Some thoughts I’ve been having recently:

Could DeepSeek and Zhipu considering their own chips imply that China’s HBM self-sufficiency has already reached a certain level, or is at least on track to get there?
Also, is the EDA moat already starting to collapse? OpenAI reportedly completed Jalapeño in just nine months using its own in-house AI chip design tools. Is it really plausible that Chinese labs don’t have something similar? In fact, there are claims that the chip design tools developed by Chinese labs may be even more practical in real-world use than OpenAI’s.
The fact that OpenAI is specifically launching GPT-5.6 inference on Cerebras makes me think they may see more demand for ultra-fast inference than people realize. Jensen has described markets like Cerebras as niche, but OpenAI’s move seems to suggest otherwise.
Broadcom is a great company, but I think Hock Tan may be downplaying MediaTek too aggressively. He says MediaTek’s TPU has been delayed, but wasn’t Broadcom’s own TPU also delayed in some areas?
Amazon’s 20% GPU price increase is not because of memory. It is because of rising demand.
I also have doubts about whether Broadcom can continue to defend its high margins. MediaTek’s TPU margin is around 35%, while Broadcom’s is around 65%. Broadcom may be able to maintain high margins for the first couple of years as it wins new customers like Anthropic, but won’t margins eventually trend lower as competition intensifies?
Morgan Stanley updated its TPU / AMD GPU shipment estimates, but personally, I do not understand how MS is arriving at this scale for AMD GPU shipments.
Morgan Stanley now models AMD shipping 2.4 million GPUs this year and 2.7 million next year.