David Sacks Says Anthropic-Written Safety Bill's 10²⁵ FLOP Threshold Would Hit Startups

The Anthropic bill fight just got a number attached, and it is a big one: 10²⁵.

David Sacks posted on X that a new Anthropic-written safety bill is circulating on Capitol Hill. His charge: as written, it will "sweep in most of Silicon Valley." He says the bill defines the frontier as any model trained on 10²⁵ FLOPs, the EU standard from 2023. Biden's executive order, California's SB 53 and New York's RAISE Act used 10²⁶, ten times higher, according to Sacks. He puts today's frontier at 10²⁷ and cites GPT-6 Astra, reportedly trained on more than 100,000 Grace Blackwell GPUs.

His math on the low end: a 10²⁵ model can be trained on as few as several hundred GB300 chips in three months. The definition is also cumulative, he says, adding FLOPs from fine-tuning and reinforcement learning, which would capture almost all meaningful post-training. That means thousands of custom enterprise models, and in his words "large compliance costs across corporate America for no safety gain."

He also turns Dario Amodei's own claim against him: Anthropic says it designs rules to disadvantage frontier companies like itself, and Sacks says this definition does the opposite.

Context: Politico reported on October 16 that Anthropic pitched lawmakers its own draft, which would limit when the federal government could block an advanced AI model's release and preempt some state AI safety laws. That report did not include the bill text or an Anthropic response.

My take: a threshold fight is a money fight in a trench coat. Whoever draws the line decides who pays for compliance. Watch that number.