DeepSeek orders 160,000 Huawei Ascend 950DT chips for inference-only deployment while Broadcom's AI semiconductor revenue outpaces Nvidia at 221% growth — signalling a structural shift in how AI hardware is procured. Today's briefing covers rare-earth export halts, AMD's 192GB local inference spec, and the solidifying split between Western and Chinese chip ecosystems.
Audio is available on Spreaker — see link below.
DeepSeek has ordered a hundred and sixty thousand Huawei Ascend 950DT chips for a new inference cluster in Ulanqab, and that single procurement decision tells you almost everything about where China's AI infrastructure actually stands right now. Here's the important distinction.
Pull back to the wider market and the competitive picture shifts further. Broadcom's AI semiconductor revenue grew two hundred and twenty-one percent year on year in fiscal third quarter, reaching sixteen point seven billion dollars.
On the local inference side, AMD's Ryzen AI Max PRO 400, codenamed Gorgon Halo, now officially supports a hundred and ninety-two gigabytes of unified LPDDR5X memory, with a hundred and sixty gigabytes allocated to the GPU. That's enough to run a three-hundred-billion-parameter model at four-bit quantization without touching cloud infrastructure.
Microsoft's Project Zenith adds another layer. It's a preconfigured Windows setup targeting thirty-billion-parameter local inference, and the first device is the Lenovo ThinkCentre X Ultra at three thousand six hundred and ninety-nine dollars, shipping November twenty twenty-six.
The other development worth watching is the rare-earth situation. Chinese suppliers have been refusing shipments to US customers since early August, tied to Beijing's sanctions response involving the Rare Biosphere Alliance.
Step back and the through-line is clear. Export controls that began in October twenty twenty-two were designed to slow Chinese AI progress.
Chapter summary auto-generated from the verified script. Listen to the full episode for the complete content.