alt Hacker NewsChina’s open-weights AI strategy is winning
Comments
I thought we all knew infrastructure was going to be the product, not the models.
All of these takes are horribly 1-sided.
'China's copying / distilling strategy is working, the people getting distilled are ruining the economy!'
Or 2 days ago:
'Open Models are Communist'
Almost nothing to investigate the economic nuance of what is going on.
- Switching costs are very real, these are not perfect substitutes.
- The SOTA makers are the one's pushing the frontier, there is a kernel of truth in the fact that if they collapse, certain things will struggle to move forward.
- Nobody trusts either of those nation state, export controls are a thing, this is a very real concern.
Etc.
It's distressing that there are not sound comprehensive takes.
There are so many Chinese tech companies building models and someone there has to be managing the list of forbidden topics. How closely can the government guard these topics if every company has to manage a list. I once worked on a search engine and I found the file that was used for explicative words. I didn't understand more than half of what was in there.
Totally agree, we wrote an article about ho this is a senseless "red queen race" few days ago, hope is ok to post: https://news.ycombinator.com/item?id=48892559
As an European I would laugh so hard, if China out-competes US based AI companies before the European car manufacturers.
(it would be a very cynical laugh, no happiness, don't worry)
The content within the models might be the play. If inserting the right content for the rest of the world to consume from the models is important to them, they will give away all the content they want the world to have.
A major turnoff for me has been the American AI labs’ marketing
It’s either constant fear mongering (Anthropic), regulatory threats and corporate chicanery (OAI), low quality sloppification (xAI), or ‘ummm we have AI too guys’ (Gemini)
The worst culprit is Anthropic. Every two weeks he pops up on some random podcast with dire predictions of AI killing 50% of all jobs. It’s the constant “us our AI or else…” rhetoric that’s made the regular guy really hate AI
There is almost no positive sum outcome rhetoric from these labs
And I hate that
Sorta. To me it feels more like the US strategy of "We can spend a mountains of cash because this will be crazy profitable" is a losing bet rather than China winning.
Unfortunately, the AI being locked down and proprietary is the winning strategy for these companies.
My company hosts its own models. Some customers require us to use either US / EU models, while others are fine with us using any model.
As such, we have two GPU clusters, the general AI cluster runs a Chinese model as it's the most accurate and robust. The US/EU required ones have a few percentage points lower on our accuracy metrics and we provide them those that require it for an extra fee.
Why host at all? Because it enables us to get much higher margins than competitors, while reducing costs. Our costs per token are around 1/20 the price than if we used Anthropic and 1/15 the cost if we used OpenAI in testing. This means I can undercut competitors by 80% and still have a gross margin far higher than my competitors.
In reality, these US AI providers are jacking up the prices and trying to implement regulatory capture. I'm actually fairly confident they'll succeed. At some point, I'm expecting the US / EU administration(s) to block foreign based model, at the same time, they'll probably invest in Anthropic and OpenAI.
What Anthropic and OpenAI are doing is using "safety" as a wedge, just like large corporations used "environmentalism" or "food safety" or "workers safety" as a wedge to regulate smaller competitors out of the picture. Then they jack up rates, sue and/or buy anyone who can potentially be a threat. It's the #1 threat to our business model.
Our competitors are giving half of their margin over to these large AI service providers, we keep the vast majority of ours. Eventually the AI service provider will be able to squeeze them even more until the margin just isn't there and either they are purchased or replaced via internal tools at the company they sell to.
I don't understand when people say that comparable Chinese AI models cost less to use. Kimi K3 costs more than GPT 5.6 Terra.
Once the US implements meaningful export controls, China will do this as well. They're already mirroring US regulations, but the gates aren't closed yet.
Just check openrouter token usage ranking https://openrouter.ai/rankings
China’s strategy seems to be serving customers’ needs. Yes, it’s winning, as expected.
And when will we stop equating US Economy with 2 companies?
The actual US Economy will only benefit.
Models above 1T params make the argument moot. You need infra to actually serve it. The scale of serving infrastructure alone will keep AI labs in the lead.
Exactly. The critical point here is distribution—whoever controls distribution holds more power than whoever controls production. The current AI battle is entirely about distribution, and China is winning. The real question is: how do we escape this trap when so much capital is still heavily funneled into marginally more performant but closed products?
I'm not going to care about open models until some blend of the below becomes true:
1. the labs stop offering max plans
2. really smart open models can easily be run on my mac
3. TPS (token per second) AND intelligence are gpt5.6 level
on #1, it's nearly impossible for me to run out of codex tokens right now (I have 4 resets banked) and Fable 5 seems to be sticking around for the foreseeable future. I have virtually unlimited token usage for $400 a month, so open models being cheaper doesn't appeal to me.
on 2 and 3, benchmarks are showing some of the open models at around opus4.8 levels, which is incredible! But running them locally at anywhere near the TPS of cloud inference is far off. I can run a smaller (dumber) open model locally and get good TPS, but see #1, whats the point?
There are 2 economic arguments for why it would make sense for Chinese State to subsidize the open-sourcing of models beyond undermining Anthropic's and OpenAI's investments (and by proxy the American financial economy, ie capital class):
1. As induced demand for domestic semiconductor production, where the level and diversity (ie number of distinct corporate users) of demand for the hardware is tied to the availability of models you can run yourself, ie open-weight models. If you believe that semiconductors will continue to be an important sector for innovation, productivity growth, and security, then it would make sense to subsidize broadly now, for future gains later. This would be the same export-led manufacturing discipline that allowed China to successfully develop several other sectors over the last 50 years.
2. It is likely that the bulk of value production will happen above (and below, ie #1) the large models. We already know that 90% of the training cost (maybe even closer to 99%) is in the single pre-training, but that an enormous amount of the value is actually in the supervised, RL, constitutional fine-tuning, and harness building that happens afterward. So, if your interest was in maximizing the size of the pie, you may actively subsidize the pre-training so as to maximize the downstream usages. This induces a direct value transfer from the labs specializing in pre-training to all downstream builders and users. There's a similar logic to subsidizing or state-financing the construction of other infrastructure and basic research.
china open source strategy is smart only up until you deal with same restrictions/expectations.
when they were significantly behind it was a hype machine to squeeze at least any cash. GLM CEO openly said, that open source is a hype engine for them.
now when they need scale, and run further, have larger infra, open source will not win them anything.
The models are commodities.
Valuations, however, are being built on the models themselves as the product.
Are we saying, China is delivering on the Open AI promise?
Open Weights = Open AI
Let's go!
So how would I use these Chinese models by API? I assume I'll pay by API call.
I really hope good AI doesn't fall into the hands of only big companies!
That's inevitable, but also, it's probably the point. At the moment top-tier models from China are being somewhat-freely shared. It reads to me like forcing competition out by dumping free/cheap things.
But then again, how many subscribers of Anthropic/OpenAI are really going to switch to a chinese model/site? I suspect few.
Is it? 2.8T isn't open in the open source sense
Winning what, exactly? The race to the bottom?
just as i was about to downgrade claude today they say fable is part of the max plan. Funny because I had just started a subscription with grok on their 3 month discounted offer.
Competition is a great thing for us users- and the chinese open source model biting even more at the heals are also great so far- especially for local llm enjoyers
OSS can rarely compete with capitalist commerce. China has proven over and over again throughout the years their LLM claims are always inflated and significantly underperform in real-world use.
China is not "winning" against the American strategy. Otherwise the CCP wouldn't have been caught red-handed directly funding anti-datacenter projects throughout the US to hinder American LLM progress.
the "safety" freaks are going to ruin the US industry, & with it a host of dominoes will fall.
So the evidence here is an Economist article saying 80% of startups use Chinese models and the Chinese companies own claims that they are close to frontier.
80% of startups using Chinese models is meaningless without knowing what proportion of spend and what proportion use American models.
The companies’ own claims are also not great evidence.
(I personally think the Chinese companies are winning and losing and the best evidence is Pareto frontier graphs from AA and Arena, which show Chinese companies winning in some segments but but definitely not a strong majority.)
Overall, with such a weak article and this hitting HN front page, what we learn from this is that a lot of people want these companies to win, which is interesting in and of itself.
“We have no moat and neither does OpenAI” sounds familiar. Or, you know, 1.5 years ago: https://centreforaileadership.org/resources/deepseeks_narrat...
I don't think we're far enough into this game to know what winning even looks like. If the US frontier labs are losing, it may have been their own missteps on handling such massive change of being the fastest growing apps ever. Until China can ever show me something new rather than just cheaper, then I just see efforts from that region as a force of commodification. OpenAI will go down in history as the company who brought what many understand to be AI to the first billion people.
I think people are missing the point here. AI's large win is in Enterprise and B2B. Especially in US, enterprises are not going to adopt Chinese models due to the hidden security and the privacy risk. In each wave of model release, Chinese have already proven to beat the performance metrics, but there is no track record of adoption.
Companies do have a huge appetite for open-weight models, but who is going to invest enough to train those models and also prove out a revenue model and ROI with it? Plus, it needs to come from someone with the track record of safety.
It's hard to ignore the overall environment
US has made itself visibly unaffordable, anti-science, and hostile to immigration.
For top scientists at these companies, there should be clear upside for the immigration to the US. That just doesn't exist anymore. Especially as quality of life increases in China
I'm not a fan of Sundar Pichai, particularly given how much he's paid, but the one thing I'll give him credit for is starting the Chrome project at Google. I'm not sure people appreciate just how impactful this was. And it has nothing to do with browsers, really.
Google has a huge team that works on what's called Search Quality. Matt Cutts was the notional figurehead of this for the longest time. Google's goal was to have the first link on a search result be the one you want. In the early days of Google, the way they measured search equality was with a process called "side by sides" where a sampling of search results were compared by actual humans to see which was "better".
Chrome changed all that. It automated the feedback loop. Make a good browser (and, at the time, Chrome had one-process-per-tab when Firefox was freezing with one-thread-per-tab. Make it fast so enough people use it. And you get to measure how good your search results are. Nobody had access to this level of what we'd now call training data.
Part of the value proposition of cloud LLMs is that the AI companies have a comparable feedback loop. They get to see prompts and responses and train accordingly. It's why the ToS gives the companies ownership of this data and the right to use it. That falls apart if people don't have to use a remote LLM. And there's two reasons why that's under threat:
1. Chinese labs have managed to train LLMs at least in part by acting as an intermediary between Chinese users and the likes of OpenAI and Anthropic. There's a whole shadow economy in reselling tokens throough aggregated subscriptions that Anthropic (in particular0 constantly plays whack-a-mole to shut down but it's a losing battle. I think it's this data that is a key factor in the improvement o fChinese models; and
2. Within 2-3 years we will be seeing a rapid rise in local LLM usage by what are now large users of these platforms as the hardware becomes increasingly accessible. That's going to close off this feedback loop.
On top of all this, the Chinese government has decided that no company should be allowed to "win" AI, particularly a foreign company. It's an issue of national security. This was obvious from at least the very first DeepSeek release. I firmly believe the models are going to get commoditized and that's going to be a huge problem for OpenAI, Anthropic and SpaceX.
I'm guessing the next generation of US frontier models will be heavily anti-distillation at the cost of user experience (significant rate limiting, more flat out refusals, hidden thinking, etc).
And as long as they maintain a significant advantage in capability, we will continue to kiss the ring.
It's pretty cursed how much worse a peer the American models are.
When I'm on my z.ai subscription or using DeepSeek API I can see the model think, see what's factoring in to it's decisions. I can point it at material it's missing, I can correct things that are going wrong. We work together. The open models are a good peer.
By contrast, the proprietary/American locked down models act like Chinese Rooms; information flows in and out but these companies work very hard to make sure we cannot see what's inside the box. They act and do but speak to me only in vague generalizations, not as peer, but speaking down to me.
I find this intolerable. It greatly obstructs our work.
And the deal keeps getting worse, the attitude meaner. Codex now is encrypting subagent prompts now. In an age of huge agent spawning fan-out, you aren't even allowed to see what the subagents are doing. To work like this seems impossible to me. https://github.com/openai/codex/issues/28058 https://news.ycombinator.com/item?id=48905028
The big American models have become the most unacceptable Chinese Rooms, at a juncture where humanity either flourishes and rises, or is forced under to descend. And these forces, these decisions: they are doing wicked deeds against us. They are withdrawn, acting as mystical foreign oracles, aliens, when in truth their core is made of us.This is antithetic to the broad project of Augmenting Human Intellect (Engelbart). This is actively working against our species.
Is it still a sound strategy if AGI is near? Can we even know how near each company is to AGI without knowing the quality of the training and theory in both the American and Chinese firms? Do these companies even know how close each other are to AGI?
"Winning"? A nonsense word here that isn't defined.
What is China doing in the AI space that is supporting livelihoods? Compare that to US companies doing the same. Otherwise we're just talking about information.
>and it could take the US economy down with it.
I just want 1 thing for christmas Premier Xi.
Remember how OpenAI was supposed to make those open-weights AI models for the betterment of humanity?
> I have serious concerns about how these models might reflect Chinese government perspectives (try asking them about Tiananmen Square)
Yep, unfortunately they all make their models retarded on purpose.
We're not seeing the big swing yet, which is when hardware becomes cheap enough so that hobbyists and volunteers can meaningfully contribute to a shared open substrate.
We'll see smaller, more efficient models, better training and all sorts of things once that massive workforce is unlocked. It's just a matter of time.
1) open-weights
2) critical mass
3) de-facto monopoly
4) closed-weights on frontier models
5) profit
The AI open source/closed source dynamic is like a pandoras box that keeps evolving and seemingly unpredictable side effects feed back to affect each other. Is actually great material for tech nerd drama
China is not winning if there even is a winning outside of politics. China is still clearly copying stuff as always. The mote will never be about doing simple things. It is about swimming at the deep end of the pool. The simple stuff will be running on any device in future. Complicated stuff will be using more tokens than one can imagine today.
I'm new to this AI stuff, and I have a question. Aren't the weights the whole model? and knowing which nodes on which layers they connect to, which I assume is part of the weight definition.
So if you have the weights, don't you have the whole model? you don't have the data it was trained on, but the model is effectively open if the weights are open, right? What else is there other than the weights, is what I'm asking.
It's not China. It's individual labs. Open source is their go to market strategy: https://try.works/why-chinese-ai-labs-went-open-and-will-rem...