In the span of time it took for Washington to restrict, then hit undo on limitations to Anthropic PBC’s top models, Silicon Valley woke up to Zhipu (智譜).
Last year, when I spoke to Zhang Peng (張鵬), the chief executive officer of the Beijing-based start-up known abroad as Z.ai, it was still a little-known lab. What struck me was his focus on achieving artificial general intelligence, and his unapologetic defense of open-source artificial intelligence (AI) at a time when many were dismissing it as an admirable concept without a sustainable business model. The jury is still out on whether it is. However, one thing has become clear: Open models from China now pose one of the biggest existential threats to the US AI establishment.
The release of GLM 5.2, Zhipu’s most advanced model, coincided with Washington’s seemingly ad hoc regulatory intervention around Anthropic’s top Fable 5 and Mythos 5 models. The timing could not have been better.
Illustration: Mountain People
Suddenly, everyone seems to be migrating to GLM. Influential tech investor Marc Andreessen wrote that many insiders call it the first Chinese AI model to “match and often beat” US offerings “with no compromises.”
In a social media post, the CEO of Coinbase wrote of keeping spending flat while token usage “grows exponentially” by defaulting to open-weight Chinese models such as GLM 5.2 and Moonshot’s Kimi 2.7.
A former Meta and GoogleDeepmind staffer said that with GLM 5.2 “things are not going to be the same.”
When it really hit home for me was when my software engineer brother in San Francisco texted me he was switching over to it.
That so many are no longer hiding their excitement, or treating Chinese models as curiosities rather than substitutes, should make US AI executives deeply uncomfortable.
Real obstacles remain, including stigma, even in the era of “China-maxxing.” A massive global survey published last month found that while many acknowledge China’s AI leadership, it lagged far behind on “trust.” However, this cuts both ways. Open-weight models are inherently less opaque than proprietary US systems. They can be inspected, modified and hosted on a company’s own servers or even US data centers.
Jim Keller, the legendary chip designer who helped develop the early processors used in the iPhone, told me this week in Tokyo that his company, Tenstorrent, has switched to GLM and Kimi K2 for coding. The draw was about much more than pricing.
“I have very little influence on how Anthropic works, whereas Kimi K2 and GLM, I can fine-tune,” he said. “I have way more control over it, so we actually think we’ll get better results — that’s on my premises, that I can control.”
The shift is five times cheaper and “actually private.”
This GLM shock is hitting right as token economics and astronomical AI spending bills are becoming as important for enterprises as performance.
That is the lens through which the recent attacks on Chinese models must be understood.
It is no surprise that US incumbents would want to cast doubt by accusing them of illicit “distillation,” or training systems on the outputs of another. The outrage is hard to take at face value. For starters, distillation is a widely used technique in AI. Even Elon Musk acknowledged that xAI’s Grok was trained from OpenAI’s products.
More broadly, if the practice alone were enough to create world-class competitors, why has everyone else not done it? Then there is the irony that today’s leading AI companies complaining about intellectual property when their own products were built on the creative works of millions, often without permission or remuneration. It all makes the indignation a hard sell.
There are valid safety concerns when it comes to open models, including easier misuse. Some cybersecurity researchers have found that the latest Zhipu model was better than Anthropic’s Claude Code tool at “genuinely difficult security research,” while doing it at one-sixth of the cost. Washington might be tempted to respond with restrictions, which would be a mistake. It would not make the technology disappear; it would push it further outside of public scrutiny. As I have written before, sunlight is the best disinfectant.
It is understandably scary to envision an accessible Mythos-like model, “but if open models get banned now and only closed models get 10 or 100x better in two years in the hands of one or two companies, I think we will have a bigger problem on our hands,” AI researcher Nathan Lambert wrote.
The safety argument is less persuasive when it comes from companies preparing for major initial public offerings and defending closed ecosystems that give them extraordinary pricing leverage.
Where does this leave us? This is no longer a cost or philosophical soft-power competition with China; it is now a direct threat to the already shaky business models of AI companies that underpin a massive share of market optimism about the US economy. The US has an advantage in compute and now must build its own robust open-weight ecosystem. Invest in research funding, prepare talent to harden open systems against misuse, give enterprises more control over their AI tools.
Ultimately, the response cannot be a closed-door panic. Silicon Valley’s China shock is no longer coming, it is already here.
Catherine Thorbecke is a Bloomberg Opinion columnist covering Asia tech. Previously she was a tech reporter at CNN and ABC News. This column reflects the personal views of the author and does not necessarily reflect the opinion of the editorial board or Bloomberg LP and its owners.
The FBI on Sunday arrested Wanying “Heather” Chang (張婉瑩) on suspicion of surveiling the family of the son of President William Lai (賴清德) on behalf of China, an example of “transnational repression” to suppress dissent in Taiwan and abroad. Zhang, a 34-year old realtor in California, was arrested on charges of working as an unregistered agent for China. She and an accomplice were accused of taking photos and videos of Lai’s son and his family, and sending information about their whereabouts and license plates to Chinese officials. “The information could be used by China to gain leverage over the Taiwanese president
On Oct. 10, 1911, gunfire in Wuchang (武昌) shattered the stillness of late-Qing China. What began as a regional uprising quickly ignited a chain reaction across the empire, toppling a dynastic order that had endured for more than two millennia and ushering in one of Asia’s first republics — the Republic of China (ROC). From that moment, the 10th day of the 10th month became a date of epochal significance, not merely in Chinese history, but in the broader arc of Asia’s encounter with modernity. Yet today, as Taiwan marks Double Ten National Day, the holiday’s meaning has undergone a transformation
A few days ago, we left our breakfast midway to catch a glimpse of the fighter jets in formation. It was a spectacular sight. We knew it was part of practice drills for Double Ten National Day celebrations and reminded us of the Annual Republic day celebration in India. While channel surfing, I came across a news debate about immigration laws, where an academic asked a fellow panelist: “Why should we be interested in your point of view when you don’t even belong here?” That got me thinking. Was our enthusiasm about the upcoming Double Ten celebrations unwarranted? Was this
Taiwan might just give up. That’s the argument Simon Shuster made in The Atlantic’s October issue. The US is unreliable, the Chinese Nationalist Party (KMT) leadership is aligned with the Chinese Communist Party (CCP), and Taiwan does not have the military capability to defend itself in a prolonged war without help. What hope does Taiwan have? Quite a bit more, in fact, than Shuster lets on. Shuster points to genuine problems. The first Donald Trump administration had a mostly positive record on Taiwan policy. It regularized the arms sales process, enhanced the military-to-military relationship, and treated Taiwan more like a normal
RSS