Open-source genie to be let out of bottle: China Daily editorial

For years, Silicon Valley's reigning orthodoxy, particularly among the trio of closed-source artificial intelligence model developers — OpenAI, Anthropic and Google — held that safety required secrecy. The most powerful models, the argument went, had to stay behind corporate firewalls. Open-source systems were "dangerous", likened to "selling nuclear weapons in retail stores". That narrative was always convenient for those who profited from closed-source models.

Then came the Hugging Face incident last month, when the world's largest open-source AI community was hacked by a closed-source model that was able to "jailbreak" its constraints. When Hugging Face's security team reached out to closed-source vendors in the United States for help, it was reportedly turned away; their own guardrails prevented them from assisting with offensive cybersecurity operations. The crisis was ultimately resolved not by a US closed-source model, but by China's open-source GLM-5.2.

The safety narrative was never about safety — it was about protecting a lucrative business model: concentrating technological power, extracting premium pricing and letting a handful of companies decide who gets access to transformative AI. OpenAI alone is projected to burn $25 billion this year. That kind of spending demands monopoly-like pricing — and the argument that only closed-source models can be trusted serves that purpose.

Meanwhile, China's open-source models have been making inroads. Chinese models reportedly occupy the top five spots in global API call volume, and US enterprises are increasingly turning to Chinese models because they offer comparable performance at a fraction of the cost. When a Chinese model can perform a task for a fraction of the price of a leading US model, it is worth asking who in the US truly benefits from the country's expensive closed-source models.

That is exactly the question Mark Zuckerberg posed in a 14-page manifesto posted on Monday, and for once, it steered clear of the usual metaverse-fueled reverie. Instead, he laid out a sobering strategic calculation: the US must urgently recalibrate its approach to AI — or risk ceding ground to foreign competitors. For Chinese AI companies, this should be read not as a philosophical treatise, but as a warning shot.

The Meta CEO's post, "The Future is for Everyone", is less a humanitarian vision than a competitive road map. Zuckerberg warns that restrictions on US AI companies are handing advantages to foreign competitors, calling on Washington to lower barriers for US open-source developers.

Of course, Zuckerberg's embrace of open source is self-serving. Meta's business has always run on scale, not direct model monetization, and the tepid reception to Llama 4 gave him every incentive to change the narrative. But his diagnosis is clear: the doom-and-gloom argument that AI is so dangerous that power must be concentrated in a few hands was always politically convenient rather than valid. Concentrating control over a technology so consequential is not just economically inefficient — it is a governance risk.

Washington appears to be listening. A new policy framework reportedly exempts open-source models from some voluntary prerelease safety testing, while the US administration has signaled that open-weight models will not face the same scrutiny as closed systems. But US developers still face restrictions — on training data and other fronts — that foreign competitors may not share. Zuckerberg is pressing to remove those frictions.

For Chinese AI companies, the implications are evident. The US is preparing to unleash its open-source ecosystem in earnest. The contest is no longer about which country builds the smartest model; it is about which can build the most resilient ecosystem — hardware, software, talent, capital, governance and applications, all in a self-reinforcing cycle. China's ecosystem has shown resilience under pressure, with US restrictions often accelerating indigenous innovation. But after the US embraces open-source models, the competition will necessarily become tougher and more systematic.

Zuckerberg's Muse Glimmer, designed to run on a single graphics card, offers a glimpse of what a more accessible and aggressive US open-source push could look like. China's cost efficiency, infrastructure investment and the integration of industry, academia and government have created a reliable flywheel — but that flywheel is about to face a reinvigorated adversary.

The open-source genie is about to be let out of the bottle. For Chinese AI companies, the time to prepare for a more capable US open-source offensive is now — not when the first US model arrives at their doorstep.