DeepSeek withholds latest AI model from Nvidia, AMD | Hacker News Reader

DeepSeek withholds latest AI model from Nvidia, AMD

reuters.com

This is getting played in (parts of) the US press like it's DeepSeek trying to hide the fact that it trained its model on "illegal" hardware but in fact it's really likely about trying to give Huawei an early advantage.

the article is framing this as DeepSeek hiding sanctions violations but it's probably simpler. Give Huawei early access before Nvidia can optimize for the model and you help build a domestic chip ecosystem.

I don't get this.

Models have become a commodity, hardware not so much; why would Nvidia care? It's the models who are competing primarily, not hardware manufacturers.

Also, models are software, so easy to switch out. AI hardware otoh has a approx. 3-5 year useful life, I don't see how this affects hardware already running, or getting built.

HW manufacturers work on the software side to make sure that the popular models run well with their systems. From drivers to patches submitted to vLLM etc.

I have an NVIDIA Spark machine, and NVIDIA has a whole team building software, docker images, etc. to make these things relatively easy to use for LLM research etc and running local models because it just makes sense to pull people in like that to keep the market captured using their HW by keeping the software up to date.

Even more so for AMD who is behind on the software side by years and failed to do the above in the early days and lost out for it.

The DeepSeek 3 series models were quite popular, and quite capable. The new ones will likely be as well. Many people will want to host and run them. By making them run initially better on Huawei hardware than on NVIDIA it will encourage API hosting providers to buy Huawei hardware, and to get things like vLLM and llama.cpp working better on them.

Moderately interesting from a geopolitical perspective but beyond that not so much.

DS4 could be an interesting release though. Hoping they come out with a coding plan and increase context dramatically. The current 8K doesn’t really do it

I don't quite get it if that is bad thing for Nvidia and AMD. How would them be able to optimize their GPUs with a model?

The model running initially (maybe) better on Huawei than NVIDIA hardware means that hosting providers will have some motivation to buy more Huawei hardware instead and also that software developers will learn to work with Huawei HW.

I personally hope we see a Huawei or similar competitor to the Strix Halo and NVIDIA Spark lineup for "prosumer" LLM work.

Via drivers I’d assume

Reply on news.ycombinator.com