Chip bans? LOL! Chinese web giant Tencent says it has enough GPUs for future AI model training Chinese web giant Tencent says it has enough high-end GPUs to train new AI models for years, in part because it’s found more efficient ways to do so. Speaking on the company’s Q1 2025 earnings call, company president Martin Lau said Tencent has “a pretty strong stockpile of chips that we acquired previously”. The company will …
As title says, US is pushing China to innovate by putting various restrictions on them.
There is a decent chance that with all the engineering resources they have over there, they may figure out a better way of doing AI/chips/whatever else they are being restricted on.
In whatever thing they innovate, they will presumably have first mover advantage.
“Over the past few months, we start to move off the concept or the belief of the American tech companies, which they call the scaling law, which required continuous expansion of the training cluster,”
This is the Mythical Man-Month all over again. It is as if people do not learn from history?
We saw this earlier with the USSR. Americans used bigger computers to solve mathematical problems, Soviet mathematicians used smarter methods. And it showed that whole classes of problems were intractable when approached with more computing power. Meanwhile, smarter mathematics resulted not only in solved problems, but also much more insight into even weirder mathematical problems. Even today, the effect on mathematics from the old soviet countries is visible.
When looking at 'AI' in the US, I see a lot of 'bigger guns' being applied, but disappointingly little insight coming out. And we already know that the newest foundational models build with the latest data centers would need more data than is available in the world.
Worse, increasingly the data out there is already generated by LLMs instead of humans. And AI generated data is the fast-food of training data, if not the hard-liquor.
In the end, the restrictions do what everyone expected, they force China to develop their own hardware as well as better mathematics and more efficient software.
When all you have is a hammer, everything looks like a nail.
The US is pretty awful for the bigger is better fallacy, you only have to look at the awful efficiency of their "muscle cars" where many struggled to get to 50HP a litre naturally aspirated while European cars were pushing 100HP per litre NA in far lighter engines and chassis for an insight into the mindset.
Faster, lighter, smarter seems to have flown over the heads of the US AI companies too.
"Brute force is simple. Optimizing is hard and cost a lot of money. USA has never been much into Optimizing."
Exactly all of this !
I'm coming from a parallel super-computing background when those were very expensive per GFLOPS (the 90s) and have seen plenty of *very* retarded code that was really wasting precious CPU cycles due to ignorance and laziness to optimize. I even saw code running slower on capable CRAYs than on your average SPARC station ! This was the very debut of parallel systems, but still.
I think we're in the same situation here with AI: the infancy. Running current LLM costs PIOPS while in the future we will realize we can achieve the very same on medium range systems.
"50HP a litre naturally aspirated while European cars were pushing 100HP per litre NA"
Is that imperial or metric horse power? ;)
My brain was trying to grok 50HP a litre as a fuel efficiency measure but it's engine power (HP) per cylinder capacity (L.) In other contexts the more likely unit would be Watts per litre I would imagine. Never knew there was a metric horse power (75kg - 1m - 1sec) A SI horse?
Subtilty not really a hallmark of US both historically and now. Particularly now.
I think the relevant measure for cars is Km/liter (Miles/Gallon for the SI challenged).
A heavier and larger engine can often boast better fuel efficiency. For instance, I assume an Airbus A380 will boast a superior horsepower per gallon (cannot find the relevant specs). But a heavier engine does not make it a superior choice for a car.
It is well known that Americans suffer from small car claustrophobia (or so I was told in a TV ad). They do prefer larger, more heavier car with fancy gadgets. Which means that even the most fuel efficient engine will result in a bad km/l (MPG) efficiency.
"Is that imperial or metric horse power? ;)"
Technically, there's no such thing as an American horse. They were all imported. Immigrants, if you like. And 99/9% of the rest of the world is metric, so an "imperial immigrant horse" as a unit of power is the anachronism :-)
In the case of AI, a very easy "smarter method" is to spend some effort selecting your training data rather than simply diverting the sewers of social media down your throat and hoping for the best. Since the big US players appear to be doing the latter (and spending time and money batting off the resulting legal challenges), I am quite willing to believe that someone else can do more with less.
A fact that American authorities don't seen to be capable of comprehending. I wonder how much was genuinely invented in that country by their own people.
I remember being fascinated by reading about the creation of the first blue LED - at a time when world plus dog said it was impossible... it was a Chinese guy that did it.
When my ex was picking up her Masters (unrelated qualification) from Warwick Uni about 15 years ago, about 80 - 90% of those picking up their awards in Science and Maths had Chinese names.
This doesn't mean they were from China, but given that Warwick has a huge amount of overseas students, I would hazard a guess many were.
I've heard it said there are three "classes" of Chinese university students.
The very best, who get government grants to go to their choice of the worlds top universities.
The best, who go to the many very good Chinese universities.
The rest, who didn't get into a Chinese university and go to the worlds other technical based universities, sometimes with government help.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
