Artificial Intelligence thread

[nugroho]

It´s not difficult to understand why the Chinese govt is not thrilled about the prospect of Chinese AI companies importing more US GPUs from Nvidia. Why help cement US hegemony in chip sales?

View attachment 167844


The US compute advantage is real and massive. Yet, while compute is important, it is also not everything. Meta had a huge budget but didn't deliver much. DeepSeek and others show that high talent density punches way above its weight. I would not bet against Chinese AI companies, but we have to acknowledge that they will have compute disadvantages compared to their US peers for years to come. It's just silly to pretend otherwise.

And here's broken down by country. Note: this includes Chinese companies buying NV GPUs.

View attachment 167846

" EPOCH " ?? something fishy

 

[HighGround]

" EPOCH " ?? something fishy

The information is not necessarily wrong, but it does not adequately reflect the reality.

A lot of "Chinese compute" is not located physically in China.

 

[tphuang]

Qwen's open-source multimodal model is quite impressive, but its Qwen3 MAX is already significantly behind, and thinking still hasn't been fixed. I haven't used it for a long time.

different AI labs serve different purposes. Qwen is the cornerstone of all the edge AI stuff.

 

[iewgnem]

China‘s total computing capacity is around 788 EFLOPS, while US get around 10 million H100 which is roughly equivalent to 10,000 EFLOPS, so 10x is reasonable.
But the total revenue of AI model is around 1/10 of the cost of buying these chips, so 10x also means huge deficit.

Please, Log in or Register to view URLs content!

Please, Log in or Register to view URLs content!

Do you realize, considering Chinese AI companies are clearly fully matching US offerings, that what you're actually implying is American AI is so bad they need to spend 10x more on compute to match China.

 

[lockedemosthenes1]

Do you realize, considering Chinese AI companies are clearly fully matching US offerings, that what you're actually implying is American AI is so bad they need to spend 10x more on compute to match China.

not exactly, because perhaps training AI model doesn't need 10x computing power, and a large part of the H100/200 chips are for model inference.

 

[iewgnem]

not exactly, because perhaps training AI model doesn't need 10x computing power, and a large part of the H100/200 chips are for model inference.

You can check the token rate on Chinese LLM youself you know

 

[Michael90]

Of the ones you mentioned only Shein completely moved its HQ outside.

Which makes sense, since SHEIN doesn’t sell in China . It’s was made for selling outside China . So it makes sense to base their HQ outside China

 

[Franklin]

Can you really spy on others through NVIDIA chips and servers or is that just a myth.

 

[Eventine]

Please, Log in or Register to view URLs content!

Bloomberg (original article under pay well) quotes Justin Lin, head of Qwen research at Alibaba, as saying that China's R&D efforts are constrained by compute:

Justin Lin, head of Alibaba Group Holding Ltd.’s Qwen series of open-source large language models, has put the odds of any Chinese company leapfrogging leaders such as OpenAI or Anthropic through fundamental breakthroughs at less than 20% over the next three to five years, according to a Bloomberg report.

Lin’s assessment reflects a broader sentiment within China’s AI ecosystem, with similar views reportedly shared by peers at Tencent Holdings Ltd. and Zhipu AI, one of the country’s most prominent LLM developers.

“A massive amount of OpenAI’s compute is dedicated to next-generation research, whereas we are stretched thin — just meeting delivery demands consumes most of our resources,” Lin said at the event, which was co-organised by Zhipu AI and Tsinghua University, as reported by Bloomberg.

The imbalance, he suggested, raises a fundamental question about innovation itself. “It’s an age-old question: does innovation happen in the hands of the rich, or the poor?” Lin added.

Note that this does not necessarily mean China doesn't have enough chips, and Western journalists may as usual be reaching to over-interpret Lin's comments as reflecting the "success" of Western sanctions.

In reality, it may just reflect the difference in data center investment: the US has driven a gigantic investment bubble in data center capacity (Open AI alone has invested $1.5 trillion and single-handedly locked up 40% of global DRAM capacity). The US has essentially bet the farm on hyper scaling towards AGI with trillions of dollars of investment. Will China follow (with all the risks of the bubble popping as most AI companies are not profitable) or will it be more cautious via a fast follower approach? I'm betting on the latter until we see clear evidence that AGI is at hand.

 

[iewgnem]

Please, Log in or Register to view URLs content!

Bloomberg (original article under pay well) quotes Justin Lin, head of Qwen research at Alibaba, as saying that China's R&D efforts are constrained by compute:



Note that this does not necessarily mean China doesn't have enough chips, and Western journalists may as usual be reaching to over-interpret Lin's comments as reflecting the "success" of Western sanctions.

In reality, it may just reflect the difference in data center investment: the US has driven a gigantic investment bubble in data center capacity (Open AI alone has invested $1.5 trillion and single-handedly locked up 40% of global DRAM capacity). The US has essentially bet the farm on hyper scaling towards AGI with trillions of dollars of investment. Will China follow (with all the risks of the bubble popping as most AI companies are not profitable) or will it be more cautious via a fast follower approach? I'm betting on the latter until we see clear evidence that AGI is at hand.

Sama buying up all those RAM wafers is for next gen research I see.

20% chance leapfrog with supposedely a fraction of compute is a flex if you look at it another way.