China's AI models, especially by Kimi will get a lot larger. I anticipate 5T first and then close 10T as the SuperNodes continue to get more powerful. Moonshot get its compute from Alibaba Cloud.
At the moment it seems like HBM, although its mostly everything chip related, probably 2035+ if we say everything goes perfectly (for china to serve everyone).What is the constraint here for scaling up? SMIC capacity, or the HBM? What is the realistic timeframe of gaining enough inference capacity to serve the world in a competitive fashion?
That's why there is talk of banning Chinese model.
It isn’t like China forced them to go all in on AI…
That is a loaded question. In Kimi's case, it wants to be capital efficient, so if demand suddenly 2x or 3x, it's going to be compute constrained until it can sign new contracts.What is the constraint here for scaling up? SMIC capacity, or the HBM? What is the realistic timeframe of gaining enough inference capacity to serve the world in a competitive fashion?
What’s there to stop people from downloading open source models? Are they gonna wall off the internet?That's why there is talk of banning Chinese model.
They can download them, but they still need a server to host them, especially now that models are often over terabytes in size, making self-hosting impossible for individual users.What’s there to stop people from downloading open source models? Are they gonna wall off the internet?
VPN time baby.They can download them, but they still need a server to host them, especially now that models are often over terabytes in size, making self-hosting impossible for individual users.