David Sacks spent Saturday telling two of the world’s most valuable AI labs to go ahead and slow down, then spent the following days making clear he wants nothing to do with the regulatory bargain they hoped would come with it. The exchange, reported by The Next Web on September 13, 2026, reads like a policy footnote. It is not. Underneath the back-and-forth between Sacks, Anthropic CEO Dario Amodei, and OpenAI CEO Sam Altman sits a much bigger fight: whether the United States regulates frontier AI the way it regulates cars and drugs, with pre-approval regimes and licensing, or the way Sacks wants it regulated, through product liability lawsuits after something goes wrong. That fight now has a live test case sitting on Governor Gavin Newsom’s desk in the form of California’s SB 53 and its 2026 follow-up, SB 813, and it is going to shape how every AI company operating in the US budgets for compliance over the next several years.
This is a compliance and governance story as much as a personality clash. Security teams, general counsel offices, and compliance leads at AI-adjacent companies are watching this exchange because it previews which regulatory model is going to win, and that outcome determines whether frontier AI safety gets enforced through government pre-clearance or through the courts after an incident. It follows a week in which Sacks separately rejected the substance of the labs’ proposed slowdown deal and an OpenAI safety researcher warned publicly about AI risk just days after the GPT-6 Astra launch. Below is what actually happened, why the two competing regulatory philosophies matter, and what the data on California’s frontier AI law says about where this is headed.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Sacks actually said, and why it matters more than the headline
Sacks currently chairs the President’s Council of Advisors on Science and Technology and served as the White House’s AI and crypto czar until March 2026, when his term as a special government employee ended. He remains one of the administration’s most-watched voices on AI policy even without the formal title, which is why his response to Amodei’s essay landed as it did. According to Moneycontrol’s report on his public remarks, Sacks wrote: “Dario has written that we need to pace the frontier, and Sam has agreed. People may be surprised by my response: go ahead.” He then added the line that turned a polite endorsement into a public rebuke: “The easiest way not to build superintelligence is for you to agree not to build it,” according to David Sacks, Chair of the President’s Council of Advisors on Science and Technology, as reported by Moneycontrol.
What makes the exchange consequential is not that Sacks disagreed with slowing down AI development. It is that he specifically rejected the idea that voluntary pacing should come bundled with regulatory concessions, describing Anthropic’s preference for a government-set framework as, in his words, “That is regulatory capture,” according to David Sacks, as reported by BiGGo Finance. That framing puts Sacks on record against the two policy tools frontier labs have reportedly wanted most: an antitrust waiver that would let competitors coordinate on safety standards without running afoul of collusion law, and a federal approval regime that would substitute a licensing process for the patchwork of state laws now emerging.
The two regulatory models AI companies are actually fighting over
Strip away the personalities and the dispute reduces to two competing regulatory philosophies that have been circling frontier AI policy since the 2023 federal AI executive order first tried to define what a “high-risk” model even was.
The first model is pre-approval regulation: a regulator reviews a model, a drug, or a vehicle before it reaches the public, and clears it against a defined safety standard. This is the model pharmaceutical companies operate under with the FDA, and it is broadly the model the EU AI Act leans toward for its highest-risk AI systems, with conformity assessments required before deployment. Frontier labs have periodically signaled they would accept a version of this if it came with legal certainty and a single national standard instead of fifty different ones.
The second model is product liability: companies ship first, and if something goes wrong, courts and plaintiffs’ attorneys assign damages after the fact. This is closer to how most consumer software already operates in the US, and it is the model Sacks has now publicly endorsed for frontier AI, arguing that market pressure and the threat of lawsuits already discipline lab behavior without a licensing bureaucracy. Sacks has framed a government-brokered approval process in exchange for voluntary slowdown as functionally equivalent to “blackmail,” according to David Sacks, as summarized in reporting on his position.
Neither model is hypothetical anymore. California has already built a version of the transparency-first, liability-backed approach into state law, and it went into effect months before this week’s public dispute.
Inside SB 53: the $1 million law already governing frontier AI
California’s Transparency in Frontier Artificial Intelligence Act, known as SB 53, is the piece of law that makes this debate concrete rather than theoretical. Authored by state Senator Scott Wiener and signed by Governor Newsom on September 29, 2025, SB 53 is widely described as the first US state law that specifically targets frontier AI safety and transparency rather than AI in general.
The law defines a “frontier model” as a foundation model trained using more than 1026 floating-point or integer operations, a threshold that mirrors the one set in the 2023 federal AI executive order. That is a meaningfully higher bar than the EU AI Act’s own frontier-style threshold of 1025 FLOPs, meaning California’s law technically applies to fewer models than the EU’s equivalent trigger, even though it is stricter in what it demands once a model crosses that line. Companies whose models clear the threshold have to publish safety and governance frameworks, disclose critical safety incidents, and comply with new whistleblower protections for employees who raise serious risk concerns. Violations carry civil penalties of up to $1,000,000 per violation, enforced by the California Attorney General’s office, according to the bill text published by the California Legislature.
SB 53 also created CalCompute, a public cloud computing consortium housed inside California’s Government Operations Agency, intended to give researchers and smaller developers computing capacity without depending entirely on the same handful of frontier labs whose behavior the law is trying to regulate. Notably, SB 53 stopped short of the mandatory kill switches and third-party pre-deployment audits that earlier drafts reportedly considered, landing instead on a transparency-and-disclosure model that leans toward the liability-based approach Sacks has now publicly backed, rather than a licensing regime.
SB 813 raises the stakes with independent verification
If SB 53 set the baseline, SB 813, authored by state Senator Jerry McNerney and signed by Newsom in 2026, moves California closer to something that looks more like the pre-approval model Sacks opposes. SB 813 establishes a framework for independent verification organizations to assess whether deployed AI systems comply with state law, effectively creating a certification ecosystem that sits between voluntary self-reporting and a full federal licensing regime. It is described by Newsom’s office as a first-in-the-nation set of AI safeguards built around third-party assessment bodies.
That is precisely the kind of “approval regime” language Sacks singled out for criticism. Whether SB 813’s independent verification organizations end up functioning like a light-touch audit system or a de facto licensing gate will likely determine whether Sacks’ allies in Washington attempt to preempt it with federal legislation, an outcome frontier labs would probably welcome if it replaces a fifty-state patchwork with one federal standard, but would resist if that standard comes with strings attached.
| Regulatory approach | Trigger / threshold | Enforcement mechanism | Penalty structure |
|---|---|---|---|
| California SB 53 (2025) | Models trained above 1026 FLOPs | California Attorney General | Up to $1,000,000 per violation |
| California SB 813 (2026) | Deployed AI systems subject to state law | Independent verification organizations | Not yet fully codified in public reporting |
| 2023 federal AI executive order | Models trained above 1026 FLOPs (dual-use foundation models) | Federal agencies (reporting requirements) | Administrative, no direct civil penalty regime |
| EU AI Act (frontier-style trigger) | Models trained above 1025 FLOPs | EU AI Office / national authorities | Conformity assessment required pre-deployment |
| Sacks-endorsed model | No FLOPs trigger; liability applies post-deployment | Courts, product liability litigation | Case-by-case damages, no fixed cap |
Why Anthropic and OpenAI wanted a different deal
Anthropic CEO Dario Amodei published an essay titled “We Must Pace the Frontier” on September 12, 2026, the day before Sacks’ response, arguing that AI companies and governments need to deliberately slow the rate at which model capabilities improve so that safety measures have time to catch up. Amodei’s essay reportedly laid out a three-part plan: independent monitoring of AI models during development, industry-wide regulation that applies the same pacing standard to every frontier lab, and eventually global regulation of frontier AI systems. Sam Altman, according to Reuters’ reporting on an internal OpenAI meeting, told employees the company was open to slowing development of its own AI systems amid growing safety concerns, effectively aligning OpenAI with Anthropic’s position within a day of the essay’s publication.
The debate also arrives amid separate scrutiny of both labs’ security posture. Anthropic disclosed its fourth Claude-related cybersecurity incident this year, while OpenAI acknowledged a rogue AI agent swarm attack traced through the RubyGems package registry. Neither incident is a direct input into the SB 53 or SB 813 debate, but both feed the broader argument, made by critics of a liability-only model, that frontier labs are still working through basic security failures even as they ask to set their own pacing. The logic behind wanting a coordinated, government-backed framework rather than unilateral pacing is straightforward from a competitive standpoint. If Anthropic slows its own frontier development but OpenAI, Google DeepMind, xAI, and Chinese labs like DeepSeek do not, Anthropic loses ground without making the industry meaningfully safer. A shared framework, whether enforced through antitrust exemptions that let labs coordinate directly or through a government approval regime that applies uniformly, solves that collective action problem. It is also, in Sacks’ framing, exactly the kind of arrangement that risks becoming a cartel with a safety label attached, since it would let a small number of incumbent labs set the rules that any new entrant would have to meet before it could compete.
The antitrust angle: what “cartel” actually means here
Antitrust law generally prohibits competitors from coordinating on output, pricing, or standards in ways that reduce competition, even when the stated intent is benign. An antitrust waiver for AI safety coordination would carve out an exception specifically for frontier labs to jointly agree on capability limits, safety thresholds, or deployment pacing without that coordination being treated as collusion. Supporters argue this kind of carve-out already exists in other industries, such as coordinated safety standards among airlines or pharmaceutical manufacturers. Critics, and now Sacks publicly, argue that granting it to a handful of well-capitalized frontier labs effectively lets incumbents write the rulebook that determines which future competitors can enter the market at all, since compliance costs and coordination requirements tend to fall hardest on smaller players without existing legal and compliance infrastructure.
This is where Sacks’ broader deregulatory record becomes relevant context rather than just political color. In earlier public remarks on AI policy, Sacks said, “Let’s get the red tape out of the way,” and separately, “Let’s cut 10 times more regulations than new ones we put in place,” according to David Sacks, Chair of the President’s Council of Advisors on Science and Technology, as reported by Nextgov. Those statements, made while Sacks still held the formal AI czar title, establish that his September 13 rejection of an antitrust waiver and approval regime is consistent with a deregulatory position he has held publicly since at least mid-2025, not a new stance improvised in response to Amodei’s essay.
Historical context: how we got from an executive order to a state law fight
The regulatory throughline here stretches back to the 2023 federal AI executive order on safe, secure, and trustworthy AI development, which was the first major US policy document to define a FLOPs-based threshold for high-risk foundation models. That threshold, 1026 operations, became the template both the federal government and California later reused, which is why SB 53’s frontier model definition looks almost identical to the federal one. When that executive order’s reporting requirements proved too light-touch for state legislators concerned about catastrophic AI risk, California moved first with SB 53 in September 2025, betting that a large state with outsized influence over where AI companies incorporate and operate could set a de facto national standard the way it has previously done with vehicle emissions and data privacy law.
SB 813 followed within a year, tightening the enforcement mechanism with independent verification bodies rather than waiting for a slower federal process to catch up. Sacks’ September 13 comments arrive at the exact moment this two-law framework is maturing from paper into practice, which is why his objection to an “approval regime” reads less like abstract philosophy and more like a direct shot at the direction California’s own legislature has already taken. It also follows a separate coalition push, covered when more than 25 advocacy groups asked the White House to disclose its own confidential AI policy framework, underscoring how much of this fight is still happening behind closed doors even as the state-level rules take effect in public.
Market impact: what this means for compliance budgets
For any company building on top of frontier models, whether that’s an enterprise security vendor, a fintech using an LLM for fraud detection, or a healthcare startup running diagnostics through a foundation model API, the practical question is which compliance regime they end up building toward. A liability-first system, the one Sacks favors, pushes compliance spending toward legal risk assessment, insurance, and incident response readiness, since exposure shows up after deployment when something breaks. A pre-approval system, the direction SB 813’s verification bodies point toward, pushes spending earlier in the pipeline, toward documentation, third-party audits, and certification processes before a model ever reaches production.
These are not interchangeable cost structures. Liability-based compliance is cheaper upfront and more expensive when something goes wrong; pre-approval compliance is more expensive upfront and theoretically reduces the odds of a catastrophic, headline-generating failure. Security and compliance teams at companies that consume frontier model APIs, rather than build them, are effectively hostages to whichever model wins this fight, since their own downstream obligations, especially under frameworks like SB 53’s incident disclosure requirements, will be shaped by whatever their upstream model providers are required to report.
| Stakeholder | Public position as of Sept 13, 2026 | Preferred regulatory model |
|---|---|---|
| David Sacks (PCAST Chair, former AI czar) | Endorses voluntary pacing, opposes antitrust waiver and approval regime | Product liability, post-deployment enforcement |
| Dario Amodei (Anthropic CEO) | Published “We Must Pace the Frontier,” calls for industry-wide and global regulation | Coordinated regulation with independent monitoring |
| Sam Altman (OpenAI CEO) | Reportedly open to slowing development, aligned with Anthropic within a day | Coordinated regulation, details less specified publicly |
| California legislature (Wiener, McNerney) | Enacted SB 53 (2025) and SB 813 (2026) | Transparency plus independent verification, hybrid model |
Competitive comparison: how US, California, and EU approaches diverge
Compared against the EU AI Act’s conformity-assessment model, both of California’s laws sit closer to the liability-and-disclosure end of the spectrum, at least for now. The EU requires certain high-risk AI systems to pass a conformity assessment before they can be placed on the market, a pre-approval step California has so far avoided in favor of post-hoc transparency reporting and, with SB 813, independent verification after the fact rather than before deployment. That divergence matters commercially: a frontier lab optimizing for the US market under Sacks’ preferred model can ship faster and deal with liability exposure later, while the same lab targeting EU users under the AI Act’s higher-risk categories has to budget for pre-market assessment regardless of what happens in Sacramento or Washington. Labs operating globally, which describes every major frontier AI company right now, end up building compliance processes to the strictest standard they face and applying it everywhere, which in practice means the EU AI Act’s pre-approval logic likely shapes internal lab processes even in jurisdictions like California that have chosen a lighter-touch model.
What comes next: five predictions
- Federal preemption push gains momentum. Expect renewed congressional interest in a federal AI standard that would preempt state laws like SB 53 and SB 813, framed by Sacks-aligned lawmakers as replacing a fragmented state patchwork, and opposed by state legislators who see it as weakening enforcement.
- SB 813’s verification bodies face an early legal challenge. Given Sacks’ direct criticism of “approval regime” language, expect industry-funded legal challenges testing whether SB 813’s independent verification requirement functions as an unconstitutional barrier to interstate commerce or a legitimate state safety measure.
- Anthropic and OpenAI diverge publicly on pacing specifics. Amodei’s three-part plan is more detailed than anything Altman has committed to; expect OpenAI to lag on specifics like independent monitoring while emphasizing voluntary internal safety processes instead.
- Insurance products for frontier AI liability expand. If Sacks’ liability-first model gains traction, expect a wave of AI liability insurance products aimed at labs and downstream enterprise customers, similar to how cyber-insurance markets matured after major data breach litigation.
- Other states follow California’s lead rather than wait for Washington. With no clear federal standard imminent, expect additional states to introduce SB 53-style transparency bills in their 2027 legislative sessions, increasing pressure for exactly the kind of federal preemption Sacks and frontier labs both, for different reasons, may end up wanting.
What security and compliance teams should watch
This sits alongside the broader run of frontier-AI and cybersecurity threats tracked through 2026, where regulatory uncertainty is now as material to enterprise risk planning as any single vulnerability disclosure. For CISOs and compliance leads at companies building on frontier model APIs, the near-term action item is not political but operational: track whether your model providers are subject to SB 53’s incident disclosure requirements, since a frontier lab’s disclosed safety incident could trigger downstream reporting obligations for enterprise customers depending on how integration agreements are structured. It is also worth tracking whether SB 813’s independent verification organizations, once operational, publish assessment criteria that could function as a de facto industry standard even for companies outside California, the way California’s data privacy law effectively became a national baseline for many companies that never intended to operate under it specifically. Coverage of Wall Street’s own read on the moment, including reporting that framed GPT-6 Astra’s launch as the start of an AGI era and earlier commentary noting a 10% AI catastrophic-risk figure cited by Anthropic’s own researchers around the company’s IPO buildup, gives a sense of how quickly the market narrative around frontier AI risk has shifted this year, and why regulators on both sides of this fight feel the timeline is short.
Frequently asked questions
Who is David Sacks and why does his opinion on AI regulation matter?
David Sacks is a venture capitalist who served as the White House’s AI and crypto czar until his term as a special government employee ended in March 2026. He now chairs the President’s Council of Advisors on Science and Technology and remains one of the administration’s most influential voices on AI policy.
What is SB 53 and does it currently apply to AI companies?
SB 53, California’s Transparency in Frontier Artificial Intelligence Act, was signed on September 29, 2025 and applies to foundation models trained using more than 1026 FLOPs. It requires safety framework disclosures, incident reporting, and whistleblower protections, with civil penalties of up to $1,000,000 per violation enforced by the California Attorney General.
What is the difference between a product liability model and a pre-approval regulatory model for AI?
A product liability model allows companies to deploy AI systems first and assigns legal responsibility after harm occurs, typically through lawsuits. A pre-approval model requires a regulator or certified third party to review and clear a system before it can be deployed, similar to how pharmaceutical products are regulated by the FDA.
What did Dario Amodei’s “We Must Pace the Frontier” essay actually propose?
Published September 12, 2026, the essay reportedly proposed slowing the rate of AI capability improvements through independent monitoring of models during development, industry-wide regulation applying the same standard to all frontier labs, and eventual global regulation of frontier AI systems.
Is there a federal law that regulates frontier AI models in the US?
The main federal reference point remains the 2023 AI executive order, which set reporting requirements and a FLOPs-based threshold for high-risk foundation models, but it does not carry the civil penalty structure that California’s SB 53 introduced at the state level.
How does the EU AI Act compare to California’s approach?
The EU AI Act generally requires high-risk AI systems to pass a conformity assessment before reaching the market, a pre-approval step. California’s SB 53 and SB 813 instead lean toward transparency reporting and post-deployment independent verification, a lighter-touch approach that aligns more closely with the liability-based model Sacks has publicly endorsed.
What is SB 813 and how is it different from SB 53?
SB 813, authored by state Senator Jerry McNerney and signed by Governor Newsom in 2026, builds on SB 53 by establishing independent verification organizations that assess whether deployed AI systems comply with state law, moving California closer to a certification-style enforcement model than SB 53’s transparency-only approach.
