"We try very hard to make proposals that disadvantage (slow down) frontier AI companies while *advantaging* smaller competitors."
That is Dario Amodei, this past weekend, explaining why Anthropic's regulatory agenda is good for the little guy. Read it twice, because the asterisks are his.
The claim is elegant. It is also the most revealing sentence in the entire fight — not because it's a lie, but because it quietly assumes the one thing under dispute: that Anthropic gets to decide who counts as "frontier" and who counts as a "smaller competitor."
The most revealing sentence in the fight assumes the one thing under dispute: that Anthropic gets to decide who counts as frontier and who counts as a smaller competitor.
Start with the charge, because it's sharper than the defense. Leo Schwartz put it plainly on The Information's broadcast: "if basically Anthropic is defining what frontier is, and the system that's created is written with Anthropic in mind, then it might be impossible for any other company to even build a frontier model."
That's the fear in one sentence. Whoever holds the pen on the word "frontier" can wall off the frontier.
Last week, this thread was still mostly a technical argument about "covered frontier model" thresholds — training-cost cutoffs, revenue exemptions, the machinery of who's in scope. This week it went public and personal, routed through investor Gavin Baker, Amodei's essay, and a chorus that pulled in Yann LeCun and David Sacks.
Here's how Amodei answered the definitional charge: he didn't. He reframed it.
He called Baker's premise — regulate-and-concentrate versus distribute-and-democratize — a "false choice." He argued that carefully designed rules can constrain the biggest labs while leaving room for smaller ones. And then, crucially, he relocated the whole dispute onto different terrain.
"I think it is fundamentally a crisis of trust," Amodei said. "Ordinary people don't trust companies, governments, or the tech industry and always suspect that we are cooking up some new way to screw them over."
Watch the move. The accusation was structural: whoever authors the threshold controls the moat. The answer is psychological: people don't trust institutions, that's a decades-old problem, and AI is "just the latest iteration of it."
Trust is the solvent here. It dissolves a question about power into a question about feelings — and feelings can be earned back by, in Amodei's phrase, "actually curing cancer" rather than promising to.
But the definitional question doesn't go away just because it's been re-narrated as a trust deficit. And this is where the reporting matters more than the rhetoric.
Business Insider documented in July what Anthropic is actually doing at the state level. The company isn't pursuing a single national ceiling. It's pursuing "a strategy of one-upmanship that encourages states to impose increasingly tougher AI guardrails, rather than align around a single set of regulations."
That's the mechanism. Not one rule, but a ratchet — state by state, each tougher than the last, each defining "frontier" a little differently.
OpenAI, per the same reporting, is doing the opposite: pushing lawmakers toward common ground, a single harmonized framework. Anthropic's state chief Cesar Fernandez said it outright — the company is "not just looking to support the same bill across the country in every single state."
Now put the ratchet next to Amodei's exemption defense. He reportedly points to SB 53 and SB 1047, both of which exempt companies below certain revenue or training-cost thresholds. See — the small guys are carved out.
But an exemption clause only protects you against a definition you already know. A fifty-front ratchet of ever-tightening, ever-diverging thresholds is a different animal. The company that reportedly helped author each new tier holds a compliance map nobody else has drawn.
That's the structural point the trust framing is built to bury. Whether or not Amodei intends to lock rivals out is genuinely beside the point. Competitive capture isn't a function of intent; it's a function of who controls the process. And the documented process is: Anthropic, definition by definition, state by state.
You can grant Amodei full sincerity and the problem survives. "We try very hard" is a statement about motive. The moat is a statement about position. They don't touch.
Which is why the trust framing is the cleverest thing he's done. It makes the whole dispute unfalsifiable. If rivals object, they're the untrusting Silicon Valley bubble that reflexively equates "regulation = regulatory capture = concentration of power" — a shorthand Amodei explicitly waves away as "overly simplified." If the public objects, that's the decades-old trust deficit, not anything specific to Anthropic's lobbying. Every objection gets recoded as a symptom of the very crisis Amodei has volunteered to solve.
Here's the tell that this is a maneuver and not just a worldview: watch the other incumbent try the same trick and fail.
Mark Zuckerberg published 6,500 words this week titled "The Future is for Everyone," promising that "everyone will have an exceptionally capable personal agent." TechCrunch's Rebecca Bellan called him "almost like an anti-Dario."
The two men are pitching opposite products — Amodei sells trusted rules, Zuckerberg sells trust-me-it's-open — and both are bouncing off the same wall. Nobody's buying. Russell Brandom's line was that the Zuckerberg manifesto is "exactly why people don't like AI."
The reason both fail is identical, and it's the reason the trust framing is doomed even as it's clever: the public will not grant either man the authority to define the game. Not the rules of it (Amodei), not the openness of it (Zuckerberg). Muse Glimmer, Meta's "for everyone" model, doesn't even run on a MacBook — you need specific hardware. Not for everyone. Not yet.
So where's the clean exit? The only philosophically consistent answer in this fight is River, which General Catalyst co-led into a $1.1 billion round to build, in General Catalyst's words, "open-weight models that individuals can train and improve continuously." No central pen, no threshold to author. Weights in the open, ownership pushed to the edge.
And here is where you have to hand Amodei his single best point, because he anticipated exactly this exit and blew a hole in it.
Open weights, he argued, "are nowhere near a sufficient solution because they simply shift the concentration somewhat to those with the most compute and chip." AI is "structurally a technology that tends to concentrate power." Free the weights and you've just relocated the moat from the model to the datacenter.
River's own backers, in reply commentary rather than a formal pitch, unintentionally confirm the trap. The remarks note that continual learning means "each model instance will live in its own hardware," that "8B humans need at least 8B AIs," each running on "1-8 GPUs." Count the chips in that sentence. The democratized future is a chip-acquisition problem.
So nobody in this fight actually escapes concentration. That's the honest bottom line, and it's the one thing Amodei and his open-weight critics agree on even as they disagree about everything else. The moat is either in the rulebook or in the silicon.
The real question isn't whether AI concentrates power. Everyone concedes it does. The question is who gets to hold the pen that decides where — and Amodei's answer, delivered as a lesson about trust, is that it should be the person you can trust to hold it.
Which is exactly the authority nobody in this argument has earned yet.