Hacker News Reader

Comment by neom | Hacker News Reader

Not many folks talking about this: https://www.tomshardware.com/tech-industry/artificial-intell...

The WH has said it hasn't approved any sales, but it's not clear China is buying, and it seem they are making good progress on their huawei ascend chips. If China is basiclly at parity on the full stack (silicon, framework, training, model), and it starts open weighting frontier models at $0.xx/M tokens, then yeah, moat issues all around one would imagine? Not surprised to see Anthropic complaining like this: https://www.anthropic.com/news/detecting-and-preventing-dist... - but I don't know how you go back from it at this point?

View on HN

Not surprising, Nvidia's margin was just a huge incentive for companies/countries to develop their own solutions. You don't have to be 100% as good if you're 80% cheaper. It's unsurprising that this is being driven by Chinese companies/labs who often have a lot less funding than the US, and the big tech companies (Google, Microsoft, Amazon) who will benefit the most from having their own compute.

I've never believed in Nvidia's moat, and it seems OpenAI's moat (research) has gone and surprisingly is no longer a priority for them.

Feels a bit crazy saying this but I can imagine a weird future where we have some outlawed Chinese tokens situation under some national security guise. No clue how that would work but nothing surprises me anymore.

Shit is about to get alot more cyberpunk than we're used to, thats for sure

It seems like it’s really only China that’s pursuing the route of doing more with smaller/cheaper models, too, which also has a lot of potential to give the whole bubble a good shake.

To me it seems like the most obvious thing to do. More efficient models both make up for whatever you lost by using cheaper hardware and let you do more with the hardware you have than the competition can. By comparison the ever-growing-model strategy is a dead end.

I think you might be underestimating the use of small models in proprietary systems. The progress from China is very visible because it's very open, but the big tech companies are doing this too for cost savings.

China doesn't need to buy it. They can continue their policy and look good.

They've already found a better route. Buy it elsewhere e.g. in Singapore. Train their models there using Nvidia hardware.

Ship the result and fine tune back in China.

So "China" is and has always been buying it. No difference. The politics can keep raging.

  it seem they are making good progress on their huawei ascend chips

This is interesting to me. I thought that the reason for deepseek delay was because of the insistence ( by the politicians) to use huawei chip[0]. But that was last year August.

Anything changes in between?

[0]: https://www.reuters.com/world/china/deepseeks-launch-new-ai-...

GLM-5 was trained entirely^ on Huawei Ascend chips. https://z.ai/blog/glm-5 / https://tech.yahoo.com/ai/articles/chinas-ai-startup-zhipu-r...

(^edit, I don't know for certain entirely is accurate - edit again, found a chinese source saying their image model is end to end ascend, or at least, domestic: https://zhuanlan.zhihu.com/p/1994775762516080044 & https://www.guancha.cn/economy/2026_02_12_806895.shtml)

And evdn this information might be not very reliable because both US and China government wouldnt be happy about fact that some models might happen to be trained on some "shadow datacenter" full of Nvidia GPUs.

Reply on news.ycombinator.com