China Isn’t Catching Up in AI. It’s Breaking Away.
By Veer Solanki · · 1343 words
Topics: AI, AI Agents, AI Chip Restrictions, AI Competition, AI Export Controls, AI Geopolitics, AI Models, AI Policy
America’s approach to Chinese AI for years was relatively simple: Keep the best chips out, deny access to crucial technology, and keep piling more companies onto entity lists if necessary. Over time, the gap should just become too wide. Except now, China appears to be doing something else altogether. This week alone, Beijing has blocked Meta’s purchase of the Chinese AI agent startup Manus, causing the deal to collapse.
After its introduction, Manus – whose autonomous agents we all struggled to avoid for much of the year, thanks in large part to its ability to clearly demonstrate what the future of agents looks like beyond demos – quickly established itself as the sort of thing no tech giant could miss.
Meta, apparently, did not. Soon after, the restrictions spread to offshore technology transfers. This last part is particularly important. Historically, many of the limitations in this AI fight have come from the US, not China; Washington has blocked chip sales, restricted exports and sought to keep frontier tech from Chinese labs.
China, by banning the acquisition of a leading domestic AI company, seems to be doing the inverse: the problem is not just what can come in, but also what can go out.
And then there is Alibaba. On Thursday, employees were apparently notified that Claude Code would be added to a list of high-risk software products, and its use within the office would be prohibited from July 10th. This is a major step given how useful Claude Code has proven even for those of us not invested in the political disputes of Anthropic and instead only in a great coding tool. The cited reason is tied to the odd security issue that roiled social media earlier this week: the accusation by some researchers that Anthropic had embedded code in Claude Code capable of detecting whether someone was physically in China or affiliated with Chinese AI labs. This claim took off on Reddit and GitHub before the expected backlash ensued in Chinese tech forums.
Now, take anything with a grain of salt, because security research spreads online rapidly and explanations don’t always stick, but Alibaba’s decision is undeniably real and, it has the potential to create a significant new hurdle for Anthropic. Many Chinese developers had apparently continued using Claude Code regardless of politics or available local models – good code is good code, ultimately. But with fears of spying running through local institutions, that calculation can change dramatically.
And all the while, Z.ai, the Beijing AI company, launched GLM 5.2. The timing is a bit too good.
The US momentarily put a ban on non-US individuals purchasing the company’s latest models (Fable 5, Mythos) on Wednesday. (The restrictions were rescinded on June 30, and it was a relatively brief affair.) I covered the specifics of Anthropic’s plight separately, given the brevity of the incident. But the window of opportunity, at least for some Chinese researchers, was there.
According to Z.ai, GLM 5.2 comes surprisingly close to the likes of Claude’s best model (Opus 4.8) and GPT 5.5 when engaging in tasks that go beyond mere benchmarks.
In multi-day projects that demand significant coding, the company estimates it comes within one percentage point of Opus and performs better than GPT 5.5 and older iterations of Opus (Opus 4.7) on various tasks. I’ll throw out that Z.ai is promoting its own metrics, and I’d be a bit too trusting to take anything beyond one percentage point at face value. But that’s not the most interesting part to me.
What I find more compelling about GLM 5.2 isn’t whether it’s the absolute best model in the world. Let’s assume Claude remains a bit stronger. Fine. If Claude is not readily available, restricted to a few institutions, and geographically blocked, that top one-point advantage loses some of its appeal.
Z.ai offers GLM as an open-source solution with an unusually long context window and has been openly advocating for borderless access. This doesn’t seem coincidental.
We saw something similar earlier in the year. DeepSeek R1 was released in January 2025 and caught a lot of the community by surprise – a domestic Chinese AI company, facing these restrictions, had managed to release a cheap and open-source reasoning model. At the time, R1 was the shocking thing. Eighteen months later, maybe the more relevant question is whether it gave everyone else the blueprints for what comes next. Money is following the same trend.
Alibaba, Tencent and Baidu all participated in a round of more than $2.8 billion in funding for Kling AI, Kuaishou’s video generation subsidiary.
For firms that are fierce rivals on almost every front of China’s tech landscape, all three names in one funding round seem rather peculiar. Kling now boasts a valuation of $15 billion before this latest funding, and counts the likes of BlueFive Capital from Abu Dhabi among the more than three dozen other investors. The company expects an IPO on the Hong Kong Stock Exchange within a year. There is, of course, a business case to be made here – Kling’s revenue in the most recent quarter has quadrupled from a year earlier, and the company’s video output is one of the very few in the world that can compete with OpenAI’s Sora.
But it is back to Alibaba, Tencent, and Baidu for me. Three rival companies all jumping in to invest in one domestic AI startup does not suggest the industry has suddenly gone all lovey-dovey.
Rather, it shows they know where the demand is going and do not want to miss out.
Even so, in the context of all the other news, it creates a different sort of picture – as local capital flows toward domestic AI products, the foreign variety faces growing trust issues and access hurdles. Let’s take all the pieces of this puzzle.
A US company tries to buy a significant Chinese agent startup, and Beijing rejects the offer. An American AI’s domestic use gets put into question, and a popular Chinese agent becomes restricted from use. An open-source coding tool goes live from China, as the availability of US counterparts becomes unstable. Three giants of China’s tech scene coalesce to put more than $2.8 billion into one domestic video AI company.
I’m not convinced this marks the moment China surpassed the US in AI; that would be too easy a headline, and frankly, probably an inaccurate one.
Many parts of American AI still lead, and Chinese companies are still subject to formidable hardware constraints. But the dynamic feels changed. For years, the story was the restriction, not the innovation.
China has been mostly playing catch-up, building tools that often felt reactive. But what we’ve seen in recent days points to something new. Companies are defending their burgeoning talent, fostering the development of local alternatives, and channelling significant capital into established domestic businesses with proven user bases and revenue streams.
Some of that activity was bound to happen, regardless of external pressure, but I suspect much of it was accelerated by it.
This leaves the US with a complicated position. Every move made to restrict a Chinese company can inadvertently boost another one. Blocks on American software can strengthen local alternatives. Uncertainty around foreign tech pushes companies toward the seemingly more reliable open source options.
By cordoning off a market, eventually those operating within it become content to forgo the old technologies altogether.
The Alibaba story, perhaps more than anything else we saw this week, illustrates this phenomenon best. Claude Code did not become functionally inferior on Thursday. People simply became less convinced that they wanted to continue using it.
The race in AI has for a long time been largely one of pure performance metrics – who has the smartest model, whose model codes the best, who spends the most money on compute.
China may still be behind in parts of the AI race. I’m not sure “catching up” describes what is happening anymore, though.
It looks increasingly like China is building a race where American companies don’t automatically get to set the rules.