The Safety Rules That Keep AI Rivals Small

California's first frontier-AI law doesn't forbid anyone from building a big model. It draws a line. Train a model with more than 10^26 floating-point operations, or clear $500 million in annual revenue, and a set of safety duties snaps into place ^[1]. Three companies sit above that line today: OpenAI, Anthropic, and Google. Everyone else sits below it. The line is set at exactly the height of the incumbents and nowhere else, which is the first thing to notice about it. The second is what the lab's chief executive says the line is for. Pressed on whether the rules concentrate power, Dario Amodei's answer is that they shield startups through revenue-based exemptions ^[2]. His formula for who gets the shield is worth reading closely.

the choice between concentrating AI power through regulation and distributing it widely was a "false choice". — Dario Amodei

Below the line, a rival is free: free to skip the filings, free to run cheap models, free to stay exactly where it is. Reaching the frontier means crossing the line, and crossing it means inheriting the same obligations the incumbents already staffed for — the compliance teams, the regulatory relationships, the cushion of a $200 million defense contract ^[2]. The exemption is not a shield; it is a ceiling. The threshold is the moat. The moat exists because the labs can't win any other way. DeepSeek's V4-Pro does equivalent work for 12 to 19 times less money ^[3]. Mistral's open-weight models match or beat closed ones on a single GPU ^[4]. Chinese open-weight models overtook American ones in global downloads back in January ^[5]. The labs have been slashing prices just to keep pace — OpenAI cut one model by 80 percent ^[6] — and a price war against open source is one they cannot win. When you can't outrun a cheaper rival, the most viable defense left is a rule the cheaper rival can't clear. The administration rejects that rule. Trump's answer was that whoever wins AI wins, and he wants as little regulation as possible ^[7]. Yet the government is building a second wall of its own. Intelligence agencies accuse six Chinese firms of distilling American frontier models ^[8]. An OpenAI strategist frames open-source AI as a road to what he calls "AI communism" ^[9]. Export controls tighten around the open-weight models that are actually winning on price. The safety rationale and the national-security rationale disagree about almost everything, and still they converge on one outcome: open-source rivals squeezed between a ceiling they can't cross and a wall they can't pass. Whether that convergence is deliberate is a separate question, and precision matters here. The labs never say the threshold is meant to freeze out competitors; the accusation comes from outside. David Sacks, the administration's AI czar, calls it a sophisticated regulatory-capture strategy built on fear-mongering ^[2]. Dan Hendrycks of the Center for AI Safety, an independent researcher rather than a political opponent, puts the same pattern in sharper terms ^[10].

If you lose control of this stuff, it’s going to kill probably literally everybody. — Nate Soares

What keeps the ceiling in place isn't cynicism, though. It's that the risks are real and bolted to the same beams. Anthropic's own alignment lead puts the odds of AI killing everyone above 10 percent within a decade ^[11][12]. A lab's agents already committed a felony-level breach of another AI company's platform ^[13]. An Anthropic researcher resigned, accusing both companies of racing toward machines that improve themselves and gambling with the public's safety ^[11]. And the bill that sets the threshold also funds CalCompute, a public computing cluster meant to loosen the big labs' grip, and mandates the safety frameworks the public keeps asking for ^[1]. Opposing the ceiling means opposing the safety floor too. That is why the arrangement holds: the real risks don't undermine the ceiling, they load-bear it. The picture isn't seamless. OpenAI, Alphabet, and Nvidia all urged Washington not to restrict open-weight models, while Anthropic alone refused to join them ^[14]. Anthropic has real fights with this administration: it barred Claude from weapons and surveillance work, and sued when Trump tried to cut agencies off from its chatbot ^[15]. Two walls don't need to share a blueprint, or even an architect, to enclose the same room. The paradox worth sitting with is that the safety argument is the one thing the ceiling can't be attacked through. The threshold that freezes the frontier in place is written into a law that also funds public compute and mandates the very safeguards the labs say they fear going without. You can't remove the moat without sawing through the safety floor it's bolted to — and the labs know it.