DeepSeek chatter centers on chips, compute caps and a 100,000-card training cluster

teortaxesTex · x · 2026-07-24

A long repost argues that Wenfeng’s investor-call comments about DeepSeek should be read skeptically, especially on compute and chips.

The speaker says the claim that he will not do chips “if possible” is hard to square with the fact that chip work has reportedly been underway for years.
They also question the claim that the government gave him no money, citing a reported $150M investment.
On compute, the post says $3B tops would translate to about 120K H200s at $25K each, but export caps make that hard to acquire in one shot.
The attached job listing for an AI Computing Cluster Performance and Reliability Engineer explicitly references a 100,000-card cluster, emphasizing training-oriented stability, fault recovery, observability, and ongoing hardware operations.
The takeaway is that the real story is less about rhetoric and more about the operational reality of building and keeping an enormous training cluster alive.

More from Companies & People