> When entrepreneurs walk into the offices of Andreessen Horowitz (a16z), a big American venture-capital firm, the odds these days are that their startups are using AI models made in China. “I’d say 80% chance [they are] using a Chinese open-source model,” says Martin Casado, a partner at a16z.
This is very different from what the author portrays. It may be the case that many pre-funded startups are using Chinese open-source models (somewhere in their workflow). But what percent of startups that survive more than a year (either with funding or revenue) are still doing this?
I imagine their pitch is: "look at how well we're doing using open source Chinese models! We'll do even better once we raise money to be able to afford frontier models!"
The way the author presents this quote makes me think he had a preferred narrative and found quotes to back it up. Or he's just a very uncareful reader.
"Well, not quite. I'd say 20-30% use open source. Of those I'd say 80% use Chinese based models. So closer to 16-24%."
Reads pretty differently.
Just anecdotally though, my company is not a startup, well established and well known and has already started investigating, purely for dev purposes (not product), using Chinese models - this was spurred by costs rising much faster than expected.
So while I agree that I don't think it's anywhere near the 80% level across the board - I wouldn't be surprised if it starts moving that way.
A VC partner meeting with early-stage founders is focused on the viability, uniqueness and defensibility of the IP tech stack not what tooling the coders are using. The developers could be using Claude or GPT 5.6 to develop a tech stack based on open weight models.
> I care about having open technology that can be run in the public interest, aligned with the public’s values. Threads like public AI, federated services, and open research have traction but need backing. Getting there in the US needs more nuanced strategy and support than we’re seeing today.
> and Chinese models are poised to take the lead.
makes it sound like the second part is a continuation of the first quote from the same source, but actually the second link is just some random person’s substack post from almost a year ago.
If you saw the engineers you’d see 80% Macs and 20% Linux laptops.
The statistic would technically be true.
I use Chinese open weight models a lot, but they’re not what I reach for when I’m doing important coding work.
The ai libraries we use let us switch models with just a configuration change.
We used to pay OpenAI >1m$/month for fraud classification, NER, etc. Sadly the US companies no longer care about non-coding-agent uses.
I imagine uptake will continue to increase as the corporate infra improves. Right now it's still bad - for example, AWS Bedrock is awful, models are months late and implemented with basic errors. Google Vertex is even worse. Finding a decent provider is the hardest part.
At posthog we see if a customer is using an llm, they use more than 1 model. The typical pattern is frontier models for a small percentage of 'harder' tasks and then one of these chinese models for more standardized procedures. As you get better at standardizing procedures you are able to use the chinese models for more and more work so token usage goes up, but the $ spend on top models has still been growing
I put my foot in the mobile comparison the other day, and will again. If you were to go back and be a mobile dev in 2010 by all means specialize on one platform, but play with both as a professional interest to stay realistic. Here it's important people have access to US/Chinese/Other, open/closed, local/cloud and that this remains. Don't become a blind Claude guy or a open weights fanatic: that way lies disappointment.
Deepseek is barely behind frontier models while 10x cheaper and 99% discount for cache.
"People with mostly bad ideas/execution use Chinese models." Is the point being made.
If you slice it to some measure of success, is the statement "Successful start-ups/companies use Chinese models." still true?
Data would be needed to argue the 80% skews unsuccessful
The Ai is writing code, not executing a startup. The code was never the hard part of startups
- fine-tuning
- running in your environment
It's like picking AWS vs GCP. Yes it is a business decision, but one that will not likely affect the outcome of the business.
We don't know that, that's the point of my statement about changing the question.
Do successful companies opt for the US/Closed models? If they do or don't it's just a correlation but it means something. Maybe it's just causal of companies being able to get more funding because the ideas are better so they opt for the more expensive model (assuming it's better).
1. Different people using Ai have different outcomes
2. There is much more to agents than the model
3. Companies are not successful because of the code, look at how many shitty products we endure
Can you explain the basis for your insistence that models matter?
Can you name another technology that determines success/failure rates?
It's not my insistence that they matter, it's my insistence that How many companies use which model isn't a measure of success. My insistence is that a better measure to determine _if_ models matter, is to ask which models successful companies use. It's not a perfect measure, as I mentioned, it would simply be a correlation but a causal link doesn't exist with out a corollary one.
> Can you name another technology that determines success/failure rates?
I'm not sure what you're getting at with this question but of course. Electricity, machines, computers, etc.
Speaking of electricity and, as an example, you could run a similar thought experiment with companies who chose to use AC or DC power when that was a thing that needed to be chosen between. It turned out, there were niches where each made sense. So the actual question here is probably less about is open/closed better but rather which situations are better for which model. Obviously you can't fine tune a closed model so if you need to do that, your options are limited.
But the big price is AGI and who gets there first, right?
At least not what I think most people imagine when someone says “advanced general intelligence”
Which is what we all called AI before the nomenclature goalpost moved
Model ai startups start from OSS models, and use them extensively for different purposes as their work would usually be banned by proprietary labs.
Application ai startups don’t want to fight the model game, so they either pick the best or let the user control it.
everyone else seems to be thinking the quote is about AI-assisted coding but I read it as the model being used within the product itself
If you're putting a lot of your money and time into a business, do you really want it built on a service only hosted by one company that will turn it off eventually and you have no recourse?
If you build something against an open model you can take that and run it anywhere. If your favorite model provider stops hosting it, you can go elsewhere, you can go rent GPU instances, you can even shell out and buy hardware to run it yourself if you've got the capital and it makes economic sense. Change some API keys, update a URL in your config, and you move on.
If the government decides that proprietary model is too good and so it gets shut off, what do you do? If a proprietary provider decides it's not worth it for them to continue hosting that model, what do you do? If that provider silently updates the proprietary model and it makes your app broken, what do you do?
The model I use to vibe code with I am just going back and forth with. Since Im building it as I go using an agent doesn't make sense but I guess that's where all the token usage comes from? Pardon ramping up my skills via vibe coding this one idea for about a month and have never hit any quota and or have gotten anywhere near my limit.
On the other hand, when developers are developing, they mainly use US AI because the quality is better.
When developing AI related services, they prioritize Chinese models due to lower API costs.
It seems like the article didn't make this distinction.
So the claim that Chinese AI is the top choice for service level AI isn't entirely wrong.
All of a sudden in almost all social media channels I'm seeing this type of content and then its usually upvoted to the top.
Non-gatekept forums like this are exceptionally easy to astroturf.
It will cost you more than it saves to use smaller Chinese models to code; because of the repeated work. That has been slowly changing recently, but with much larger Chinese models, however those models are so expensive they're much more price-competitive iwth the US competition.
But for actually providing end-user AI features, particularly simpler ones, the US isn't even in contention. The costs and limitations just outright kill those features conceptually.