80% of US AI startups rely on Chinese open-source models for survival. Investors from Andreessen Horowitz are shocked. The top 16 on the global open-source list are all occupied by Chinese entries.

In the Reddit community, a post titled "Chinese Open-Source AI Models Sweeping Through US Startups" has gone viral. According to The Economist, Martin Casado, a partner at Andreessen Horowitz (a16z), revealed that up to 80% of US AI startups are now using Chinese open-source models instead of those from OpenAI or Anthropic during their fundraising pitches.

If a few years ago people regarded open-source models as "also-rans," today's landscape is quietly being reshaped. As the poster quipped, "If 80% of US startups are using Chinese models, you can safely assume that globally, this proportion might approach 100%."

"Now, when entrepreneurs go to pitch at Andreessen Horowitz (a16z), one of the top venture capital firms in Silicon Valley, their startups are very likely to be running on Chinese models. Martin Casado, a partner at a16z, said bluntly, 'I'd say there's an 80% chance they're using Chinese open-source models.'"

If the AI bubble is really about to burst, you have to wonder: How many of today's popular AI stars like OpenAI will still be around? Have they become "yesterday's news," and will the future trend be a more streamlined, cost-effective model built on free and open-source AI? If 80% of US startups are choosing Chinese open-source models, you can safely assume that in other parts of the world, this proportion might approach 100%.

Silicon Valley originally thought they'd soon have the next AI unicorn, another Google or Meta to conquer the world. Maybe that will happen one day, but for now, it seems that Chinese open-source AI is about to spread across the globe instead.

The Economist's report also mentioned that China has a more obvious advantage in open-weight models. In a series of intelligence tests, Chinese models released this year have outperformed similar US open models, such as those from social media giant Meta. Moreover, their capabilities are approaching those of the most powerful proprietary models.

Notably, Meta, once highly regarded in the open-source community for opening its Llama model, has now shifted its strategic focus to the so-called "super intelligence" and has become more cautious about opening its models. With Alexandr Wang joining Meta and driving the strategic adjustment, the Llama project is gradually moving towards a closed-source model.

Martin Casado isn't just a VC who makes random revelations. He's a general partner at a16z, in charge of a $12.5 billion infrastructure fund for a long time. He has invested in several well-known startups in the AI and infrastructure fields, including World Labs, Cursor, Ideogram, Braintrust, etc.

Early on, he co-founded the network virtualization company Nicira, which was sold to VMware for $1.2 billion in 2012, regarded as one of the landmark events in the commercialization of software-defined networking (SDN). Casado is widely considered a visionary in the enterprise infrastructure field, having contributed to the formation of the modern cloud computing stack.

In the past few years, Casado has paid increasing attention to open source. In July last year, he and Ion Stoica (co-founder of Databricks and a professor at Berkeley) co-wrote an article in The Economist, calling on the industry to recognize the importance of open source in AI. "Even the largest and most powerful models should be open," they said.

Last month, Casado further shared his views on the podcast "20VC." He pointed out that the industry is experiencing a trend from "open source" to "closed source." Although public opinion generally supports open source, the actual ecosystem is gradually tightening, he said. "I think there will be less and less open source," he added.

He said bluntly that if you ask, "Is open source more dangerous than closed source?" the answer is yes, because China does have an edge in the open-source field. "We're really seeing a large-scale spread of Chinese open-source models now," he said.

"The right approach for us isn't to close off, but to further promote our own open-source efforts," Casado emphasized.

But in reality, industry giants are doing the opposite. Meta initially won praise from the open-source community by opening its Llama model, but as it adjusted its strategy, the scope of opening has gradually shrunk, and the project is moving towards a closed-source model - the opposite of what Casado advocates. Meanwhile, although OpenAI says it will "re-open source," its newly released proprietary model, GPT-5, has underperformed, raising doubts about the sincerity of its open-source commitment.

Casado specifically pointed out that the biggest difference between open-source AI and traditional software is that open source doesn't mean easy replication. To truly reproduce a large model, you have to rebuild the entire data and training pipeline, which is almost impossible. After all, training a top-tier model requires hundreds of millions or even billions of dollars in investment, and it can't be given away for free.

In his view, open source will still be an important part of the entire ecosystem. Historically, open-source software has accounted for about 20% of the market value, but in the AI field, this proportion is much higher. In this sense, the open-source AI ecosystem is actually healthier than in the era of traditional software.

All of the Top 16 Global Open-Source Large Models Are Chinese

Software engineer and self-media author Rohan Paul recently noticed a shocking phenomenon: On the Design Arena rankings, the top 16 open-source AI models are all from China. The best non-Chinese model, GPT-oss-120B, only ranks 17th.

Paul sighed, "China has almost completely dominated the open-source field. It's really remarkable that teams like DeepSeek, Zhipu, Kimi, and Qwen can achieve this while sticking to open source."

Design Arena is currently the world's largest crowdsourced AI-generated design evaluation platform. Its core mechanism doesn't rely on automated metrics but measures the quality of models through the preferences of real users. The evaluation method is similar to a "battle": the platform randomly shows responses from two different models, and users vote for the "better" one.

Different from benchmarks like MMLU and SWE-Bench, which rely on objective task completion, Design Arena is closer to the real user experience. Once a new model is launched, it can quickly accumulate a reputation score through this "human-voting battle" method.

On the Design Arena open-source list, the top 16 are all domestic large models. This means that in terms of user experience, Chinese open-source models have completely outperformed their overseas competitors.

Meanwhile, this list also provides an excellent window to observe the power of open source. Open-source models are catching up with the cutting edge at an astonishing speed: Open-source models that can run on consumer-grade GPUs are only about 9 months behind the most advanced closed-source models on average. And after the release of the o1 model, the open-source community caught up in just 4 months.

Netizens' comments also confirm the view of VC tycoon Martin Casado: In a competitive environment with limited capital, the spread of open source is not only a technological phenomenon but also an inevitable choice for startups.

User TradingPulse X wrote in the comments, "Chinese models offer better value for money than Western ones. If you're a startup, it's almost an obvious choice. Cash flow is the lifeline of a startup."

Another netizen, Dylan Bishop, exclaimed, "It's so amazing 👀." Rohan Paul replied, "Yes, it almost happened overnight. The entire startup ecosystem that relies on open source now depends so much on China."

Of course, there are also some skeptical netizens who think the Design Arena rankings are unreliable. But as netizen lordmostafak said, regardless of their origin, the open-source models that can truly solve problems will be the winners in the future AI ecosystem:

"Personally, I think the GPT-OSS model should have ranked higher, but there's no denying that the market is now dominated by Chinese models. I don't care where they come from - as long as they work. As the saying goes, 'It doesn't matter if a cat is black or white, as long as it catches mice, it's a good cat.' Interestingly, this saying itself comes from a Chinese person. It's kind of ironic!"

Reference Links:

https://www.reddit.com/r/LocalLLaMA/comments/1mz4hrg/all_of_the_top_15_os_models_on_design_arena_come/

https://x.com/rohanpaul_ai/status/1959710355208499692

https://www.youtube.com/watch?v=v0-7wlTdGck

This article is from the WeChat official account "InfoQ." Compiled by Tina. Republished by 36Kr with permission.