Lila Soto: Hugo, I've had a weird week — I keep refreshing the same news story, which almost never happens to me.
Hugo Vance: The Huang interview.
Lila Soto: How did you— yeah, the Huang interview. July 22nd, Mike Allen, Axios. Jensen Huang calls Kimi K3 excellent — that's the word he used — and tells Washington not to ban it. And I mean, Nvidia's stock had already fallen more than twenty percent from June peaks because of that exact model. He's defending the thing that just cost him.
Hugo Vance: What I'd want to know is whether that reads as conviction or as theater.
Lila Soto: That's the question — and here's the fact that I think sharpens it. Kimi K3, Moonshot AI's model, ranked first in blind developer testing. Above GPT-5.6 Sol. Take the brand names off and developers just... chose it. So Huang also posts on X — open models strengthen safety, accelerate innovation, enable sovereignty — and I'm wondering what's underneath that. Is he reading something the market missed entirely, or is the CEO of the world's dominant chip company just... performing calm?
Hugo Vance: You see, the 2.8 trillion parameters matter here. That's not a model someone releases to signal virtue. That is frontier scale. Which makes Huang's calculation more interesting — and more suspect.
Lila Soto: More suspect how?
Hugo Vance: Because the CEO of the dominant infrastructure supplier arguing publicly against restricting a competitor's access to that infrastructure — that is not a neutral position. That is a position with a very direct line to Nvidia's revenue.
Lila Soto: But that's almost too clean, right? Like — the conflict is obvious, so maybe everyone's already discounting it. What if his read is correct despite the interest?
Hugo Vance: That's where I'd pump the brakes. Think of it this way. Imagine the owner of the world's largest highway system tells you that more cars on the road — even cheap ones built overseas — is good for business. He's not wrong about the logic. But he also owns every toll booth. That is Jensen Huang. Nvidia profits whether it's an American model or a Chinese one running on its chips. The revenue case is identical either way.
Lila Soto: The toll booth thing — yeah, that lands.
Hugo Vance: And then look at who signed the joint industry letter defending open-weight models. Nvidia. Microsoft. Meta. Mistral. That is a coalition with a financial interest in adoption expanding. OpenAI and Anthropic — conspicuously absent. They went the other direction, lobbying for restrictions on Chinese AI. Which means Nvidia's own largest customers are formally on the opposite side of their chip supplier.
Lila Soto: Wait — Anthropic, whose model Kimi K3 allegedly distilled from, didn't sign? And they're lobbying against Huang's position?
Hugo Vance: Correct. Michael Kratsios at the White House Office of Science and Technology Policy accused Moonshot of building Kimi K3 through distillation of Anthropic's model — Claude Fable 5 specifically. Scott Bessent, Treasury Secretary, floated sanctions over exactly that. And Anthropic's response was not to defend Moonshot. So you have Huang defending the company that allegedly extracted from his customer's model. That is not a neutral analysis. That is a revenue argument wearing a policy coat.
Lila Soto: And the market already called it — the twenty-plus percent drop from June peaks is basically the crowd voting against the rising-tide thesis in real time.
Hugo Vance: Yes. Which is the problem for anyone trying to use Huang's argument as policy guidance. You literally cannot separate the market analysis from the revenue incentive. Not because he's lying — he may well be right — but because the structure of his position makes it impossible to know.
Lila Soto: But that structure — the revenue argument wearing a policy coat — actually gets more tangled once you pull on the distillation thread. Because okay, imagine a startup founder in Austin. Her whole product runs on Anthropic's API. Expensive, but she trusts it, her team knows it. Someone Slacks her Friday afternoon: Kimi K3 benchmarks higher in blind testing and it's free. She's excited. But wait — did that model get its capabilities by extracting Claude Fable 5's outputs without permission? Because that's exactly what Michael Kratsios said.
Hugo Vance: The White House OSTP director. Named Moonshot AI specifically.
Lila Soto: Right — and Claude Fable 5 ranked above Kimi K3 on the Artificial Analysis Intelligence Index. It's the alleged source. So the thing that beat your expensive proprietary model in blind testing may have learned how to do that by quietly ingesting your expensive proprietary model's outputs. That's not just a legal question for Moonshot. That scrambles what 'open' even means.
Hugo Vance: Openness built on a contested foundation.
Lila Soto: And Scott Bessent floated sanctions over exactly this — distillation from U.S. models as an actionable offense. So Huang's framing, that open models are just... good for everyone, it kind of assumes the 'open' is clean. If the capabilities were extracted from Anthropic without permission, then Huang is defending openness that was achieved by taking from one of his own customers. I can't move past that.
Hugo Vance: Well, and here's where I'd give the hot take a partial win, actually. The performance is real. Kimi K3 tested first blind. That is not a fabrication. Whatever the method — contested or not — the model works. The Austin founder's excitement is empirically grounded.
Lila Soto: Mm — the capabilities are genuine, the provenance is the problem.
Hugo Vance: Precisely. And that distinction is what makes Huang's 'open and excellent' framing unstable. Excellent — I'll grant him. Open — that depends on whether the method survives a Bessent sanction. Those are not the same claim.
Lila Soto: And honestly, the policy knot gets tighter from here — because the U.S. position on chip export controls and the position on restricting AI models may actually be pulling in opposite directions. We'll get into that contradiction in a minute, because it doesn't resolve cleanly.
Hugo Vance: That contradiction — the chip controls pulling one way, the usage restrictions pulling the other — that's where I'd plant the flag. Washington is simultaneously running export controls to deny China access to advanced Nvidia chips, and debating whether to restrict Americans from using Chinese AI outputs. Those are not the same policy. They may actually be opposing policies.
Lila Soto: Wait — how are they opposing?
Hugo Vance: Well. If the open-weight model is already released — Kimi K3 is already out — restricting Americans from using it does nothing to slow Moonshot AI. They've already shipped it. What it does slow is American developers building on cheap infrastructure. Which, I mean, paradoxically reduces demand for Nvidia chips and U.S. data centers. You've contained your own builders.
Lila Soto: So the usage restriction actually undermines the export control logic.
Hugo Vance: Unless the goal isn't containing Chinese AI capability at all. If the goal is protecting OpenAI and Anthropic's market dominance — that is a different policy objective than national security. And it requires acknowledging that their lobbying for restrictions reflects competitive self-interest just as much as Huang's lobbying against reflects his.
Lila Soto: Nobody in the room is clean. Huang, Anthropic, OpenAI — they all have a stake in the outcome they're arguing for.
Hugo Vance: Kratsios and Bessent frame it as IP theft and national security. But Nvidia, Microsoft, Meta, Mistral — a European company co-signing a U.S. industry letter, which is itself remarkable — they reject that framing entirely. The conflict of interest problem runs in every direction. That's the calibrated take. Not that Huang is wrong. Not that Anthropic is lying. It's that no one in this debate is arguing from a position without a direct stake.
Lila Soto: The honest verdict is that the policy knot exists because the interests created it — not the other way around.
Hugo Vance: And that's what keeps happening, yes. You started this week refreshing the same story — Huang calls Kimi K3 excellent, defends the thing that cost him. We've turned it over every direction. And the shape hasn't changed. Jensen Huang is simultaneously the most credible market analyst of global AI demand and the executive with the most direct financial stake in the answer being 'yes, buy more chips.' Those two things will not separate. Not now, not after a Senate hearing, not ever.
Lila Soto: And I guess — okay, I'll half-concede it. He might be reading the demand curve correctly. Also he profits if he's reading it correctly. Both true, forever, no resolution. What I'm sitting with is the other half — whether 'open' and 'closed' are even the right categories anymore. If distillation is how frontier capability spreads now, the whole vocabulary Washington is writing policy in might already be obsolete.
Hugo Vance: Well. Washington is trying to write policy about a model that's already released, built on a model that's already deployed, using a legal category — IP theft — that courts haven't defined for AI yet. The question of whether Kimi K3 is excellent innovation or excellent theft may be the same question. And nobody with a financial stake in the answer is the right person to ask.
Lila Soto: Which is... every person currently in the room.
Hugo Vance: Every single one. Good week to keep refreshing.
