California Report on Frontier AI Policy 2025-06-17

Citation

Governor’s AI working group (lead authors Fei-Fei Li, Mariano-Florentino Cuéllar and Jennifer Tour Chayes, per the March 2025 release), “The California Report on Frontier AI Policy”, published by the Office of Governor Gavin Newsom, June 17, 2025. https://www.gov.ca.gov/wp-content/uploads/2025/06/June-17-2025-%E2%80%93-The-California-Report-on-Frontier-AI-Policy.pdf. Wiki copy: raw/legislative/California Report on Frontier AI Policy 2025-06-17.md. Read 2026-10-15.

They are lines 10, 99-100, 169, 176-178, 195, 204, 212, 217, 222, 231, 250-254, 274, 300, 412, 415-417, 423, 521, 551 and 582-588 of a full transcription. [...] marks the cuts. The excerpt does not show the author list. The authors above come from the March 2025 release about the draft.

Bottom line

• The working group report does not mention CalCompute in the excerpt. Its one direct passage on compute is a summary of public comments: researchers lack access to compute, and New York has Empire AI.
• The Governor and the author say SB 53 builds on this report.

Background

The Governor says he “called on world-leading AI experts” in the fall of 2024 (SB-53 signing message 2025-09-29). A draft came out on March 18, 2025 (Governor AI working group release 2025-03-18). This is the final report. Senator Wiener based the July 2025 SB 53 amendments on it (Wiener SB-53 expansion release 2025-07-09).

Key claims

• Key principle 3 (line 10). “early design choices are critical because they shape future technological and policy trajectories.”
• Compute scale (lines 99-100). Frontier models have “estimated annual cost increases of approximately 4x in computational resources (compute) used for training and 2.5x in training dataset size.”
• Systemic risks. These include “global AI R&D concentration, market concentration, single points of failure, environmental risks”.
• Case-study lessons. “Early design choices create path dependency”; “Transparency is critical for generating holistic evidence”; “Trust expertise, but verify claims through third-party evaluation”.
• Policy windows. “Policy windows do not remain open indefinitely.”
• Third-party evaluation. It has “unmatched scale”, “unmatched diversity” and is “distinctively independent”.
• Impact thresholds. The report gives the European Union Digital Services Act threshold of “at least 45 million monthly active users in the EU” as an example.
• Foundation models need compute (line 169). Foundation models “are resource-intensive to produce, requiring significant amounts of data and compute to yield capabilities that can power a variety of downstream AI applications”.
• Benefits of frontier AI (line 176). “Frontier AI breakthroughs from California could yield transformative benefits across a range of practical applications in fields including but not limited to agriculture, biotechnology, clean technology, education, finance, medicine and public health, and transportation.”
• Policy can unlock benefits (line 178). Policy “can empower and attract top talent in California, support upskilling in key areas that will be important as frontier AI transforms a wide range of sectors, build entrepreneurial and impactful cross-sectoral initiatives to bridge siloes of expertise, and generally enable a policy environment that promotes innovation. That includes startups and the broader ecosystem of actors who can leverage frontier AI for tremendous benefit.”
• National competitiveness (line 195). “Emerging technologies from California, including semiconductors, cloud computing platforms, and consumer internet services, have all strengthened U.S. national competitiveness.”
• Evidence-generating mechanisms (line 204). “The evidence gaps motivate our exploration of evidence-generating mechanisms like public- facing transparency, third-party researcher access, and adverse event reporting in subsequent sections.”
• Evaluation is immature (line 212). “Overall, current evaluation tech- niques are nascent and ad hoc, and they have not yet achieved the scientific rigor [^137] expected in other domains”.
• Inference scaling (line 231). “This improvement comes from inference scaling, which means using more computing power during the actual operation of AI models, not just during training.” Also: “The amount of compute required to build a model with a given level of performance has declined.”
• Openness (line 412). “Open models are often associated with benefits to innovation, competition, privacy, and science, but they may be more prime for misuse by malicious actors”. Also: “Open model weights enable deeper inspection, which is necessary for several forms of scientific research, including on safety and security”.
• Researcher access barriers (line 423). “Third-party AI evaluators require access to prominent foundation models and permission from developers to conduct risk assessments.” Company terms of service “often legally prohibit AI safety and trustworthiness research”. The report adds: “in March 2024, over 350 leading AI researchers and advocates signed an open letter calling for a safe harbor for independent AI evaluation.”
• Public feedback (line 551). “We received over 60 feedback submissions”. Academia, industry and civil society each gave “at least 25% of all submissions.”
• Public comments, “The AI talent pipeline” (lines 582-588). “These comments suggested initiatives to improve workforce training for workers across the AI supply chain, in- cluding via collaboration with the state’s educational institutions.”
• Public comments, “The governance of open and closed models” (lines 582-588). “Some highlighted that the open release of foundation models, including via broader open-source development practices, is critical for both spurring innovations and increasing safety and transparency. Others highlighted the risks of open models.”
• Public comments, “Resourcing the AI ecosystem”.

  Resourcing the AI ecosystem. Several comments stressed the importance of public funds to advance research on AI, highlighting the strengths of California’s research institutions, the initiatives in other states such as New York’s Empire AI, and the deficits in access to compute for researchers.

Evidence and method

The “Resourcing” passage reports what public commenters said. It is not a finding of the working group.

Limits

• The excerpt has no CalCompute passage.
• The compute access passage is a summary of comments, not the authors’ recommendation.
• The “AI talent pipeline” and “governance of open and closed models” passages also summarize comments. They are not findings of the working group.
• None of them names CalCompute or a public compute cluster.