The Open AI Revolution Isn't Coming From Silicon Valley
China's open source AI models are closing the gap with frontier US labs faster than predicted, triggering panic, lobbying, and trade battles in Washington.
Okay so I’ve had a version of this half written for like a week now, kept putting off finishing it because the story kept changing under me. Something new broke again this weekend actually, so whatever, let’s just go with what’s confirmed right now and I’ll probably have to come back to this in a month. Fair warning too, this one took way longer to pin down than usual, half my tabs were contradicting the other half on dates and chip names, I actually gave up and restarted my notes twice.
Quick setup if you haven’t been following. China’s open source AI models have gone from cute niche thing to actually closing the gap with the big US labs, and it happened faster than pretty much anyone predicted. DeepSeek’s R1 back in January 2025 was the first real gut punch, outperformed several prominent US models on independent benchmarks, did it for a fraction of the training cost. That’s the part that actually spooked people, not that a Chinese lab made a good model, that they made a good model cheap. It’s kept going since. Alibaba’s Qwen family passed Meta’s Llama for global open source download share earlier this year. Chinese open source models went from basically nothing on OpenRouter in late 2024 to something like 30% of all usage on the platform by early this year. This month Moonshot AI dropped Kimi K3, a 2.8 trillion parameter model with a million token context window, and Zhipu’s GLM-5.2 launched under an actual MIT license, reportedly rivals OpenAI and Anthropic on coding and agent tasks according to Reuters, at a fraction of the cost to run.
The why behind all this is kind of boring honestly. Cheaper training methods, distribution across every hardware tier so more people can actually run the things, constant shipping instead of one big release a year.
At the World AI Conference this year, Xi Jinping stood up and offered five thousand AI training placements, joint regional centers, tool access, specifically framed as aid to developing economies. China’s bundling free open AI models into its foreign policy. So on one side you’ve got a state most of the west calls authoritarian, giving away frontier-adjacent AI tools to poorer countries with no license fee attached. On the other, American AI companies that talk constantly about openness and democratizing access, while their actual products sit behind subscription paywalls a lot of the developing world can’t touch.
I don’t think that means China’s the good guy here, plenty to be skeptical of in a government-adjacent AI strategy, doesn’t take much imagination to guess why a state might want its software running in as many countries as possible. But it’s worth sitting with the fact that the free and open thing is currently coming from the direction people didn’t expect, and the closed and paywalled thing is coming from companies headquartered in countries that talk the most about freedom and openness.
This past weekend the New York Times reported OpenAI and Anthropic have been lobbying Washington, privately, to restrict Chinese open-weight models, citing national security risk and something called distillation, the accusation that Chinese labs are training partly off outputs from American systems. Publicly, Sam Altman’s been out signing letters in favor of open weight AI as a category, OpenAI actually joined an industry letter defending open-weight models last Friday, the same week they were reportedly also privately pushing regulators to restrict the Chinese ones specifically. Anthropic sat that letter out entirely, while apparently also lobbying against Chinese open models behind closed doors. Treasury Secretary Bessent’s quote on this was “open source is not open season on American IP,” which is a good line, I’ll give him that, though it’s also doing a lot of work reframing “we’re losing the open model race” into “they’re stealing from us.”
The coalition pushing back against that private lobbying is a weird alliance if you look at who’s actually on it. Nvidia, Microsoft, Meta, Google, Hugging Face, over 200 startups signed the pro-openness letter, argued restricting open weight models just concentrates power in a handful of frontier labs. David Sacks, Trump administration adviser, called the whole thing regulatory capture dressed up as safety concern. When a Trump appointee and 200 startups land on the same side of an argument, that’s usually worth a second look.
There’s a separate mess with the actual hardware, and this is where the monopoly question gets a real answer instead of a vibe. Nvidia’s chip exports to China have been through several complete flip-flops in about twelve months, banned, unbanned, re-restricted, partially unbanned again with a 25% tariff attached, and at one point Trump personally said the previously-restricted chip was fine to sell because it was “obsolete” anyway, a strange thing to say about a chip everyone’s fighting over. Jensen Huang’s been personally lobbying the administration for months, trips to the White House, trips to Beijing, every inch of access he wins back a new tariff or condition shows up to claw some of it away. Congress, separately, has a bipartisan group pushing for a full blanket ban on semiconductor manufacturing equipment exports to all of China, which would be a much harder line than anything currently in effect.
Here’s the thing about Nvidia specifically though, the company you’d assume benefits most from all this, they’re actually getting squeezed from both directions. US policy keeps changing every few months so Chinese buyers can’t plan around Nvidia long-term even when chips are technically legal, orders don’t materialize the way they would under stable policy. On the China side, Beijing’s own government told domestic AI companies not to even buy the H20 chips when they were cleared for export, security concerns of their own, and separately China’s antitrust regulator opened a case against Nvidia claiming they violated terms from a 2020 review. Nvidia’s stuck between two governments that don’t trust it for opposite reasons, not exactly the position of a company winning a trade war.
I keep going back and forth on how to actually wrap this up, honestly, because there’s real evidence of panic here, not just vibes, two of the biggest AI labs in the US publicly championing openness while privately lobbying to restrict the specific open models actually competitive with them. About as clean an example of only liking open source when you’re winning it as you’re gonna find. And there’s real evidence the US is using export controls and tariffs as a blunt instrument that ends up hurting its own chip industry’s access to a massive market, while China’s own government throws up its own obstacles right back, which makes the whole who’s-the-monopolist question messier than either side’s talking points.
What isn’t settled, and I don’t think anyone’s actually answered this convincingly yet, is whether restricting Chinese open models protects US national security in any real sense, or just protects OpenAI and Anthropic’s subscription revenue dressed up in security language. Not the same goal even if the resulting policy looks identical on paper.
US officials are apparently still debating case-by-case restrictions versus a blanket approach and haven’t landed anywhere yet, so this one’s not done. I’ll probably have to redo half these numbers by the time I actually get around to a follow up, given how this week went.