[Der_Einzige]

They always say "no, it'll be GOOGLE!" and try to bring up TPUsv6 (you still can't even play with ironwood/TPUv7 yet) as though they're competitive with blackwell let alone Reuben, as though anyone except google internal engineers get real value from them, and as though GCP doesn't have well over 1 million+ haswell/blackwell GPUs.

These same people unironically believe that Gemini was trained with zero GPUs for any part of the process (including all experiments), and it was all done on TPUs. They cite this as evidence that CUDA/Nvidia is dead.

It's stupid, it's wrong, but it's what they claim.

I'm going to be a lot wealthier than the Nvidia bears.

And if you don't believe this, just take one look at TPU pricing (remember, Trillium is TPUv6) and tell me this is competitive with Nvidia/Oracle/any neocloud with a straight face.

https://cloud.google.com/tpu/pricing?hl=en

[danielmarkbruce]

I'd argue the Nvidia moat is the most difficult moat to understand unless one is really in the weeds using GPUs. "all you need to do is design a similar chip" seems to fit in peoples heads nicely when they are even a little ways removed from the details.