Sacks Slams Anthropic AI Safety Bill Over Stale FLOPs Threshold
Synopsis
White House AI and Crypto Czar David Sacks accused Anthropic of circulating a Capitol Hill safety bill built on a 2023 EU compute threshold — 10²⁵ FLOPs — he says is three years out of date, would sweep in enterprise AI, and would entrench incumbents over startups.
Key Takeaways
A new AI safety bill attributed to Anthropic is circulating on Capitol Hill and defines 'frontier' models as those trained on 10²⁵ FLOPs — the same figure used in 2023 EU AI Act drafts.
David Sacks argues the true frontier in 2026 is 10²⁷ FLOPs ; Biden's executive order , California SB 53 , and New York's RAISE Act all used the higher bar of 10²⁶ FLOPs .
The bill's cumulative FLOPs clause — counting fine-tuning and reinforcement learning — could pull most enterprise AI post-training work into a frontier regulatory regime.
Sacks contends the stale threshold favours incumbents like Anthropic that already exceed it and can absorb compliance costs, while hobbling open-model competitors.
He calls on Anthropic CEO Dario Amodei to publicly clarify whether the company genuinely believes 10²⁵ FLOPs still represents the frontier.
The controversy highlights a structural flaw in static compute-based regulation: definitions written into law become increasingly misaligned with actual technological risk as hardware advances.
A compute threshold lifted from a 2023 EU draft may soon define what counts as 'frontier AI' in the United States — and White House AI and Crypto Czar David Sacks is calling it out as both technically obsolete and strategically convenient for one company in particular. On Saturday, October 17, 2026, Sacks posted a detailed critique of a new AI safety bill he says was authored by Anthropic and is now circulating on Capitol Hill.
The 10²⁵ FLOPs Problem — and Why the Number Matters
At the heart of the dispute is a single number: 10²⁵ FLOPs — floating point operations used as a proxy for how powerful an AI model is. The bill, as Sacks describes it, borrows this figure directly from the EU AI Act, a standard written in 2023 when the frontier looked very different. By contrast, Biden's October 2023 executive order, California's SB 53, and New York's RAISE Act all used 10²⁶ FLOPs — ten times higher. Sacks argues the actual frontier in 2026 sits at 10²⁷ FLOPs, pointing to models like GPT-6 Astra, which he says was reportedly trained on more than 100,000 Grace Blackwell GPUs.
The gap is not trivial. A 10²⁵ FLOPs model, Sacks writes, can be trained on 'as few as several hundred GB300 chips in three months.' That places it firmly in the range of mid-tier enterprise workloads — not the bleeding-edge systems the safety debate is ostensibly about. Anchoring a 'frontier' definition to a three-year-old number means the regulatory net drags far below the waterline of genuine risk.
The Cumulative FLOPs Clause That Could Ensnare Enterprise AI
The threshold itself is only one concern. Sacks flags a second, less-discussed design choice: the bill's definition is cumulative. It counts FLOPs not just from initial training but also from fine-tuning, reinforcement learning, and other material modifications. In practice, that means almost every meaningful post-training operation — exactly the work enterprises are scaling aggressively — could push a model across the regulatory line.
The enterprise trend, Sacks notes, is moving toward 'thousands of smaller custom models built for specific workloads.' These are not frontier systems by any safety-relevant measure. Forcing them into a compliance regime designed for the most powerful models in the world would impose large costs on corporate America with no corresponding safety benefit.
Sacks vs. Dario: Incumbents, Startups, and the 'Safety' Label
Anthropic CEO Dario Amodei has previously claimed, publicly, that his company designs rules intended to disadvantage frontier incumbents — including Anthropic itself — and benefit smaller competitors. Sacks inverts that framing directly. A stale 2023 cutoff, he argues, does the opposite: it advantages companies that already clear the bar and can absorb compliance costs, while handicapping the open-model ecosystem that poses the most direct competitive threat to Anthropic's pricing power.
That is a serious charge. If accurate, it suggests the bill functions less as a safety instrument and more as a regulatory moat — a dynamic that has played out repeatedly in tech policy when incumbents help draft the rules that govern them.
A Call for Anthropic to Clarify Its Position
Sacks closes with a direct challenge: 'Anthropic should clarify if they really believe 10²⁵ FLOPs represents the frontier.' The question is pointed because the answer has a binary quality — either the threshold is technically defensible, or it is not. He acknowledges that some ideas in the bill 'could be reasonable if limited to true frontier companies,' signalling an openness to safety regulation in principle, not opposition to regulation itself.
The broader issue Sacks raises is structural: any compute-based definition written into statute will be outdated before the ink dries if it does not include a mechanism to move with the frontier. Hardware capabilities, training efficiency, and model architectures are all advancing faster than any legislative cycle. A static number is, in effect, a policy that gets looser for incumbents and tighter for everyone else — automatically, year by year.
Congress now faces a straightforward test: is the FLOPs threshold in this bill a genuine safety guardrail, or a number chosen for its convenience to the company that wrote it? The answer will say a great deal about whose interests AI safety legislation in America is actually designed to protect.
Point of View
And whether compute thresholds can ever be durable regulatory instruments. The accusation that Anthropic — a safety-focused lab with commercial incentives — drafted a bill that entrenches its own market position is precisely the kind of 'regulatory capture' argument that tends to gain traction in Congress across party lines. If the FLOPs cutoff is not updated to track actual frontier capabilities, the bill risks replicating the pattern seen in financial and pharmaceutical regulation, where incumbents use compliance costs as a barrier to entry. The Indian and global startup ecosystem should watch closely: any US law that shapes enterprise AI compliance will ripple into procurement requirements and export norms that affect developers worldwide.
NationPress
17 Oct 2026
Frequently Asked Questions
What is the FLOPs threshold debate in the Anthropic AI safety bill?
The bill reportedly defines 'frontier' AI as any model trained on 10²⁵ FLOPs , a standard from 2023 EU AI Act drafts. David Sacks argues this is three years out of date — the actual frontier in 2026 is 10²⁷ FLOPs — and the lower bar would regulate far more than genuinely high-risk models.
Why does David Sacks say the bill hurts startups?
Sacks contends the 2023 threshold favours incumbents like Anthropic that already operate above it and can absorb compliance costs, while imposing those same costs on smaller companies, open-model developers, and enterprises running customised models — none of which pose frontier safety risks.
What is the cumulative FLOPs clause and why does it matter?
The bill's definition is cumulative: it adds together FLOPs from initial training, fine-tuning, and reinforcement learning. Sacks argues this would capture nearly all meaningful post-training work in the enterprise sector — work that is rapidly becoming standard practice with no corresponding increase in safety risk.
How does the Anthropic bill compare to Biden's executive order?
Biden's October 2023 executive order on AI set its trigger at 10²⁶ FLOPs — ten times higher than the 10²⁵ figure in the Anthropic-linked bill. California's SB 53 and New York's RAISE Act used the same higher bar, making the new bill's threshold notably more expansive.
What has Dario Amodei said about Anthropic designing AI rules?
Anthropic CEO Dario Amodei has publicly stated that the company designs safety rules intended to disadvantage frontier incumbents — including Anthropic itself — and benefit smaller competitors. David Sacks argues the bill's stale compute threshold does the exact opposite.
