Marcus Vale: Dario Amodei's team just handed Congress a number that should make every frontier AI investor rethink the whole model. 28.8 million.
Ben Okonkwo: That's — hm. That's a lot to open with.
Marcus Vale: Anthropic accused Alibaba of running 28.8 million exchanges with Claude — six weeks, 25,000 fraudulent accounts — operators linked to the Qwen AI lab routed the whole thing through commercial proxies. Geo-restrictions, bypassed. Completely.
Ben Okonkwo: And Claude started misidentifying as competitor models.
Marcus Vale: Which is the tell. Because adversarial distillation — forget the term, basically a student copying every dish out of a master chef's kitchen window, never gets the recipe, doesn't need it — at that volume, what got transferred wasn't just Claude's writing style. Something structural got absorbed. And now Anthropic's at the Senate Banking Committee, asking Congress to punish the people who did it. But wait — actually, here's what nobody's saying — DeepSeek, Moonshot AI, MiniMax, they already did 16 million exchanges back in February. Anthropic documented it. Nothing changed. So why Congress, why now?
Ben Okonkwo: The gap between knowing and acting. That's the real story.
Ben Okonkwo: Right, but — okay, pump the brakes for a second. Anthropic's language is 'operators linked to Alibaba's Qwen lab.' That's directional. That's not a command chain. 25,000 accounts could be purchased from resellers, routed through proxies, and the actual decision-maker is — we don't know. Alibaba is literally suing the U.S. government right now to get off the Pentagon blacklist. The public record on this is completely one-sided.
Marcus Vale: Hold on. You're saying Anthropic fabricated the query volumes?
Ben Okonkwo: No — the volumes are real. 28.8 million exchanges, not disputing that. But 'this happened' and 'Alibaba ordered it' are two different evidentiary claims. And here's what actually bothers me — February 2026, DeepSeek, Moonshot AI, MiniMax, 16 million exchanges, 24,000 fraudulent accounts. Anthropic documented it. Platform defenses didn't materially tighten. So now the Alibaba campaign is 1.8 times larger and suddenly there's a White House letter? That sequencing is — I mean, that's not national security response logic, that's something else.
Marcus Vale: OpenAI flagged the same pattern. Google flagged it. So it's not just Anthropic crying wolf.
Ben Okonkwo: Pattern is real. But pattern isn't proof of this specific actor. And the capability-transfer claim needs — actually, this is where I want to slow down. Claude misidentifying as competitor models tells you behavioral imitation worked. Output signatures got copied. It does not tell you whether Qwen absorbed Claude's actual reasoning depth or just its formatting. Those are genuinely different threat levels.
Marcus Vale: That's the gap Anthropic hasn't closed publicly.
Marcus Vale: But look — that gap you're pointing at? That's actually where the hot take lands correctly. Export controls assume you protect the weights, you protect the chips. The whole framework assumes that. But the attack surface here is the API. Every query is a legal interaction. You can't sanction a GET request.
Ben Okonkwo: Hm. That's — yeah, that's the structural problem.
Marcus Vale: Think about Claude Tag — launched June 23rd, Slack-integrated, monitors your channels, delegates tasks, connects directly to your codebase autonomously. That agentic reasoning workflow? Anthropic says that's exactly the capability domain Alibaba was targeting. So imagine a developer running Claude Tag on a Tuesday morning — that whole workflow, being systematically harvested across 28.8 million exchanges. Geo-restriction failed because commercial proxies were — frankly, trivially sufficient. At that scale.
Ben Okonkwo: Right, the geofencing failed not because it was sophisticated and got outmaneuvered — it failed because it was never going to work against commercial proxy infrastructure. That's not a close call.
Marcus Vale: Which is why Congress getting a letter isn't really about catching Alibaba. It's — wait, actually this is the part that matters for Dario Amodei specifically — Anthropic is approaching an IPO. The valuation rests on a frontier capability moat. If Qwen absorbed even 40% of Claude's software engineering and agentic reasoning at effectively zero training cost, the R&D cost asymmetry between U.S. and Chinese labs just... collapses. That's a valuation problem dressed up as a national security argument.
Ben Okonkwo: I'll partially grant that. The structural inadequacy of export controls at the API layer — that part holds. Where I'd push back is the 'moat is broken' conclusion assumes the distillation lag is short enough to matter before Anthropic ships the next capability generation. And we don't actually know that yet.
Ben Okonkwo: Right, but that's actually — okay, that's the concession I'll make. The structural question Anthropic put to Congress is real: if commercial proxies can route 28.8 million queries past geofencing, what enforcement mechanism actually holds? Export controls on chips don't touch that. API-level interactions are legal by definition. There's no framework that catches this.
Marcus Vale: Okay, not pure IPO theater. I'll take that.
Ben Okonkwo: But — and this is the part that's almost funny — Anthropic writing to the Senate Committee on Banking, Housing, and Urban Affairs while heading to an IPO is not a coincidence. The best argument Dario Amodei has for why distillation is a national security crisis is also, word for word, the best argument for why investors should pay a premium to fund the one lab that stays ahead of it.
Marcus Vale: The moat is the threat. The threat is the moat.
