Onpode
Cover art for Big Tech warns US against restricting open-weight AI—but China's Kimi K3 is forcing the debate

Big Tech warns US against restricting open-weight AI—but China's Kimi K3 is forcing the debate

July 24, 2026 · 9 min

Juniper Vale & Mark Delaney

Moonshot AI's Kimi K3, a 2.8-trillion-parameter open-weight model from Beijing, benchmarks second overall in third-party tests and costs $3 per million input tokens — roughly half what Anthropic charges. Because its trained parameters are already publicly downloadable, the U.S. policy debate over restricting Chinese AI may be arriving too late to matter.

On July 16, 2026, Beijing-based Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model built on a sparse Mixture-of-Experts architecture. It activates only a fraction of its parameters per request, enabling lower inference costs than dense models of equivalent size.

0:009:13
Get the next episode on AI Model Competition Intensifying

Follow it free — new episodes land in your feed.

Or make your own — any topic, in minutes
About this episode

When Moonshot AI released Kimi K3 on July 16th, Nvidia's stock moved before most of the commentary did. Investors read the release — a 2.8-trillion-parameter open-weight model, benchmarking near the top of third-party tests, priced at three dollars per million input tokens — as a structural shift, not a product launch. This episode works through why that read might be right, and what it breaks in the current US policy debate. The episode digs into the Mixture-of-Experts architecture that keeps inference costs low despite the model's enormous total size, the White House allegation of 'large-scale, covert industrial distillation' from Anthropic's Fable model, and why independent experts were largely skeptical of that claim as the main driver by late July. It also examines the industry letter — signed by Meta, Microsoft, Nvidia, Hugging Face, and Mistral — opposing open-weight restrictions, and why the letter's silence on China is hard to read as accidental. The sharpest tension the episode surfaces: is the concern that China has built frontier AI capability, or that the world now has cheap access to it? Those are different problems that call for different responses, and the episode makes a real case that policymakers, industry signatories, and congressional investigators are currently running those two conversations as if they're one.

Frequently asked

What is Kimi K3 and who made it?

Kimi K3 is a 2.8-trillion-parameter open-weight AI model released by Moonshot AI, a Beijing startup backed by Alibaba, Tencent, and Meituan. It was released on July 16th, ranked second overall in third-party benchmarks, and placed first specifically in web interface building, competing directly with Anthropic's Claude Fable and OpenAI's GPT-5 variants.

How much does Kimi K3 cost compared to Claude and GPT?

Kimi K3 is priced at $3 per million input tokens. Anthropic's Opus 4.8 costs roughly double that per task. Because Kimi K3 is also open-weight — meaning anyone can download and run its trained parameters — developers can use or self-host near-frontier AI capability at significantly lower cost than leading US commercial models.

Did Moonshot AI copy Anthropic's Claude model to build Kimi K3?

White House science advisor Michael Kratsios alleged that Kimi K3 was built through 'large-scale, covert industrial distillation' of Anthropic's Fable model using chips not cleared for US export. However, independent experts assessing the model by around July 23rd largely concluded distillation was not the primary explanation. Moonshot AI issued no public statement about its training process.

Why did Nvidia's stock drop after Kimi K3 launched?

Nvidia's stock sold off after Kimi K3's release because investors recognized that a near-frontier open-weight model priced at $3 per million tokens collapses the compute cost that normally drives GPU demand. If developers can access frontier-level AI cheaply without buying expensive chips, fewer high-end Nvidia GPUs are needed to serve that demand.

Why did Meta, Microsoft, Nvidia, and Hugging Face oppose US restrictions on open-weight AI models?

Meta, Microsoft, Nvidia, Hugging Face, and Mistral co-signed a letter dated July 24th arguing that restricting open-weight AI models would be ineffective. The letter does not mention China once. Critics note the signatories have financial interests in open-weight proliferation, even as Nvidia simultaneously absorbed stock losses caused by cheap open-weight models threatening GPU demand.

Grounded in 12 sources
China's Diverse Open-Weight AI Ecosystem and Its Policy ... · hai.stanford.edu
The secret Trump administration battle to fight Chinese AI - Axios · axios.com
OpenAI and Anthropic find common ground: Open-weight AI - Axios · axios.com
Exclusive: Nvidia's Jensen Huang defends Chinese AI amid Kimi panic · axios.com
China's Moonshot AI claims Kimi K3 can rival OpenAI and ... · bbc.com
Moonshot’s Kimi K3 May Be More About Memory Than Compute - Bloomberg.com · bloomberg.com
Nvidia, Microsoft, Meta warn against 'premature restrictions' of open-weight models - CNBC · cnbc.com
Lawmakers probe growing use of Chinese AI models in U.S. companies · cnbc.com
What is China’s Kimi K3 and why is the US so rattled by it? - CNN · cnn.com
China’s Kimi K3 Hits US Stock Markets. Is the American AI Boom Over? · finance.yahoo.com
China’s Latest A.I. Breakthrough Threatens America’s Lead · nytimes.com
China’s Leader Pitches ‘Openness’ in Push to Shape A.I.’s Future · nytimes.com
Read transcript

Mark Delaney: Juniper, quick one before we get into it — did you see the Nvidia chart this week, or were you living a normal human life?

Juniper Vale: I saw it. And I immediately knew what caused it before I even checked.

Mark Delaney: Kimi K3. July 16th. Okay so — uh, let me just put the shape of this out there, because I think once you hear it, the stock move makes complete sense. Moonshot AI, which is a Beijing startup backed by Alibaba, Tencent, and Meituan, releases a 2.8-trillion-parameter model. Open-weight. Anyone can download the trained parameters, run them, modify them.

Juniper Vale: And it's not just big — it benchmarks second overall in third-party tests. First in web interface building specifically. They're comparing it directly to Anthropic's Claude Fable and OpenAI's GPT-5 variants.

Mark Delaney: At three dollars per million input tokens. Which is — wait, what does Anthropic charge for Opus 4.8?

Juniper Vale: Roughly double, per task. So when investors looked at that pricing on an open-weight model that benchmarks near the top — the implication for Nvidia is pretty plain. Fewer people need to buy the expensive chips if the compute cost collapses.

Mark Delaney: And Moonshot was reportedly in talks for a new funding round valuing them at thirty billion — up from twenty billion in May. So the market read this release as, like, a structural shift, not just a product launch.

Juniper Vale: That's exactly the question driving today — is this a technology story, a market story, or something else entirely? Because depending on which lens you use, the right response looks completely different.

Mark Delaney: But that structural shift framing — I keep wanting to say, didn't we already say that about DeepSeek? Like, wasn't that the same headline six months ago?

Juniper Vale: That's the right question, and the answer is actually — no, not quite. Think of it like this. Imagine a restaurant that serves a Michelin-star meal but only heats the specific ingredients on your plate. It's not firing up the whole kitchen every time you order. That's Mixture-of-Experts. Kimi K3 has 2.8 trillion parameters total, but only a fraction of them activate per request. So the inference cost stays low even though the model is enormous.

Mark Delaney: Wait — so the size number is sort of... misleading on its own?

Juniper Vale: It's the total capacity, not what you're paying to run. And that's actually the part that matters for the cost. DeepSeek was open-weight and cheap too, but Kimi K3 is explicitly showcased at the World Artificial Intelligence Conference in Shanghai — coding, reasoning, knowledge work — and it's benchmarking against Anthropic's Opus 4.8 directly. The scale and the open-weight combination together, that's what's sharper here.

Mark Delaney: So the headline is 'China matched US models' but you're saying the actual story is — what, cheaper?

Juniper Vale: The actual story is that a startup founder in São Paulo or Lagos can now afford frontier-level AI in their product. Three dollars per million input tokens, open-weight, download it, run it, modify it. That's not an access question anymore. It's done. The capability is already out there.

Mark Delaney: Huh. So it's not really an IP problem — it's an adoption problem that already happened.

Juniper Vale: Exactly that. And because it's open-weight — meaning the trained parameters are public, anyone can redistribute them — you can't un-ring that bell with a policy letter. Nvidia's stock moved because investors understood that before most of the coverage did.

Mark Delaney: The market clocked it faster than the think tanks. Yeah, that tracks.

Juniper Vale: But — and this is where I want to slow down — the policy response isn't tracking the adoption story. It's tracking a different story. The one where Michael Kratsios, the White House science advisor, goes out and says Moonshot AI built Kimi K3 by covertly distilling Anthropic's Fable model. Using chips not cleared for U.S. export to China. 'Large-scale, covert industrial distillation' — that's the phrase he used.

Mark Delaney: Wait — he named Anthropic specifically? Like, Fable specifically?

Juniper Vale: Fable specifically. And here's what makes it sticky — knowledge distillation is a real technique. A newer model learns by imitating the outputs of a more capable teacher model. So on its face, the allegation isn't technically absurd.

Mark Delaney: Right — but the part that doesn't fit is that independent experts, by around July 23rd, were largely saying it's... not the primary explanation. Like, uh, not 'this is impossible,' just — this doesn't account for what we're actually seeing in the model.

Juniper Vale: And Moonshot never responded. No public statement about their training process. So you've got an accusation, experts who mostly don't buy it as the main driver, and silence from the accused. That's the evidentiary foundation for restricting open-weight models.

Mark Delaney: No, I don't buy that. I mean — okay, wait, actually that's kind of the whole problem, right? Because either the distillation claim is the actual legal and moral foundation for the policy case, and if experts say it's shaky, the whole thing kind of... falls? Or it was always just cover for a bigger fear about Chinese capability generally, and we're talking past the real argument.

Juniper Vale: Those are genuinely two different conversations. And policymakers are treating them like they're the same one.

Mark Delaney: Which is going to matter a lot more once we get to who signed that industry letter and what they actually stood to gain — that part gets messier.

Juniper Vale: If the allegation is unverified, and the model is already open-weight and globally distributed — what exactly is the restriction supposed to stop?

Mark Delaney: That's — and that's where Xi Jinping just walks right through the open door. Because the same week this is all happening, he's at the World Artificial Intelligence Conference in Shanghai, calling AI a 'symphony of global collaboration.' Pitching openness as China's whole governance philosophy. With Kimi K3 on stage as the exhibit.

Juniper Vale: While U.S. officials are accusing Moonshot of IP theft.

Mark Delaney: That's not a contradiction — that's, uh, I mean that's just brilliant messaging. He's running the narrative inversion in real time. America used to be the open one. Now we're the side arguing about restricting access.

Juniper Vale: And Kimi K3 is open-weight, so the 'openness' framing isn't even a lie. It's selectively true in a way that's really hard to counter. You can't call it propaganda when a developer in Nairobi can literally download the model.

Mark Delaney: Hold on — so what does that actually look like on the ground?

Juniper Vale: Think of a compliance engineer at a mid-market fintech firm — she's building a document-parsing feature, pulls up Kimi K3's API on a Friday afternoon, three dollars per million tokens, it works better than what she was using. She ships it. She hasn't broken a law. The House Committee investigation probing Chinese AI adoption? Still underway. The Trump administration debate about restrictions? Still internal. By the time any policy lands, her team's product is already in production and her users don't know or care what model is underneath.

Mark Delaney: And that's — wait, that's the thing Meta and Microsoft and Nvidia and Hugging Face and Mistral are all technically correct about in that July 24th letter, right? Restrictions at this point are locking a barn after the horse has been running loose for two weeks.

Juniper Vale: They might be right on the merits. I'm not sure the motivation is clean, though. Nvidia's stock just sold off because cheap open-weight models threaten GPU demand — Jensen Huang's company co-signs a letter saying don't restrict this, while absorbing losses caused by the exact dynamic they're defending.

Mark Delaney: Yeah. The letter doesn't mention China once. Not once. That's — I mean, you don't do that by accident.

Juniper Vale: Six months from now — either the U.S. has sanctioned Chinese open-weight models as a national security threat, or they've settled for export friction, some tariff-adjacent friction at the edges, because the industry won that argument. Those feel like the only two paths. And neither one actually answers the question underneath.

Mark Delaney: Which is — is the threat that China has the capability? Or is it that the world has access to it cheaply now? Because, uh, those are not the same problem. Like, not even close. One you maybe stop with sanctions. The other one... I don't know what policy even touches that.

Juniper Vale: Yeah. That's it. That's the actual fork. And I don't think anyone in the room — not the House Committee, not the Trump administration debate, not Nvidia or Meta or Microsoft signing that letter — I don't think anyone's answered it yet.

Mark Delaney: Mm. No. I don't think they have either.

Juniper Vale: I appreciate you sitting with the messy version of this instead of reaching for a clean ending.

Big Tech warns US against restricting open-weight AI—but China's Kimi K3 is forcing the debate · Onpode