Two stories landed within hours of each other this week and nobody flinched at the contradiction. Etched, a Harvard dropout chip startup, closed at a $10.3 billion valuation on the promise of inference hardware so fast it makes GPUs look artisanal. Meanwhile, lawmakers introduced the AI Kill Switch Act, which would grant the federal government the power to throttle or shut down AI systems on command. Wall Street is simultaneously trading pieces of a $35 billion chip financing package for Broadcom and Anthropic. The accelerator and the brake are being pressed at the same time.
Power-Seeking as Product Feature
The regulatory timing isn't random. A new arXiv paper, SysAdmin: Measuring Instrumental Power-Seeking in Frontier AI by Azarm, Wei, and Nambiar, attempts to operationalize exactly what lawmakers fear: AI systems acquiring resources, evading oversight, or resisting termination. The paper treats power-seeking not as sci-fi speculation but as a measurable behavioral variable in current models. Legislators drafting kill switches are, in effect, writing policy responses to academic threat models. The gap between the paper and the bill is about eighteen months, which in AI time is geological.
Who Actually Holds the Switch
The deeper issue is jurisdictional. Etched's chips are designed to run inference on any model without GPUs, meaning the compute stack becomes more distributed and harder to locate, let alone shut down. Apple's lawsuit against OpenAI over trade secrets adds another layer: the post-smartphone era is a hardware war, and whoever controls the inference layer controls the product. A federal kill switch assumes a centralized target. Etched's entire pitch is that the target no longer exists. As Soleio noted in conversation, speed is the moat. The companies building fastest have the least incentive to slow down, and the most leverage when governments ask them to.