"DeepSeek says it took just two months and a little over $5 million to build one of its models, using degraded Nvidia chips that are available in China. I don’t believe that. Lennart Heim, who knows AI computing, pointed out that the cost numbers are almost certainly untrue, and the timing of this release, and the story of DeepSeek, is likely a strategic public relations campaign by the Chinese intelligence services.

More likely, the company had a large number of chips on which to train, export controls really only started biting in October of 2023, and the big new U.S. chip clusters haven’t really come online. Plus there are likely capabilities that U.S. firms are keeping private. A few open source Chinese models, though impressive, don’t mean the U.S. lead in AI is gone, or that the U.S. strategy of denying top chips to China has failed.

That said, no one denies that DeepSeek’s technical accomplishments are impressive, and the accomplishment suggest that there’s a lot about AI that we still don’t know. And a Chinese firm releasing a better cheaper product they claim was produced for less than most top AI project managers make in a year does make me feel like America is quite… Soviet."

https://www.thebignewsletter.com/p/monopoly-round-up-china-embarrasses

#AI #GenerativeAI #China #USA #BigTech #DeepSeek #DeepSeekR1 #Nvidia