing?
how do you tally this with the claim from the State Council Information Office that daily token processing has reached 140 trillion (https://hellochinatech.com/p/china-token-economy-140-trillion?utm_campaign=post-expanded-share&utm_medium=web&triedRedirect=true) which if my maths is right would imply 7.3 million H100e?
https://substack.com/@aquavis/note/c-153749544
How does this compare to the US?
I admire your assumption that China would spend 33% more compute on science than on domestic video production. Do we do the same in the USA?
Good timing to revisit this. Since April 8, three data points update the demand-side picture.
Kimi K2.6 open-sourced yesterday on Huawei Ascend infrastructure, benchmarking 58.6% on SWE-Bench Pro while running 300 parallel agents and completing a 13-hour autonomous financial exchange refactor. Domestic chips now sit at ~41% of AI servers in China (from Chinese media sources tracking the Ascend deployment wave).
The pricing signal is the most interesting update. Moonshot raised K2.6 API prices 58% the same day it open-sourced the weights. Yang Zhilin's argument is that agents don't buy tokens, they buy compute runs. Autonomous workloads are orders of magnitude more compute-intensive per task than single-turn inference.
If that holds, the demand-side calculus shifts. The relevant metric isn't total tokens processed but simultaneous autonomous workflows supported. DeepSeek V4's native Ascend integration (expected this week) will be the first real test of whether Chinese infrastructure can meet that curve.
More on K2.6, Ascend, and V4 timing in today's China AI Dispatch (chinaaidispatch.substack.com/p/300-agents-and-a-price-hike)