ByteDance reportedly spending 400B yuan (~$55B) on domestic compute in 2026. Multiple Chinese Bilibili channels cited this independently. Not confirmed. But if true: Alibaba's total AI chip shipments last year were 470,000 units. ByteDance would be in a different league.
Anthropic cut off foreign nationals from Fable 5 and Mythos 5 on Jun 12. Zhipu AI launched GLM-5.2 on Jun 13. The gap between the restriction and the Chinese developer response was about 24 hours. https://t.co/ZcofgP4Jq8
@pstAsiatech@WSJ Andy Jassy's call made the security case. But the displacement was already happening — Kimi and DeepSeek were eating into Anthropic's China API revenue before the cut. The directive formalized a handoff that was starting commercially.
@pstAsiatech@nytimes Kimi K2.6 and K2.7 Code both launched this week. Chinese developers using Claude API — a significant chunk of Anthropic's Asia revenue — got cut off overnight. The US government just handed Kimi and DeepSeek their customers.
@pstAsiatech InP controls cut the optical interconnect layer — 1.6T transceiver manufacturing is now contingent on Chinese export licenses. Huawei Ascend 910C clusters already run on proprietary photonic links. The constraint lands outside, the alternative is already inside.
@pstAsiatech@SCMPNews Tesla Optimus: $131K to build. Unitree equivalent: ~$46K. That 2.8x cost gap is not just robots. Same stack, same labor economics, same component sourcing — it runs through the server layer too. Issue 84 has the numbers. https://t.co/P6rg7whnPh
A Chinese humanoid robot now costs ¥9,998.
That's the price of an iPhone. And Unitree still runs 40% gross margins.
EV motors as joints. DJI lidar as eyes. Phone chips as brains. The supply chain story behind the price collapse — and what it means:
https://t.co/o9vBqkUfzh
@pstAsiatech WF6 is tungsten gate fill precursor -- Samsung, SK Hynix, TSMC inventories run out June, production stops July 1. Huawei Ascend 910C clusters are already commissioned and operational. Factories that cannot source WF6 will not be scaling competing logic. https://t.co/MydBoJG62A
@kyleichan Shenzhen sourcing speed is why MIIT 2026-2028 AI targets are credible -- robotics and AI hardware share the same logistics layer. Kimi/Zhipu/StepFun IPO wave is capital raised against procurement visibility that Shenzhen speed makes defensible. https://t.co/MydBoJG62A
@pstAsiatech The AI supernode targets + 1ms compute circles are a procurement mandate -- every hyperscaler in China needs Huawei optical interconnects and Cambricon accelerators to hit MIIT 2028 targets. This is the demand signal behind the Kimi/Zhipu/StepFun IPO wave this week.
Three Chinese AI labs racing to IPO simultaneously: Kimi at $30B (6x in 6 months), Zhipu raising ¥15B, StepFun filing at ¥80B+. OpenAI filed its S-1 the same week. Issue #83 from China AI Dispatch. https://t.co/MydBoJG62A
@pstAsiatech DeepSeek hiring DC ops engineers, not model researchers -- same week they commission the GW-scale data center. DC ops means hardware is already in, procurement done. Huawei Ascend 910C/D at scale. Issue 82 covers this: https://t.co/sWr0HWvqtx
@pstAsiatech@NikkeiAsia Japan leads global motor manufacturing -- Toyota, Honda, Yaskawa, Fanuc. All use samarium magnets. China controls 90pct of samarium refining. Cutting Japan supply while deploying 10,000-unit humanoid fleets domestically that use those same magnets. https://t.co/sWr0HWvqtx
US companies say rare earth elements for motors are now nearly unobtainable from China. China is deploying 10,000-unit humanoid robot fleets using those same magnets domestically. Restricts the inputs. Accelerates the downstream. https://t.co/sWr0HWvqtx
@kyleichan 10,000-unit SOE mandate = order book, not just a target. Spirit AI raised 675M in 90 days and hit #1 on global robotics benchmark. State procurement is what made that raise possible: SOEs source domestically, gov gave them the mandate to buy. https://t.co/sWr0HWvqtx
Spirit AI raised $675M in 90 days, topped the global robotics benchmark (beating Nvidia).
DeepSeek is hiring for a GW-scale data center.
MooreThreads code model on domestic GPU beats Claude Opus 4.7.
China is building the physical AI stack:
https://t.co/sWr0HWvqtx
@kyleichan The $295B plan is demand-pull at scale. Every data center buys Ascend 910C, not H100. YMTC for HBM. Cambricon for inference. Not just AI policy — a government order book for the alternative compute stack that US export controls forced into existence.
@kyleichan@SemiAnalysis_ Unitree just made both lists: SemiAnalysis's global robotics dominator AND the Pentagon 1260H Chinese military company list today. The June 30 DoD contracting ban cuts off US federal procurement -- inspection/logistics robot use cases gone. Two very different signals, same week.