Comment by PeterHolzwarth
2 days ago
Article grabbed at random that provides some more context (tho could use more):
https://www.cyberkendra.com/2026/07/deepseek-pauses-fundrais...
"The Hangzhou AI lab has told prospective investors in its second fundraising round that it is suspending the deal, people familiar with the matter told Bloomberg on Saturday, days after remarks attributed to founder Liang Wenfeng about US-China AI competition circulated widely online."
And:
"Tencent's technology outlet published a 118-item version covering AGI strategy, chip supply, pricing, and retention. In it, Liang reportedly framed China's disadvantage as an arithmetic problem rather than a talent one: "The biggest gap between us and the US is in resources.""
"The specifics were unusually candid. Liang is said to have told investors he needed 200,000 Huawei 950 chips to train a frontier model but received 16,000, adding that "Huawei's problem is still insufficient capacity" and expecting the crunch to last at least three years. He also floated narrowing the gap with US labs to three to six months using a fraction of their computing."
I wonder if it would be viable for the Chinese to pursue a huge buildout of less sophisticated logic fabs. My experience with GPGPU is that it seems to intensely skew towards memory bandwidth, and has a much lower compute intensity than graphics (by which I mean rasterization and shaders).
This seems to be holding true for AI as well. I'm sure if you have an excess of compute and a dearth of bandwidth, you can trade the former for the latter, but still, this is a fundamenta property.
A lot of talk has been said about how companies are doing 'financial tricks' to extend the useful life of GPUs by showing lower depreciation - but what if these are not tricks at all - new GPUs don't really have that much more bandwidth, and while they might be clever in some other ways, they are limited in how much they can improve fundamentals.
This has been reflected in how memory vendors' stock price has exploded, but NVIDIA stayed stagnant.
Since the Chinese are far closer to the US in building SOTA memory chips, it's possible that their disadvantages are far overstated.
The real question is whether it's easier to improve the software side instead. There are likely a lot more optimizations possible in terms of model architecture, and if there is a compute bottleneck, then it's going to put a lot of pressure on Chinese labs to address the problem using more efficient designs.
And one cool trick is that if one develops step function more efficient training, one can still “open source” the model without revealing the training techniques.
1 reply →
And that's why OpenAI bought all the future ram contracts
this bodes well for continuing to refine smaller models and open sourcing them.
There's a delusion that what America's AI companies are doing is "best"; the chinese should realize that the forefront is bloated and there's likely hundreds of speed ups viable. Pushing open weights will continue to grind down the bloat.
I was gonna say, this just puts more pressure to deliver ground breaking research with limited resources. And if history teaches us anything it’s that scarcity produces ingenuity.
http://www.incompleteideas.net/IncIdeas/BitterLesson.html
> One thing that should be learned from the bitter lesson is the great power of general purpose methods, of methods that continue to scale with increased computation even as the available computation becomes very great. The two methods that seem to scale arbitrarily in this way are search and learning.
6 replies →
> "There's a delusion that what America's AI companies are doing is "best""
Not sure if the word "delusion" is the correct word here? It has not been proven in either direction. We can all see lots of possible issues with it, but it is also possible that it could be what is needed to unlock key capabilities.
We can see that the Chinese models have been getting better, but OpenAI is out there supporting 10 million active users with their frontier models, and now we know that Deepseek can't even get what they need to properly train models.
They can't get hardware because the US has put restrictions on how much can be sold to China. There is not a technical or know-how limitation, but political. Deepseek could otherwise write some checks to NVidia for what they want.
Thanks to the import restrictions, I expect Chinese GPU hardware to be competitive within a few years.