Has China Gone All In on Open-Source, Open-Weight, and Why?
Price advantage and start of a new norm?
Hi all, I was so honored to be invited to join hosts Kevin and Aleks on BBC4’s show Artificial Human today (new series to be aired in June). We talked about our usual topics- China’s AI strategy, China’s embrace of AI, technical optimism, and the rise of DeepSeek, but one thing that caught me thinking when Kevin asked me the simple question of “what is open source vs open weight and why has DeepSeek embraced this model?” After the recording session, I couldn’t stop thinking about whether this “embrace” has resulted in results and whether it was the right move. So, I thought I’d better look into this trend and whether the Chinese AI ecosystem has indeed been leading in embracing open-source/ open weight compared to other major global players.
Before we dive in, at a high level, I want to quickly sum up a few key themes/ topics we’ve written about, and I was pleasantly surprised to see that it all largely aligns with Morgan Stanley’s findings from its May 13th China AI Blue Paper, which I will reference for this write up. (If anything, I think we’re ahead of the curve ;p)
Physical AI: China is projected to hold ~30 % of the global humanoid installed base by 2050 (302 m units), with commercial sales >50 million per by then, according to MS, and we’re seeing areas of continued AI integration. [See Electric Dreams, Robot Reality: China's EV-to-AI Evolution]
Industrial robots + AI = cross-disciplinary fusion, or what I’ve been calling a convergence of technology moment. There are over 1.7 million industrial robots in Chinese factories (51 % of global demand). Couple that with AI software. Think of the future of manufacturing and labor. [See Rise of China's Robotics Industry: from Manufacturing Arms to Embodied AI]
Super apps powering the next wave of Agentic AI: large proprietary data sets and a large base of DAUs set consumer internet big tech up for catching this opportunity. [See China's AI Application Super Powered: Open-Source Momentum Meets Super-App DNA]
Proprietary data strengthens multimodals: Google (YouTube), Alibaba (Youku), Kuaishou, Meitu, and Tencent (Tencent Video) lead the way in images, videos, and text models. We’ve examined Bytedance and Kuaishou before, which generate around 80 million new videos per day and have over a billion users combined daily. [See The Ghibli Hype: Next Frontier is in Video & Introducing China’s Leading AI Video App, Kling] & [Why Tencent's Integration of DeepSeek Into its AI App is a BIG DEAL]
Inference- price deflation continues, and model commoditization sets up the space for better consumer-facing use-case adoption and a free-to-use business model. Monetization on the consumer end will come from ad sales, value-added services, and new innovative ways. [The Great AI Power Shift - From AI Models to AI Applications
Open-source tilt and model quality parity: China supplies 40 % of the top-10 open-weight models on HuggingFace, with Qwen-72B on top; DeepSeek-R1 & peers close the quality gap on Chatbot-Arena scores. [See this article]
China continues to push forward green energy solutions to power the AI boom. AI tasks draw up to 33× the energy of task-specific software, and China already has 246 green DCs and targets 100 % renewable DC power by the early 2030s. (This, let’s see, but there are notable efforts) [AI Arms Race Far From Over: Chips is Only Half the Game, Infrastructure is the Other]
But like I said, I want to dive deeper into the open-source/ open weight topic today. I wrote about DeepSeek’s open-source week in March and the personal and mission drivers that led CEO Liang Weng Feng to choose to open-source DeepSeek here, FYI.
Closed vs Open
Morgan Stanley wrote that China’s open-source LLM ecosystem is gaining significant traction with the launch of DeepSeek, but we didn’t need the investment bank to tell us that. Hugging Face has ranked several iterations of Chinese AI startups’ open-source models highly for common sense reasoning, math, coding, and reading ability. And not just the startups ranking highly, but also the big tech models. Chinese companies currently own four of Hugging Face’s top-ten open-weight slots, with Alibaba’s Qwen-2.5-72 B-Instruct leading at number one.
In fact, in Morgan Stanley’s report - Exhibit 42, four out of ten red bars are of Chinese AI labs. Compare that to a snapshot from late 2023, where none were Chinese. China’s open-source models are catching up and leading the way.
No One Wants To Pay
According to Morgan Stanley’s research, as the graph above shows, 83% of enterprises prefer open-source models for the flexibility to tweak and run them privately. And only 21% say they would pay a premium for closed-model software. Budgets are moving accordingly: hardware’s share of next-year AI spend rises to 34%, software slips to 40%. So if you sell SaaS wrappers in China, prepare for thinner margins, or you may need to bundle unique data or tools that users can’t replicate behind the firewall.
Before, I thought only consumers would put free-to-use as top priority, but shockingly, AlphaWise China CIO Survey found that enterprises are found to prefer open-source models - and not paying for the (premium) models - as well.
And this may challenge what we’ve previously written about, that Chinese AI wrapper companies were looking to innovate within vertical industry use-cases. So, if even enterprises don’t want to pay (or pay little for APIs) for AI wrappers and instead make tweaks and adjustments based on customized needs on top of open-source models. These vertical AI applications may need to reconsider their monetization strategy or offer something SO ready to use.
Thus, price remains an advantage for many of these Chinese open-source/open-weight models, leading by example, is DeepSeek, which has its pricing undercutting every Western frontier model today ( ~ USD 0.55/2.19), and its API price is expected to continue to drop. Cheaper inference and cheap tokens will collapse the moat that closed weights have enjoyed until now. The growth of users and adopters will likely come from value-added services and sheer volume of users, which, in a way, will make even the enterprise playbook look more and more similar to the consumer-facing strategy.
And will open-source become the new norm? Venture Capitalist Marc Andreessen said that American AI labs also need to seriously consider open-source AI or “risk ceding control to China.” And as Business Insider wrote, as hardware bifurcation deepens, more attention is being shifted to software. Andreessen added that it is "plausible" and "entirely feasible" that open-source AI could become the global standard.