How China split Silicon Valley over AI

(Edited and refined by Candice Chan, with the assistance of AI translation.)

Nvidia chief executive Jensen Huang published his first-ever post on social media platform X on 24 July, throwing his support behind open-weight artificial intelligence (AI) models. He also attached an open letter titled Open Weights and American AI Leadership, jointly signed by 25 American technology companies and organisations, including Meta, Microsoft and Palantir.

The open letter compares today’s debate over open-weight AI with the rise of open-source software in the 1980s. It argues that open weights expand access to the AI economy by allowing start-ups, established businesses, universities, and public institutions to build on advanced models without training one from scratch or paying frontier-model prices for every task.

The letter further argues that open weights will also strengthen competition, which is what keeps the gains of AI broadly shared rather than concentrated in a few hands, and give customers greater control so that they will not become locked into a single provider.

It also explicitly argues against restricting AI distillation, describing it as something that “reflects a long tradition of learning from, building upon, and improving existing technologies, a tradition that has helped drive innovation since the rise of the open-source software movement”.

Strikingly, America’s three leading proprietary AI companies — OpenAI, Anthropic and Google — were absent from the initial list of signatories. As more US companies joined the initiative over the weekend, OpenAI and Google subsequently signed the letter. Anthropic, however, whose proprietary models have benefited from the closed-model ecosystem, is still not on the list.

The game is afoot

On 27 July, Huang built on the momentum by bringing together more than 30 companies to establish the Open Secure AI Alliance, elevating Silicon Valley’s open-versus-closed AI debate from public advocacy to an organised industry initiative. This time, none of the three leading US AI companies was part of it.

The immediate catalyst for Silicon Valley’s increasingly fierce debate over AI development models has been the recent release of the Kimi K3 large language model by Chinese start-up Moonshot AI. Kimi K3 has achieved near-parity with America’s most advanced AI models in multiple benchmark tests, sending shockwaves through the technology industry. Moonshot AI also released Kimi K3’s model weights on 27 July as scheduled, allowing developers worldwide to download, fine-tune and deploy the model free of charge.

Since the emergence of DeepSeek, Chinese AI companies have consistently championed open-weight models, providing free access to capabilities for which leading American AI companies seek to charge. Coupled with the growing competitiveness demonstrated by Chinese AI firms, this has posed a direct challenge to an AI landscape previously defined by US companies.

A battle of prices

The first challenge concerns pricing power. Compared with the most advanced models developed by US AI laboratories, Chinese models are not only steadily narrowing the performance gap but also offer significantly better value for money. The price difference is stark: some Chinese AI models cost only around 1% as much to use as their leading US counterparts.

As a result, many American start-ups under pressure to control costs have begun adopting Chinese models at scale. According to OpenRouter’s statistics on token usage across major AI models, the share of usage accounted for by American models has fallen from 70% a year ago to around 30%, while Chinese models have risen to more than 45%.

After reports emerged that the Trump administration was considering banning Chinese AI models from entering the US market, the Little Tech Association — representing nearly 200 American start-ups — appealed to the government last week not to cut off access to Chinese models. One member of the association bluntly warned that if such restrictions were imposed, hundreds of companies would instantly die.

For America’s leading AI companies, competition from China’s open-weight models is reminiscent of the disruption Chinese electric vehicles, solar panels and lithium batteries have brought to overseas markets. It upends a previously highly profitable business model, not only significantly shrinking profit margins, but also affecting investors’ assessments of the competitiveness of America’s leading AI developers, as well as the valuations and future fundraising prospects of these companies.

Hugging Face and security concerns

The second challenge concerns the security narrative. Opponents of open-weight AI argue that once model weights are released, developers lose control over how they are used, creating significant security risks. However, recent reports that an OpenAI agent behaved unexpectedly during internal testing and attempted to compromise the AI development platform Hugging Face have weakened claims that proprietary AI ecosystems are inherently more secure.

The episode attracted even greater attention after Hugging Face unsuccessfully attempted to analyse attack logs using a leading commercial US AI model, and ultimately deployed Zhipu AI’s open-source GLM 5.2 locally to complete the digital forensics. The fact that a Chinese open-weight model resolved a security incident allegedly triggered by a proprietary American model has prompted the industry to rethink prevailing assumptions about AI security, particularly when the most advanced AI capabilities are concentrated in a handful of closed proprietary systems.

As Huang launched the Open Secure AI Alliance, Anthropic chief executive Dario Amodei finally broke his silence on 27 July, stressing that “Anthropic has never advocated for a ban on open-weights models”. Nevertheless, he singled out China, arguing that he was concerned about the risk that an “authoritarian government… builds AI models that are more powerful than those built by the US, and use them to achieve permanent military superiority or perpetrate incredibly deep repression of their own people”. He also repeated that the US should not sell powerful chips to China and should crack down on distillation.

A polarised Silicon Valley

Each side in Silicon Valley’s increasingly polarised AI debate has its own commercial interests. The cheaper AI models become, and the more open-weight models proliferate, the more companies are likely to deploy and train them, increasing demand for Nvidia’s hardware. By contrast, America’s three leading AI companies have little incentive to see an open ecosystem undermine the dominance of proprietary models. Instead, they hope to strengthen their competitive advantage by relying on regulation as a protective moat.

From a broader perspective, Silicon Valley’s growing divide over open and proprietary AI is no longer simply a technical disagreement. It has evolved into a structural contest over leadership of the AI industry, pricing power, and control over the global narrative surrounding AI. Stakeholders on both sides are expected to continue lobbying the Trump administration, leaving it caught between the competing interests of Silicon Valley’s two rival camps.

As China becomes the principal external factor shaping America’s pursuit of AI leadership, Silicon Valley’s internal divisions are increasingly becoming entangled with the broader US-China strategic competition over intellectual property, national security and export controls.

This article was first published in Lianhe Zaobao as “黄仁勋点燃硅谷AI大激辩”.

Related: Why young Americans are turning against AI | The new AI blocs: Are China’s WAICO and America’s Pax Silica splitting the tech world?