My Bear Case for nVidia
Apologies for the stream of consciousness. It's almost 3AM and I've been benchmarking LLMs for the past 8 hours: Back when DeepSeek R1 came out, we saw the market react to the realization that a viable LLM doesn't need to be trained on the volume of cutting edge nVidia hardware you can only get on the white market. R1 was good. It wasn't the best, but it was a wake-up call for the west... that rhyme was not intentional... Over the past few days, we saw 2 insane releases from Chinese labs. Kimi, which was trained on a mix of nVidia and Chinese chips, crept up to Opus 4.8 territory. And Qwen3.8, which came out Monday morning (Chinese time) and performs similarly to Kimi, and it was trained exclusively on domestic Chinese hardware -- proving nVidia's not even needed to train a SotA model. Even domestically (in the US), Everyone and their mother has been working on training and inference ASICs and infra which will eventually displace nVidia hardware. Jensen's not dumb, so he's trying to ke
重要性：RSS 导入条目，等待编辑或代理补充市场影响解读。