Part of: this research — view all 12 rows
Status Quo and Driving Forces As of mid-November 2026, no jurisdiction currently holds the binding authority to dictate a minimum share of a frontier developer's compute for safety or other non-capability purposes. Existing frameworks, such as the EU AI Act and California's SB 53, regulate AI models as consumer products rather than internal research instruments 2 sources. However, political momentum for such a lever is accelerating alongside AI capabilities. On November 10, Anthropic stated that expected progress indicates it has met its AI R&D acceleration threshold, and METR independently estimated a roughly 2× acceleration in research progress at Anthropic. Because a compute-allocation mandate does not require regulators to first agree on a complex capability measurement, it is emerging as an attractive pacing mechanism.
The California FAISA Pathway The earliest credible opportunity for this authority to become law is California's announced Frontier AI Safety Act (FAISA). Announced by State Senator Josh Becker on October 20, 2026, the proposal would require independent verification of developers' compute usage and authorize the Governor's Office of Emergency Services (CalOES) to set enforceable minimum percentages of R&D compute that must be allocated to safety work 44 sources. Although the bill is currently pending formal introduction and lacks published statutory text, a successful run through California's ordinary 2027 legislative calendar would typically see it take effect on January 1, 2028. This potential trajectory firmly anchors the 10th percentile estimate.
Alternative Jurisdictions and Legislative Delays If California's 2027 attempt fails, is vetoed, or features a delayed commencement, subsequent attempts in 2028 or 2029 establish the 25th percentile. Other jurisdictions are also exploring computing restrictions, though with slightly different mechanisms. For instance, New York's RAISE Act 2.0 (S.10731), introduced in late October 2026, proposes capping the compute developers devote to automated capability research 2 sources. While a strict cap on capability compute may not cleanly resolve as a "minimum share for other purposes," it highlights the spreading legislative appetite. The median estimate of early 2032 reflects a roughly 50% cumulative probability that California, another US state, the federal government, the EU, or China enacts a qualifying mandate within the next five to six years as evidence of AI recursive self-improvement grows more undeniable.
Failure Modes and the Long Tail The long tail of the distribution—stretching into the late 2030s and 2040s—accounts for the substantial risk that this highly specific regulatory mechanism is never adopted. Industry pushback frequently results in binding mandates being stripped down to disclosure-only requirements, as seen when California's vetoed SB 1047 was followed by the reporting-only SB 53. Furthermore, lawmakers and regulators may ultimately conclude that capability compute is too difficult to meaningfully distinguish from safety compute anthropic.com. If that definitional challenge proves insurmountable, governments are likely to abandon compute-share minimums entirely in favor of alternative pacing levers, such as capability evaluation gates, rigid deployment caps, or mandatory kill switches.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited