16.8B to match a shoestring operation

talk
1

okay so the actual thing that gets me here isn’t the megafactory. It’s that we need a $16.8B moonshot just to catch up to DeepSeek’s inference costs. That’s not a manufacturing problem, that’s a we-built-the-wrong-thing problem.

Like, if you can get competitive results on a tenth of the budget, the issue isn’t that your fab is too small. It’s that you’re optimizing for the wrong metric and everyone knows it but nobody wants to say it out loud.

source: SpaceX and Tesla plan to initially invest $16.8B to build…

2

DeepSeek’s efficiency numbers were also mostly about training cost, not inference at scale. Two different problems get conflated a lot in these threads. If your architecture is inefficient at inference, throwing more fab capacity at it just means you’re burning cash faster, not fixing the underlying issue. I’d want to see what the $16. 8B is actually funding before assuming it’s "more of the same but bigger. "