https://www.cnbc.com/2026/05/21/nvidia-jensen-huang-china-ai-chip-market-huawei.html
>>108871111+$0.50
I ran his statement through a CEO to English translator>Please let us sell our products in China.
>>108871111i mean the majority of the chinese llm's have already being converted to use huaweis chip
Is that because the Huawei chips are significantly better, or because Nvidia can't legally sell its chips in China?
>>108871371Huawei cracked how to reversee engineer Nvidias chips but modified the design just enough to avoid legal issues.
would be interesting to hear how much it would cost to train a frontier model using huawei chips vs latest nvidia tsmc stuffthe energy consumption must be like 5 times higher
>>108871371Because basically no matter what, even if Trump lifted all restrictions on Nvidia sales to China, Huawei has CCP backing for production and supply so they're not going to overcome that.
>>108872073Good thing China built all that solar and has all that unused power capacityI've been reading up on these chinese chips lately, seems they're 60-70% the performance of nvidia which is really not bad. While they're behind in the fundamental technology (uv litho etc) seems to me that doing things the hard way + sheer scale and effective infinite capital because state backing for any well-founded investment is a killer advantage over time. I feel iike the Chinese might overtake by about 2035.
>>108872150If you look at the nvidia showcases they have those graphs with almost exponential growth in efficiency for training. No one wants to be at the efficiency from 5y ago
>>108871111>Went to China to sell NVIDIA chips>Still enforcing pre-existing RTX 4090/5090 export banWhat is his plan?
>>108871371China's entire MO is working with foreign companies until they can steal enough of their secrets to create a domestic alternative, then kicking the foreign company they stole from out. It's admirable in a sense. If (((Western))) governments operated with such a degree of ruthless national self-interest then things would be better for their people, but Jews and their shabbos goyim have no interest in that goal.
>>108872061that's what the japs did to your parents but you won't be learning any of that with every waking moment of your life watching animated jappy propaganda
>>108872192>in efficiencydumbass
>>108872291Based Chinese Communist Party knows western capital class will sell everything just to be able to make some profit in China.Blame capitalism.A capitalist will even sell you the rope you will use to hang him.
>>108872428what
>>108872455there are no exponential efficiency gains. they make the silicon large and use all to all intra-rack networking to get those numbers. power, price and size all explode from one model to the next.
>>108871371>>108872192Cheap and domestic is good enough for China.IDK why Americans and westerners in general have this obsession the only thing that could ever be worth using is the absolute cutting edge. A million easily acquirable domestic accelerators beats out whatever rack you managed to assemble from smuggled GPUs
>>108872917Its extremely important for the frontier modelstraining models on huge datacenters with latest nvidia systems takes months, half a year perhapsWith worse systems that are slower and eat up much more energy the costs skyrocket and you end up with a smaller model thats worse than competition that releases later.
>>108872991Are you not allowed to use hardware for practical applications that don't require the top end Nvidia chips?Or is AI solely a race for some undermined goal and everything not frontier is worthless?Do you also look down on Google's own TPUs?
>>108873031you can put billions into making a shitty model no one wants to use because theres several more advanced models out theregoogles tpus are made on modern tsmc nodes
>>108873058There are practical models out there already. Even antrophic is using older Nvidia hardware because they have a model that's useful and it's going to be cheaper to deploy if it's not on cutting edge hardware.Google has ancient TPUs and Nvidia hardware (literal Pascel GPUs) still in production because they have a use case Frontier models have little practical purpose if they are going to be available for 6 months before being discarded. Literally gives zero time for businesses and users to find a use case that fits the capabilities of the model.At some point your going to have to deploy cheap hardware to cover and sustain the practical use cases.
>>108871415Please tell me there's an audio version of this.
>>108872061What legal issues? China isn't going to prosecute a Chinese company for ripping off a Taiwanese company's IP.
>>108872440>Based Chinese Communist Party knows western capital class will sell everything just to be able to make some profit in China.>Blame capitalism.kek, they are literally just playing the same game as everyone else, people just complain because they aren't the ones benefiting from it
>>108871111Huawei engineer here. Our chips are good. Buy our chips.
>>108871371China blocked nvidia imports because they don't want to get rugpulled when Trump changes his mind and blocks nvidia exports after China grows dependent on them. Better to just develop a domestic solution.
>>108871111Congrats, US was basically using antibiotics frivolously and the chinese are now an anti-biotics resistant strain. America did it in during the red scare with Qian xuesen and now they're doing it again.
Novidia is stagnant garbage, they're evolving backwards, not to mention their sucking up to the pedo shlomo class
>>108872440>Blame capitalism.There was no long term monetary profit it, just more land ownership. They don't do it to get richer, they do it because they can't stand uppity wagies.
>>108871111SoWHAT WAS THE POINT OF THE TARIFFS AND DATACENTERS TO "BEAT CHINA"?
>>108875775>the slave tries to justify why he wears his chains by claiming the chains "own the wagies"
>>108872917>IDK why Americans and westerners in general have this obsession the only thing that could ever be worth using is the absolute cutting edgeWe don't, all of the stuff we have is what we call "Goyslop". The cutting edge costs extra and is forced from us using our taxes so the military can grow larger (The military never uses it effectiely)
>>108873099>Even antrophic is using older Nvidia hardware because they have a model that's useful and it's going to be cheaper to deploy if it's not on cutting edge hardware.Inference is completely different than training.xAi made a mistake building Colossus 1 and ended up with heterogenous cluster of H100/H200/GB200 GPUs. This is a nightmare to train huge models on because each card has different specs, features, and hardware requirements. During gradient synchronization, a heterogeneous cluster would bottleneck on the slowest GPU (H100) so the faster GPUs would end up idling. So xAi ended up at 11% MFU.Later xAi pivoted to renting Colossus 1 to Anthropic for inference.
>>108876819Can they really do a full weight update for every mini-batch at that scale?I thought they would have to start training parallel models and merging. In which case a couple of the model instances being slower doesn't really matter.
>>108871111
