Now Playing
DeepSeek's Chip Move, H200 Rationing & TSMC's Slowing Growth
Audio is available on Spreaker — see link below.
DeepSeek is designing its own AI inference chip while Beijing quietly rations H200 imports — and TSMC's revenue growth just missed expectations by a wide margin. Today's briefing unpacks what these signals mean for the future of AI compute.
Audio is available on Spreaker — see link below.
DeepSeek is building its own AI inference chip. That's the clearest signal yet that China's most watched AI lab is betting its future not just on model efficiency, but on semiconductor independence.
Running parallel to DeepSeek's long-term bet is a short-term policy concession that tells its own story. Beijing is set to approve limited H200 chip imports for Alibaba, ByteDance, and DeepSeek, with the total capped below two hundred thousand units.
On the supply chain enforcement side, US Commerce Secretary Lutnick pressed ASML executives directly on suspicions that a complete EUV lithography system reached China through underground channels. ASML denied the allegation with tracking documentation.
Meanwhile, three Chinese startups — Suanmiao, Transtreams, and Dongfang Suanxin — completed tape-outs in April and June of this year on three-dimensional stacked AI chips built using domestic processes. They've raised hundreds of millions of yuan to pursue this architecture as a direct workaround to US export bans.
Away from China's self-sufficiency push, Samsung's System LSI division is developing Gaia, a four-nanometre AI PC chip with an integrated neural processing unit, targeting mass production in twenty twenty-seven. It's Samsung's first serious return to PC silicon in roughly ten years.
The thread connecting all of this is straightforward. Nvidia's training dominance is not seriously threatened in the near term.
Chapter summary auto-generated from the verified script. Listen to the full episode for the complete content.