Do not have any empirical evidence, but reality is that China's semiconductor capabilities are not at par with Taiwan yet and the US is able to influence Nvidia's sales to China as well as access to other vendors (TSMC) and technologies, giving the West an unfair advantage.

Just like Chinese EVs and Chinese renewables eventually beat the West, I have no doubt that China can probably eventually pull ahead, but I think it is probably accurate to say that China is currently still behind (how far is hard to say) because they have a slight technology handicap imposed by the US.

Your comment is responding to an issue that is different from what GP said. GP was talking about Chinese open source particularly, i.e. their open source models, which AFAIK have consistently been keeping up with (albeit a few steps behind) the closed source OpenAI and Anthropic models.

Hardware capacity is a separate issue entirely.

    > have consistently been keeping up with (albeit a few steps behind) 

I mean, this sentence is self contradictory, no?

    > Hardware capacity is a separate issue entirely.

It seems like hardware capabilities are at the very heart of both training and inference which is why Nvidia, TSMC are hitting record income and capitalization. Feels like divorcing hardware from the equation is discounting a big part of winning this race.

> I mean, this sentence is self contradictory, no?

By benchmarks, the Chinese models are ahead of where the proprietary US models were ... something like 6 or 12 months ago. And all the benchmarks are a bit fuzzy anyway on whether a small gap is trivial or significant. The Chinese aren't having any problems keeping up on model quality. The gap isn't going to lead to any difference that matters unless the US pulls a rabbit out of it's hat.

Plus dollar-for-performance they might be leading in practice, it is hard to compete with self hosted.

You can keep up even if you’re behind. If someone is running a race and you’re constantly two seconds behind their time, you are steps behind but keeping up.

> I mean, this sentence is self contradictory, no?

As others have pointed out, no, not at all. For specifics, see the chart from this link posted by another commenter: https://hai.stanford.edu/news/inside-the-ai-index-12-takeawa... . If anything, Chinese models are closing the gap with US models, not falling behind.

> Feels like divorcing hardware from the equation is discounting a big part of winning this race.

Depends on what race you're talking about. When it comes to "who has the most powerful models", I'd argue it's actually not really that significant - China obviously has the power to train good models.

Probably a better way to phrase it would be keeping “pace“. Yes, they are still behind but by about the same amount as they always have, they aren’t drifting further behind. Like two marathon runners, one a a mile ahead, but both maintaining the same mile time.

China would probably be very confused that you're asserting they're not on par with Taiwan.

In semiconductors.

Otherwise, no one would need to buy from Nvidia or contract with TSMC.

This article might catch you up: https://en.wikipedia.org/wiki/Political_status_of_Taiwan

I'm aware; I'm Taiwanese :)

> There are no switching costs for users to move to a new model.

This depends on how many proprietary APIs are in the way of the model itself.

Or whether the model still works the way you want. For example, a lot of people were pretty unhappy with Claude 4.7 and preferred the way Claude 4.6 worked. If you're relying on a service, then you're stuck with whatever changes the provider decides to make. And the provider is chasing a demographic that's profitable, if you happen to fall out of that demographic then tough luck.

But if you run your own models then you're not subject to anybody's whims anymore. You have full control of how your software works and what it does.

That’s absolutely a reason to just use a web chat interface. Zero cost switching between providers when one isn’t performing. Second, your data are spread between providers which reduces surveillance effectiveness.

You can't really do any serious work with a chat interface though. Also, you're still wasting time figuring out if your providers do what you expect. If you run a local model then you know it's going to work the way you set it up in perpetuity.

This is absolutely false as a recent study from Stanford clearly states https://hai.stanford.edu/news/inside-the-ai-index-12-takeawa...

> This is simply not true?

Yes, this is purely delusional.