The Distillation War

China doesn't need to out-research America. It just needs to out-copy it — and a letter sent to the White House this week shows the copying may already be industrial scale.

On June 10, Anthropic CEO Dario Amodei sent a letter to the White House and to two senators who agree on almost nothing — Tim Scott and Elizabeth Warren. The accusation inside it is narrow on its face and enormous underneath: operators tied to Alibaba's Qwen AI lab, Anthropic says, ran roughly 25,000 fraudulent accounts to extract nearly 29 million interactions from its Claude models between April 22 and June 5, deliberately targeting the most valuable capabilities — software engineering and agentic reasoning — in what the company calls the largest known "distillation attack" to date.

Strip away the cyber-thriller framing and the real story isn't theft. It's economics. And it threatens the single assumption holding up a trillion dollars of market value.

What distillation actually is

Knowledge distillation is mundane in the lab and radioactive in geopolitics. You take a weaker "student" model and train it on the outputs of a stronger "teacher." The student never sees the teacher's weights, training data, or methods — it just watches millions of high-quality answers and learns to imitate them. Done well, the student inherits much of the teacher's capability at a fraction of the cost.

That fraction is the whole point. The frontier labs — Anthropic, OpenAI, Google — justify their valuations on a simple premise: a durable lead. Billions in compute, scarce talent, and years of research buy a moat that followers can't cross quickly. Distillation is an acid bath for that moat. If a fast follower can spend a rounding error to clone the outputs that matter most, the lead compresses from years to months.

That is the part investors should sit with. The bull case for the entire AI capex supercycle — the chip demand, the data-center buildout, the premium multiples on the labs — rests on frontier capability staying expensive and proprietary. An effective, repeatable distillation pipeline says it stays neither.

Why Dario went to Washington, not to court

The most revealing thing about the letter is what it isn't. Anthropic didn't file a lawsuit. It wrote to policymakers — and pointedly to both sides of the aisle.

That's a strategic choice, and a shrewd one. A lawsuit is a commercial dispute over terms-of-service violations. A letter to the White House reframes the same facts as national security: American frontier capability being siphoned by a Chinese state-adjacent champion. That framing tees up the remedies Anthropic would prefer anyway — tighter export controls, mandatory "know your customer" rules on API access, and a legal posture that treats model weights and outputs as strategic assets.

Hold both truths at once, because both are real. The threat is genuine: industrial-scale capability extraction by a geopolitical rival is exactly the kind of thing export-control regimes exist to stop. And the remedy is self-serving: KYC mandates and weight-as-strategic-asset rules happen to build a regulatory moat around the incumbents who lobbied for them, precisely where the technical moat is eroding. When the company losing its natural moat asks the government for a legal one, the analyst's job is to notice — not to cheer.

The caveat the breathless coverage skips

Distillation is hard to prove. Anthropic can document fraudulent accounts and the sheer volume of scraped interactions. What it cannot easily show is the other half of the chain: that Qwen actually trained on that data. That link is, for now, circumstantial. Alibaba has not publicly responded.

There's also an awkwardness worth saying out loud. The frontier labs built their first models on the scraped open web — books, code, and writing ingested without permission. "You trained on our outputs without consent" is a real grievance, but it arrives with a long shadow. None of that makes the accusation false. It does mean the moral high ground is narrower than the press release implies — and that nuance is exactly what separates sharp analysis from a stenographer's transcript.

What it means for money and the future

Three takeaways for anyone with capital or strategy exposed to this:

1. Re-underwrite the moat. If you own the AI trade — chips, hyperscalers, the labs directly — your thesis quietly assumes the frontier lead is durable. This episode is evidence it may not be. That doesn't break the trade, but it should move distillation resistance, access controls, and capability half-life from footnotes to front-page risk factors in how you value these names.

2. The competition is asymmetric, and that favors the copier. China doesn't need to win the research race to stay one cycle behind at a tiny fraction of the cost. Staying a fast, cheap second is a perfectly good national strategy when the leader is footing the R&D bill. Expect Washington to respond with controls — and expect those controls to be as much about protecting incumbents as protecting the nation.

3. Sovereignty stops being optional. If API access is now understood as an intelligence-extraction vector, then for anything sensitive — defense, critical infrastructure, classified workflows — the calculus shifts hard toward on-device, air-gapped, and self-hosted models. "Don't send your crown jewels to someone else's endpoint" is no longer a fringe security posture. It's the obvious one.

The headline will read as a spy story: 25,000 fake accounts, a Chinese AI lab, a letter to the White House. The real story is quieter and bigger. The most expensive moat in technology may be the easiest to wade across — and the first company to discover it just asked the government for a bridge.

Alpha Briefing — what's happening, and what it means for money and the future.