Comment by overgard
1 day ago
I do think open-weights models are going to "win" in the sense that they're probably going to be dominant when the hardware to run them becomes affordable. (which might be a while). Although I guess you could probably rent the GPU's yourself to hypothetically save on costs. (I'm a little skeptical -- I've heard of companies doing this and the inference bills are surprisingly high -- assuming the sources are correct. I don't know if a lot of people really want to be advertising "oh god our bill is horrible")
I'm sort of baffled by what the entities that train the open-weights models get out of it though. Is it just a direct play to undercut the US providers because they view them as a threat? I just don't really understand the business model behind it.
If you're NVIDIA then open-weight models are a classic example of commoditizing your complements; cheaper models mean more people buying GPU's to run them. [1]
Your guess is as good as mine for China though.
[1] https://www.joelonsoftware.com/2002/06/12/strategy-letter-v/
Nvidia selling more GPU's at the cost of it's datacenter business is pretty close to Kodak selling digital cameras at the cost of film.
The data center side is so bloated anything that eats into it is a huge negative. Their data center business brings in 20x the gpu market. Local open weight models will be what pops the bubble and China will do anything in it's power to enable that pop.
I wonder what the thinking inside NVIDIA is at the moment. They have countless examples to learn from here, about the danger of not being willing to cannibalize your high end products. But, of course, there’s a reason that there are lots of examples of this sort of failure.
There’s plenty of competition that would be happy to attack them from below, though…
7 replies →
> Nvidia selling more GPU's at the cost of it's datacenter business is pretty close to Kodak selling digital cameras at the cost of film.
This assumes a) AI is a zero-sum game, and b) we're actually talking about on-prem AI will replace cloud-based AI. I think neither statements are true.
AI is like compute: we'll need all sorts of it, in various sizes, everywhere. I'm sure Nvidia whats to own all the workloads.
On the other hand, I do think open weight, like open source, will win in general.
2 replies →
> Nvidia selling more GPU's at the cost of it's datacenter business is pretty close to Kodak selling digital cameras at the cost of film.
You mean they resisted the idea trying to protect their legacy business, and it ended up all but killing them?
9 replies →
Hmm this sounds like an incumbent missing a paradigm shift because they didn't want it to disrupt their core (usually enterprise) business, although riding the shift would have ultimately delivered an order magnitude larger business.
Classical example is Microsoft actively undermining mobile because it threatened selling Windows or enterprise licenses.
Or Yahoo fighting Google's model because the latter model's didn't depend on taking enterprise deals to rank results.
2 replies →
Could you explain what you mean by this? I thought all of their insane profitability and returns are from crazy margins on their GPUs. I know they started/partnered/invested in some data center businesses, but I thought they were fledgling
Their data center business is selling gpus. They wrap them in a complete platform, but that's what they are selling.
They do report it separately from consumer and business sales of gpus used in PCs.
3 replies →
Could you put that in units of HP Printers and Toner please?
You need 64 H200 super-node for inference for kimi k3. You will not do inference locally. What might pop the western hardware bubble is Chinese GPU, memory, networking companies. But even in China, these AI centric hardware is not cheap.
1 reply →
There’s export controls on nvidia. According to Jensen, they expect that Chinese models will start being optimized for Huawei
https://www.dwarkesh.com/p/jensen-huang
They already are.
No, Nvidia is worried about Chinese hardware stack. China is under sanction, what are they training and inferencing these models on? Even if this model might still be Nvidia chips, what about the next model. Everyone knows model and hardware companies are working together. There are a number of very competitive companies in China in this space. After they got this area sorted out, China will do training, inference and tokens entirely on their stack and export their entire stack.
With their models being open-weight, any two-bit firm in the world with enough capital to invest in a few servers can become a provider capable of carving out their own little niche in the economy.
The Chinese see this as a lift on the entire economy, as it comodotizes the technology to a degree in which many firms can serve many sectors of the economy, a true total-economic win worth the public investment.
The American strategy is built off of private investors believing that with enough money poured into as few companies as possible, one or two firms can come to dominate the entire market and start charging an ever burdensome "tax" on every sector it can touch. Not what I would call a total-economic win for the country.
Additionally I think there is something to the idea that they are trying to undermine foreign competition as a stall. They can’t compete economically for geopolitical reasons right now but they also can’t let American firms dominate the rest of the world in that market/technology stack.
The U.S. model is honestly more baffling to me, if the goal is broad economic growth.
But it hasn’t seemed like the U.S. powers have been interested in broad growth for quite a long time now. Just whatever can line their own pockets.
The Chinese strategy is to fool naive westerners into thinking that's their strategy (looks like its working), until they can buy time to do the American strategy.
Hence, buying up and closing up foreign competition then whining about it when it's blocked: https://www.dw.com/en/china-firm-seeks-damages-over-state-co...
While I completely agree with your take, I think everyone has been surprised by how quickly LLMs have become highly useful and extremely powerful, and by how possible it is for relatively smaller models to also be highly useful.
Given that, I would expect that in hindsight OpenAI and Anthropic would spend 40% of what they have on compute if starting over and knowing the actual landscape.
The massive capital allocation was a blind decision and they swung big.
It is still possible that techniques will be developed that create a moat where the massive hardware capex is justified, but US policies of banning competitive GPUs and blocking frontier lab releases makes such things far less likely.
"Escape velocity" for AI is when the open weight models are good enough to help drive the next frontier innovations/techniques. I think we are close to that if not already there, at which point it's a race to commoditization no matter what Altman or Lutnik wish will happen.
Don’t forget it took that huge spend to publish the papers and get to the models we have. It’s not obvious that without them we’d have LLMs springing up out of China or anywhere else.
2 replies →
China is the factory of the world. They don't need software to win. Rather they prefer software is free and they can win in hardware. So if AI inference is free, they can put it in as many hardware components as possible and sell them in the market - think toys, cars, tools with chips manufactured in china optimized for the use case. In long term you tend to commoditize hardware. We have thousands of device types of cheap x86, and with linux/bsd software on it coming from china. Why do we think GPUs will be different.
- Software is copy-able for nearly free. - Hardware is not. Even if you have the plans, you've to build the infrastructure for it, and its a fairly finite, physical item.
Both are however hard and expensive to produce. It's much better for profits to develop and sell the hardware, and copy the software someone else spent resources on.
America was the factory of the world before and England before that. The temptation of “moving upstream” is irresistible.
So, "commoditize your complements". A big part of Microsoft's old playbook (and a smaller part of its current one), so we know it can work.
We just sacrificed our entire lucrative SaaS market to this, and perhaps even big tech itself. To the altar of AI.
And now it's all going to become commoditized.
Billion dollar software will be commodity. Salesforce. There are orgs already moving to their own internal tools.
It does not seem hard now to rebuilt Google Search, Google Chrome, Gmail, Gsuite, Netlify, Vercel, Cloudflare, Vimeo, Twilio, or even Stripe. The cost barrier has to have dropped 1000x, maybe 10000x.
We have millions of engineers with the talent to do this. Many of whom are unemployed and have savings and nothing better to do. They could easily carve these markets into pieces.
We shouldn't shut down open weights. It's too late. They'll win, and that's a good thing. Big tech was a thermodynamic bubble of high energy waiting on the dam to burst, and now it has. The genie won't go back into the bottle, and that's totally fine. It's progress.
Now we need to rebuild our factories and supply chains and energy and resource inputs. Because the back half of this revolution is going to be robotics and factory automation. If we don't have the connective tissue in place, we're really going to hurt.
We'll do well if we regrow manufacturing. If we don't, we might be in for a world of trouble.
You're right about it being relatively easy ro build your own version of something right now in the proverbial garage. But startups will still have to compete with the network effects of the large firms and the increasing push for data sovereignty from non-US customers. Not to say it's not possible, just that the technology alone isn't sufficient.
> It does not seem hard now to rebuilt Google Search, Google Chrome, Gmail, Gsuite, Netlify, Vercel, Cloudflare, Vimeo, Twilio, or even Stripe. The cost barrier has to have dropped 1000x, maybe 10000x.
[citation needed]
> I'm sort of baffled by what the entities that train the open-weights models get out of it though. Is it just a direct play to undercut the US providers because they view them as a threat? I just don't really understand the business model behind it.
In China, it's because they are being heavily subsidized to do the research activity. It's not really complicated -- if you allocate public money for people do to a thing, they will do it.
China is known to spread love and kindness through markets with no self-interest, after all.
The US funds (and used to even moreso) scientific endeavors that stand to bolster the entire country all the time. Medical research is the obvious one, DARPA is the defense based one (though oftentimes defense is just a post-hoc justification for a lot of those), the Department of Agriculture is constantly researching improvements to the farming industry, the Bureau of Weights and Measures and the NSA fund cryptography research.
Research and development of new technologies often is a "rising tides lifts all ships"-type deal, which it is absolutely in the government's best interest to support.
For China, the best case scenario would of course be to control a locked-down best-in-class frontier model that the rest of the world becomes reliant on. The US seems to be beating them at that, and "Everyone is reliant on the United States" is a pretty bad scenario. A middle ground, positive outcome is that no one is reliant on locked-down closed models, so they're supporting that outcome.
It's really not that nefarious.
"public" doesn't mean "love and kindness with no self-interest".
We only think "government spending is for hippies" in the US, and only when we don't look at public spending like defense bills.
Of course it's in self interest. It's too bad that the US has gone the other direction and cut public funding for research.
> it's because they are being heavily subsidized to do the research activity
Is it actually true? This seems pivotal because currently most theories rest on the idea that individual Chinese companies are acting in China's overall economic or strategic interest. It's a tough sell to believe they all just do that through implicit desire to align with the CCP's direction. I would believe it much more easily if there were concrete incentives involved.
Yes, incentives are why people go to work in the morning. https://www.rand.org/pubs/perspectives/PEA4012-1.html
You could say the same thing about Linux in the 90's and 00's. And yet here we are.
Remember the Halloween papers?
https://en.wikipedia.org/wiki/Halloween_documents
[flagged]
You're fixating on what's a meaningless blip in the story; the technological world was built on post-WW2 US spending that ended 40+ years ago.
1 reply →
China wants the US economy to flounder. Building our entire growth model on software that can be copied and taken by a small group of people will have no possible consequences.
> China wants the US economy to flounder
It’s not that simple. If the US economy goes into a recession, it will take large sectors of the weak Chinese economy with it either directly or indirectly.
It’s probably more accurate to say they don’t want American LLMs to become dominant. The huge US data center build out doesn’t depend on Anthropic and OpenAI anyways. Those data centers can just as easily serve Qwen or GLM models.
1 reply →
China is looking after their own interests, but they absolutely don't want their largest export market to struggle. The global economy is not a zero-sum game, and the idea that it might be is the root of many of our policy issues in the US.
2 replies →
I don't think it is true that they necessarily want the US to struggle, I suspect it's more self interest. LLMs seem to be one of the biggest innovations of the last few decades, China probably just wants to make sure it's not being left behind and/or made hugely reliant on the US for what seems to be turning into a piece of critical infrastructure.
China has an effective strangle hold on some key sectors (solar, rare earths) and I am sure they relish this position and the leverage it gives them. You'd be careful not to give away that same leverage to a competing power if you can invest a few billion now and cover your bases.
Assume, crazy business, that china wants what’s best for the world. Recognizing the danger of an AI arms race rapidly producing uncontrollable superintelligence, it focuses on open weights to reduce the economic incentives for further advancement of AI beyond the “highly useful for humans” stage.
Seems like the only thing that could avert an intelligence rapid take off. Everyone wins except for shareholders.
> China wants the US economy to flounder.
Why would they possibly want their largest customer to flounder?
3 replies →
The fact that the US economy is load bearing towards unprofitable projects while things like medicare for all or universal childcare continue to not exist is just a damning indictment of the country.
I don’t think that’s true. The US is China’s most valuable trade partner.
3 replies →
The Chinese domestic market is extremely competitive on most things, including LLMs. Once one top company went open weights there, that's going to put pressure on others to follow suit. It'd be akin to if a company like Anthropic or OpenAI went open-weight, it'd probably result in a domino-effect of more US models going open-weights since otherwise that competitor is going to win a huge chunk of mind/market share for free.
It's also relatively free right now. Few people are going to run local models, and in the future it's likely that every model being released today will be obsolete. The only real downside is ease of distillation for competitors, but that's probably impossible to stop anyhow.
Undercutting US dominance in AI is huge for China. If the entire narrative is that you have to use Anthropic or OpenAI to access a decent model, then China's AI labs are sitting on the sidelines as some third rate solutions. China publishing the model weights of models comparable to the frontier proprietary models drastically undercuts closed labs dominance. Maybe these Chinese AI labs don't have the billions infrastructures some of the US players do, but they don't have to if the model is open weight. Many inference provider companies around the world have hardware that can run these models and they will happily run frontier class models for people. Starting in 7 days, people will have the option of which of many providers they want to use to access K3.
Making frontier grade models a commodity will make a competitive market where companies compete for business by improving their quality and decreasing their prices. The cost to access frontier grade models will continue be driven down the more competition that enters the market. This commoditization will challenge the valuations of Anthropic and OpenAI.
They get money from subscriptions and tokens, same as for closed-weight providers. Yes they'll lose some traffic to hosting services, but many users prefer to use the original training company since they have a guaranteed-correct implementation. Similar business model as open-source SaaS companies.
Some companies (most notably Deepseek) also manage to host their own LLMs so efficiently they undercut all third-party hosting services.
No one cares which company they use. They care about cost and does it act in a way they expect. Expecting anything else is pretty laughable and goes against human nature. You either create a moat so deep no one else can play or you have the government force users to use your products.
> I'm sort of baffled by what the entities that train the open-weights models get out of it though
if you do the training then you're in control of the output. For example, recommending your products/services or failing to mention your competitors. You could also automatically introduce backdoors into code deemed interesting, i'm sure all governments are very interested in having that influence.
Absolutely on the industrial backdoors, but also on the consumer front: try asking Qwen anything about tienamen...
That seems hypothetically possible, but kind of hard to do? I don't know a lot about training, but it seems like they wouldn't have exact control of the data that goes in at that scale (scraping the internet) so it seems like it would be kind of hard to do that in a way that's subtle. I'm kind of reminded of Elon Musk trying to put his political views into Grok and it seemed like it created huge technical problems with the model saying some really out of control things. Maybe it was just an xAI issue though.
It’s totally doable and is done right now in the form of “ai safety” guardrails. No reason you can’t have guardrails to reform responses about your competitors as negative. Or, when asked about current events, only give a one sided opinion.
> I just don't really understand the business model behind it
There’s a lot of value in the same sense there is a lot of value in controlling what Google search results are shown and what people see in the Twitter feed.
Probably some value in kneecapping the frontiers too.
> I'm sort of baffled by what the entities that train the open-weights models get out of it though
I agree. The thesis in the article is interesting insomuch as I had not heard it expressed this way before: US restrictions on GPU exports have made it feasible to train models in China but not serve them. Therefore open model is a hack to get around the export restrictions, since models can be trained internally but shipped out of the country to be served elsewhere under the banner of open weights. I don't really buy this argument - inference is much cheaper than training and they are hosting their models anyway.
I think it is more likely (a) they have the money to do it and they need it for internal reasons - these are huge companies (b) there is a lot of prestige in China associated with besting American technology (c) people are still basing logic on outdated ideas of Chinese capability which are no longer true.
So it is easier than people think for Chinese labs to do this, they need to do it anyway and there is a lot of prestige from opening the weights. It is honestly not that different to why American companies themselves have released open weight models.
> I'm sort of baffled by what the entities that train the open-weights models get out of it though.
Why is it so baffling that people want to build great things? There are plenty of people who are happy building things for a salary and have no interest in taking over the world. Do you find the whole world of open source software baffling? Linus Torvalds and Richard Hipp and Antirez created the world’s most prolific software products and released it for free.
I don’t think they need the hardware to become affordable (as a regular end user).
They need their models to be good enough and cheap enough. Then the rest will follow. Companies will figure out how to host them for you efficiently, and you pay them monthly.
I don’t think I will ever want to set up a home server, no matter how inexpensive the hardware gets. At work I still use Cursor (with Anthropic models usually) because it’s paid by my employer, but for private stuff, I’m already using cheap models with OpenCode, and it’s extremely cheap and surprisingly capable.
I think there was around half a year between where the best models became good enough (last year December?) and where the cheap models became good enough (couple of months ago?).
I've always found it weird how ludicrously poweful personal hardware has gotten. The fact that you can't buy a midrange CPU with less than 16 cores, or that a high-end gaming GPU is 100+ TFLOPS just blows my mind. The fastest supercomputer in 2004 was 70 TFLOPS. Absolute crazypants level of power, and companies were fall over each other to get people to buy it.
'Libre Office' did not 'win'.
People are happy to pay $50/year per seat to have the extra features and to not have to deal with stuff.
There's an issue at the margins here:
$1000/employee is a massive cost - it has to be deeply justified. $50/employee is like ... $2 out of your pocket. It's an incremental cost. The CFO is happy to pay it if there is a lot of value.
A lot of software is in that later category.
Imagine if gasoline was 1 cent per litre - and there was 'free gas' but it was a pain to use, and you had to check a bunch of things. You may just pay the 1 cent.
AI is not quite that yet, but these dynamics will play out eventually, for a lot of things.
> 'Libre Office' did not 'win'.
I think France and other parts of the EU are switching over to it. Although that's probably more due to Microsoft's aggressive behavior recently. Agree on the cost thing generally, but I can't help but think that when the hype to "do AI" blows over, people are going to be casting a jaundiced eye towards data security, which probably means self-hosting and sandboxing
> Although that's probably more due to Microsoft's aggressive behavior recently.
The current US govt might have played a role as well.
1 reply →
My guess it's to mess with the proprietary vendors economically. It's in China's interest to squeeze the US AI business.
The business model is making your country more innovative, which makes its citizens richer
Americans used to do that too, spectacularly.
Just consider the two alternatives: one is your industrial sector with this incredible new automation and analysis tool available for free. The other is one where it has to pay huge chunks of its resources to overseas companies.
Would anyone besides those within China themselves use a completely closed and hidden model from China for their critical business needs?
If the outputs are independently verifiable at low cost, and the US models refuse to even try because somebody sneezed nearby and it sounded like "antigen" and not "achoo"... yeah sure. Whatever works to get the job done.
The hope is that AI will open up whole new sectors of economic activity. If you have to chose between exploring that space while potentially being exposed to Chinese tampering, versus just sitting on your hands and doing nothing... well then you take that risk.
I think part of it is definitely to weaken US providers and the US economy as a whole. China has a completely different domestic economic structure and motivations from western cultures... it's probably closest to a fascist economy mixed with a Maoist cultural ideology behind it. There's definitely winners and losers and the state tends to have tight controls over everything though.
I also think the restrictions on OpenAI and Anthropic are somewhat short sighted. In that the guardrails dramatically limit efforts towards securing your own software in many ways. Yes, it's also "dangerous" and maybe there should be a means of identifying "domestic" or otherwise "secure" accounts for those allowed to use the models without the same guardrails in place.
Does it weaken the US though? It weakens the big providers but most of the economy is in companies buying and this helps them save money.
Yes, it damages its image, this is further made evident given the amount of propaganda that follows each time. Why would you invest in claude or codex if you just read how China's stuff is better?
1 reply →
Isn't most of the economy riding on the stock of these handful of companies though?
6 replies →
GDP growth in the United States is in AI and healthcare. AI capital expenditure is around 5% of total US GDP. Housing right before the 2009 market collapse was around 6.7%.
Biggest issue I see is housing has more real value than AI expenditure. Demand is real and isn't based purely on a few companies valuation or marketing spin. Nearly 2 decades later we still haven't caught up to construction rates before the 2008 collapse. When the bubble pops it's going to really suck.
> Is it just a direct play to undercut the US providers because they view them as a threat?
It's China putting pressure on a financing strategy in the US that was always a house of cards. It could also be China democratizing something that should have ALWAYS been democratized. Maybe both when the history books on this get written and absorbed by the winners of the LLM wars.
i'm building thigns with open models: 128GB AMD 395+; expensive 72GB blackwell; old NVIDIA 48+48GB; cards.
If it weren't for the massive memory cartel of OpenAI/Anthropic et al, both Mac and AMD would be selling these things.
I repeat, the models are building, modifying and deploying almost anything on github.
Chinese companies are cut out of the market by the US gov restrictions, so making the models open is a survival tactic.
It also has the benefit of crashing the US economy.
China does not want to crash US economy. US economy strong means more purchases for our goods. But US want to crash Chinese economy by: 2018 trade war, 2019 Huawei sanctions, sanctioning at least 1000 large influential Chinese companies, 2020 Tiktok theft, 2022 total ban on GPUs, semiconductor equipment going to China regardless of end use, banning all sorts of Chinese goods such as EVs, cars, batteries, drones, even DJI cameras, 2020 fake lie about Xinjiang to implement full ban of anything made by Xinjiang, 2025 50%+ tariff on Chinese goods. I don't think building LLM was thought has important until Anthropic went on full China hating, banning anyone on a Chinese IP using the service, even companies headquartered in China's oversea offices. Did you know that a regular person in China trying to use Claude to ask some innocent question gets banned? Talk about everyone created equal and democratic world.
You leave us with no choice but to build technologies ourselves. If its bad for Anthropic, well, you could owned the market in China, oh well, you reap what you sow.
Crashing China's largest export market is devastating for China's economy.
2 replies →