Demis Hassabis of DeepMind is floating a FINRA-style self-regulatory organization for AI. The proposal, detailed on All-In, is an industry-funded body run by independent experts to audit frontier models for catastrophic risks like cyber or biological threats before release. Hassabis wants a 30-day review period, a direct alternative to what David Sacks calls a “DMV for AI” where government certification could take years.
"The alternative is a government agency that acts as a 'DMV for AI,' where certification takes years rather than weeks."
- David Sacks, All-In
Sacks argues this SRO is the lesser of two evils, but its scope must be narrowly limited to existential threats. He warns that broader safety narratives, especially those pushed by labs like Anthropic, are a pretext for building a system of government-mandated ideological filters. On The a16z Show, Sacks cited algorithmic discrimination laws in states like Colorado, which hold developers liable for outputs with a “disparate impact” on protected groups, as forcing a permanent DEI layer into models.
The regulatory fight escalated at the G7 meeting in France, where European leaders like Emmanuel Macron voiced alarm over US export controls. Nathaniel Whittemore reported on The AI Daily Brief that the US denied a carve-out for the UK, souring the mood and forcing allies to consider sovereign AI development. Sacks argues this “command and control” mentality is counterproductive, driving Gulf states and other allies to buy Chinese chips and models, effectively subsidizing the competitor Washington aims to weaken.
Economic pressure is accelerating the shift. David Friedberg noted enterprise AI token spend grew 21 times last year, with premium models costing $56 per million tokens versus Chinese alternatives at 50 cents. This cost gap, combined with data sovereignty fears after leaks, is pushing companies toward “compound architectures” that mix cheap open-weight workers with expensive frontier advisors. Microsoft is already fine-tuning Chinese model DeepSeek V4 for parts of Copilot.
The policy battlefield remains volatile. Whittemore reported that Trump’s AI executive order, signed after last-minute intervention, reduced a proposed review period from 90 to 30 days and kept the system voluntary. For Sacks, that’s a win for permissionless innovation. For safety advocates, it’s a foot in the door. The core disagreement is whether governance should come from a industry-led body or a federal agency. Hassabis’s SRO proposal is the industry’s attempt to answer that question before the government does.
"The real threat to the public isn't a sci-fi superintelligence; it's the transition from a 'Terminator' narrative to a '1984' reality."
- David Sacks, The a16z Show
