For years, the AI narrative in Silicon Valley has revolved around monolithic, proprietary models: comparing typically GPT-5 with Claude 4.5 or Gemini 3 around the industry benchmarks. The prize of being even part of the race is huge: it’s the chance to build the next OpenAI, the next Google, or the next Meta, valued in the hundreds of billions. To this day, it seems, the ultimate focus is to reach “superintelligence” or AGI (Artificial General Intelligence) through massive, closed frontier models.
But whispers in the VC community have just turned into a shout: The future of practical AI adoption is being built on something far more streamlined, cost-effective, and critically, open-source models, and it’s coming from China.
A study by the Massachusetts Institute of Technology and open-source AI start-up Hugging Face found that the total share of downloads of new Chinese-made open models rose to 17 per cent in the past year. The figure surpasses the 15.8 per cent share of downloads from American developers such as Google, Meta and OpenAI — the first time Chinese groups have beaten their American counterparts.
Martin Casado, a partner at Andreessen Horowitz (a16z), recently revealed in The Economist a startling reality: up to 80% of the AI startups pitches to the firm are US startups using Chinese open-source models. This dramatic shift is not a side trend; it’s the new operating principle for cash-conscious innovators worldwide.
How Startups Are Using Open Source AI Today (and Why)
For early-stage companies, the choice to adopt open-source models, also often “open-weight” models, which provide numerical parameters rather than the full training source code, is driven by survival and efficiency:
- Cost-Effectiveness: Chinese models offer “better value for money than Western ones,” making them an “obvious choice” when cash flow is the lifeline of a startup. Open source offers a streamlined, cost-effective model built on free AI.
- Adaptation and Customization: Open models spur greater adoption by allowing companies, governments, and researchers to more easily adapt them to the “nooks and crannies” of specific use cases. This enables hands-on experimentation, fine-tuning, and deployment for regulated industries or those needing data sovereignty.
- On-Premises Deployment: Open-source models help users run their AI tools directly on premises rather than relying solely on proprietary cloud services.
- Accelerated Advancement: By making architectures and weights accessible, the broader community can examine and improve upon the systems, accelerating AI advancement and preventing key knowledge from being siloed within one high-resourced group.
This reliance on cost-effective alternatives means that the entire startup ecosystem that relies on open source now depends heavily on China.
The Video Generation Battleground
The shift is visible across model types. China has particularly leaned heavily into developing AI video-generation models. According to Hugging Face data for 2025, while the US is the primary developer of most AI model categories, China dominates the share of downloads for Video Generation models.
This dominance highlights a key area where Western open models are playing catch-up. The leading models in the West are OpenAI’s Sora 2 and Google’s Veo3, with Israel’s LTX-2 (by Lightricks) being the only open-source, open-weights model. But the Chinese models on the other hand, like Qwen3, Kling and Wan, are all accessible via platforms like Fal.ai and are gaining popularity, partly due to the high cost of the Western models.
The Chinese Open-Weight Advantage
When Deepseek first announced that its open-source model, trained for an alleged $5M on less advanced chips, is performing at the level of ChatGPT-3 for a fraction of the cost, Nvidia lost $65 billion in market cap in a single day. Since its controversial debut, Deepseek has gone viral.
While US giants like OpenAI and Google have maintained full control of their most advanced technology, profiting through subscriptions and enterprise deals, Chinese groups have been offering wider access to their models. This strategy appears to be a major success:
- Market Dominance: China has quietly upstaged America in the open-model domain. A study by MIT and Hugging Face showed that the total share of downloads of new Chinese-made open models rose to 17% in the past year, surpassing the 15.8% share held by American developers (Google, Meta, OpenAI).
- Performance Edge: Chinese models released this year have outperformed similar open US models, like those from Meta, in intelligence tests. On the Design Arena rankings, the world’s largest crowdsourced evaluation platform, the top 16 open-source AI models are all from China, signalling they have completely outperformed their overseas competitors in user experience. DeepSeek, Zhipu, Kimi, and Qwen are among the key teams dominating this leaderboard.
- Pace and Innovation: Cut off from advanced Nvidia chips due to US export controls, Chinese labs have focused on being more creative in their development approach. They are shipping models on a weekly or biweekly basis, offering many variations, which is a “paradigm shifting” pace compared to the semi-annual or annual releases from US labs.
Will we see more open source models in the west?
The spread of Chinese open-source models is forcing a reckoning. While American labs are betting on high-valuation, proprietary models, their Chinese rivals are focused on driving mass adoption through openness.
As Martin Casado emphasises, the right approach for the West is not to “close off” but to“further promote our own open-source efforts”. Historically, open-source software has accounted for about 20% of the market value, but in the AI field, this proportion is much higher, suggesting a healthier open-source ecosystem than in the traditional software era. For venture capital, understanding which startups are leveraging these free, high-performing models, and how they plan to monetise ancillary services like customisation and support—is essential to navigating the post-DeepSeek shock era. The rise of Chinese open AI proves that it doesn’t matter if a cat is black or white, as long as it catches mice, it’s a good cat.
I'm a former general partner at google ventures, head of Google for Entrepreneurs in Europe and founding head of Campus London, Google's first physical hub for startups.
I'm also the founder of Techbikers, a non-profit bringing together the startup ecosystem on cycling challenges in support of Room to Read. Since inception in 2012 we've built 11 schools and 50 libraries in the developing world.
- In 2026 Gaming companies must deal with AI supercycle and VC winter to find new moats - March 12, 2026
- The Researcher’s New Job Is Writing the Spec - March 10, 2026
- Weekly Firgun Newsletter – March 6 2026 - March 6, 2026
