The big problem with huewei's current chips is the scope of it. To train deepseek v.5 they are chaining 160k together. And the programming they are using instead of cuda is more time intensive (if one node goes out they have to pull every node and check by hand) so having extra gpu if even as fallback is advantagous

The author hasn't signed the social smart contract — the cast may be restricted on farcaster.xyz, so the Farcaster app isn't offered below.

Open this cast in

Quilcast Quorum Vibeshift Opencaster Neynar Terminus Firefly HomieHouse Fatline Linkcast