推文

@GenAI_is_real · 2026-10-11 23:37

As discussed with some of my friend, If the Reflection acquisition talks are real, I don't think Jensen is being a bagholder at all. The strategic logic is actually pretty elegant. NVIDIA's best world is not one where a handful of frontier labs buy GPUs, everyone else buys API tokens from those labs, and the labs eventually optimize their own hardware stack. In that world, NVIDIA ends up with a small number of extremely powerful customers who have every incentive to reduce their dependence on GPUs. And we can already see the direction of travel: Google has TPU, AWS has Trainium, Microsoft has Maia, Meta has MTIA. At enough inference scale, custom silicon starts becoming economically rational. Open models change the structure of the market. Instead of: NVIDIA → frontier lab → API → everyone else you get thousands of companies fine-tuning, serving, and specializing their own models. A hyperscaler may justify spending years optimizing an ASIC. A smaller company usually just wants something that works, has good software support, and can run the model today. That is a much healthier customer base for NVIDIA. So supporting open-weight labs like Reflection is not just about selling more GPUs this quarter. It helps prevent the model layer from collapsing into a few vertically integrated API providers. The more companies believe they can own their model stack, the more independent compute buyers exist. And the more independent compute buyers exist, the harder it is for the ecosystem to route around NVIDIA. Seen that way, open models are not a side project for NVIDIA. They are part of the moat.

曝光 688 · 评论 0 · 点赞 6 · 书签 2 · 曝光/时 774.7744181153043