Dylan Patel: prediction

Transcript context

…Like they can release a 1e27 model by 2026? And a 1e28 model in the works. They totally could do this just with foreign chip supply. It’s just a question of centralization. Then the question is, do you have as much innovation and compute efficiency wins when you centralize? Or do Anthropic, OpenAI, xAI, and Google develop things, and secrets shift a bit between each other, resulting in a better long-term outcome versus nationalization in the US? China could absolutely have it in 2026-27 if they desire to, just from foreign chips. Domestic chips are the other question. You have 600,000 of the Ascend 910B, which is roughly 400 teraflops or so. If they put them all in one cluster, they could have a bigger model than any of the labs next year. I have no clue where all the Ascend 910Bs are going, but there are rumors about them being divvied up between the majors, Alibaba, ByteDance, Baidu, etc. Next year, you have more than a million. It's possible they actually have 1e30 before the US because data center isn't as big an issue. A 10 gigawatt data center… I don't think anyone is even trying to build that today in the US, even out to 2027-28. They're focusing on linking many data centers together. There's a possibility that come 2028-2029, China can have more flops delivered to a single model, even once the centralization question is solved. That's clearly not happening today for either party. I'd bet if AI is as important as you and I believe, they will centralize sooner than the West does. So there is a possibility. How many more wafers could they make and how many of those wafers could be dedicated to the 910B? I assume there's other things they want to do with these semiconductors.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.