Comment by dhx
3 days ago
I take it from [1] (transcript of recent DeepSeek CEO discussion with investors) that DeepSeek would disagree on the immediate catastrophic impact to the likes of OpenAI or Anthropic. The reason is even though technology parity mostly exists, only OpenAI, Anthropic et al have the inference capacity to gain market share and generate revenue. Chinese vendors don't have the chips needed to scale up inference and gain market share, and the DeepSeek CEO doesn't think this would happen in optimistic circumstances in the next 3 years, but thinks it might be possible in 5 years.
In summary, regardless of country of origin, availability of inference capacity is the moat protecting the likes of OpenAI and Anthropic, not technology superiority.
[1] https://www.fredgao.com/p/deepseeks-liang-wenfeng-breaks-his
That merely pushes the valuation onto the hardware makers, not the companies that have the temporary preferential access to their hardware.
That makes them at best temporary middlemen.
It only justifies their long term valuations if they can leverage that temporary monopoly for technological superiority (they can't) or lasting market share (they can't).
Chinese models prove there's no technical advantage, and the software side is heavily commoditized so there's not much advantages to market share either.
The question mark in my mind over the technological superiority is whether the additional volume of data they see due to capturing the top of the market allows them to do recursive self-improvement in a way nobody else can match, before any of the other labs can figure it out. That's the only runaway outcome I can see.
If you have exponentially increasing use of your harness, then it's true that every day you capture exponentially more data, but it's also true that every day exponentially more data will slip through the cracks of your would-be monopoly and that data arrives at your competitors via various channels (competitor harnesses, subsidized reselling, etc)
The very exponential that you are relying on to give you runaway improvement is also giving exponentially increasing data to your competitors. All else being equal your competitors stay a step behind but you never develop a monopoly either. That's the best case for Anthropic/OpenAI. In reality, training data is just one variable, exponentials don't last forever, and your competitors will get better at capturing a bigger slice of training data.
If user data would become such a key ingredient (which it might, i actually remember noam shazeer talking about the importance of user data), i think chinese labs can still get it from china, as keep in mind it ahs a billion people behind the great firewall banned from using us llms. And btw broadly for any gap like this, you really gotta consider that if its becoming a bottleneck, chinese labs will find a way to buy it from one of the labs unless theres strict regulation at the government level
But is that data good? That's the question. As in, is my usage at work:
a) indicative of problems that aren't already out there in the wild? (no) b) are the responses I'm getting so good and novel that the model can improve itself? (no)
It's the garbage in garbage out idea, just scaled up. If the model gave a bad answer, and I didn't catch it, and you now train on that I/O pair (my perhaps crappy prompt, the bad output), then you're not going to improve anything.
1 reply →
Yes, RSI seems to be the new AI industry McGuffin of 2026, just as agentic capability has become table stakes and scaremongering has become a punchline.
The Chinese models are adopting licensing quite rapidly and Xi will soon enough close them for security reasons. The most widely used model, integrated across Bytedance apps and operations, has never been open and is most closely associated with the state.
I would add that it is not just capacity, but also negotiation ability. With scale comes the ability to negotiate better prices than everyone else. Even if you can find capacity for your smallish user base, your inference cost can not match these companies unless you have a technical advantage for your inference cases. Squeezing the hardware requires request batching and caching which are far easier at scale and sustained user activity.
Export controls have highly motivated China to figure out how to make state of the art chips entirely in country.
It’ll certainly take years but I would not bet against China’s ability to manufacture something.
Thanks for sharing.
Is lack of inference chips due to the trading blocks by trump administration? What if Trump agrees to sell chips to china, would they collapse then? That's not a very strong position to be at
Most discussion in recent years about chip fabrication shortages, expansion, etc has focussed on leading nodes (<7nm) and AI/computer chips. But perhaps more quietly in the background, China has been rapidly building other semiconductor capacity such as power semiconductors used in electric vehicles, wind turbines, solar modules, train traction systems, etc. For example, Chinese-produced motor vehicles (37% of global motor vehicle production in 2025) in a year or two are targeted to use 100% domestically produced chips, and this production is decreasingly dependent on imports, even for factory tooling.
The report at [1] is a good summary of long term trends for China's rise in domestic self-sufficiency for semiconductor manufacturing. The report predicts "At current pace, China may achieve self-sufficiency in semiconductor manufacturing by 2027-2028, though trailing at leading-edge nodes". By contrast, before the first Trump presidency in 2017, a chart shows China importing 30% of all globally manufactured semiconductors (and increasing). Other reports on semiconductor fabrication equipment sales show the means, which is China having been and continuing to be in number (1) position for expenditure on semiconductor fabrication equipment.
The reports at [2] and [3] are also a good summary of long term trends for semiconductor foundry capacity predictions to 2031. A prediction is made that China's current 12% global semiconductor foundry supply capacity (across all semiconductor categories) in 2025 will expand to ~30% by 2031.
[1] https://www.yolegroup.com/product/report/china-semiconductor...
[2] https://www.yolegroup.com/product/report/status-of-the-semic...
[3] https://www.yolegroup.com/press-release/the-global-race-for-...
is that releveant if people can host their own models? that activity still undermines the valuation / diminishes the US companies 'moat' ?
People can host these models is doing a lot of lifting here, these are models that depends on 5 digits on specialized installation to run on.
IMHO, this has the impact of softening the impact of data centers sitting unused in the long term if they can still serve open weight models, even if Anthropic or OAI have to scale down their expansion rate to pay the bills.
Regardless, reality has to give at some point; these valuations don't make any sense. We've been valuing GenAI as disruptive work, when in reality they're much closer to cloud providers with a beefy, one-pony-trick R&D department.
they have the capacity via market manipulation; so you know, they only have things they've bought on the governments future debt obligations.
so, you know, they're as vulnerable as utilities at this point, if only there were people who gave a shit more about society than greed.