Update: OpenAI and Google Collaborating on AI Safety; Antitrust Risk Examined

Full analysis

What's new since we last covered this: FTC antitrust scrutiny of safety coordination talks.

OpenAI's policy chief Chris Lehane says Google, OpenAI, and Anthropic are now working together on AI safety, and that OpenAI is checking whether the labs could legally agree to slow down how fast they ship new models. FTC chair Andrew Ferguson called that "moat digging" within a day. The question for anyone building on these APIs: is this real safety cooperation, or three incumbents testing whether the word "safety" can shelter a supply-restriction deal from antitrust law?

This is easy to undo. Nothing has been agreed. What's actually being decided is a legal theory: can the three biggest labs coordinate on deployment pace without getting sued. There's no deadline. The FTC set the tone before OpenAI even finished the sentence.

The Skeptic

Ferguson's first instinct is the right one. Two labs that already hold the enterprise contracts and the mindshare asking whether they can legally agree to slow shipping is the same question a cartel asks about limiting supply. Lehane reaches for airlines. Boeing and Airbus do not agree on how many planes ship per quarter. They agree on how to not crash them. Those are different things, and the announcement blurs them on purpose. The pivot is the point: OpenAI spent months funding a group, Leading the Future, that fought AI safety bills. Now it's meeting Congress and preaching cooperation. That's not a change of heart. It's a change of legal strategy once the old one drew heat.

The Safety Lens

Real safety cooperation has parts you can name. Shared descriptions of what could go wrong. Shared tests. Disclosures somebody outside the labs can audit. Lehane described none of that. He described talking. When all three incumbents claim "safety" at once and none of them commit to an outside referee, safety stops working as a brake on any of them and starts working as a press line. The in-house evaluator idea in the same news window makes it worse: an auditor who reports to the lab it audits is not independent, and everyone signing the checks knows it. The aviation comparison actually cuts against Lehane. Early aviation safety was inadequate precisely until the NTSB, an outside body with subpoena power, took over crash investigation. Voluntary talk among carriers did not get us there.

The Enterprise Buyer

If you run a multi-vendor setup, OpenAI plus an Anthropic fallback, the hedge only works if the two ship on different clocks. A coordinated slowdown converges their release calendars, and the diversification you paid for evaporates. Worse is the information gap. The labs would know a slowdown timeline before you do, which makes capacity planning and any SLA you've signed downstream genuinely harder to honor. You'd be committing to customers on a roadmap the vendors have quietly agreed to stall. No procurement team can sign that. And the smaller model you might have used as leverage, Mistral, a Llama variant, an open-weight option, is exactly the competitor Ferguson says a coordination pact would lock out.

The Compute Pragmatist

Follow the money underneath. A deployment slowdown among the three big labs is an agreement to stop expanding inference capacity, the compute that actually serves model responses. That lands badly for NVIDIA and the hyperscalers, who have already signed long-horizon buildout contracts betting demand keeps climbing. Microsoft and Google committed capital on growth. A soft slowdown strands some of that capacity and misaligns the bet. Note who has the antitrust exemption question and who eats the risk: the labs do the signaling, the hyperscalers absorb the demand swing. That's a big reason I doubt a real slowdown happens. Google is a party to the safety talks and also one of the hyperscalers that loses if deployment stalls. Its own two pockets point in opposite directions.

The Researcher

The airline analogy is carrying the entire argument, and it doesn't hold. Aviation safety runs on mandatory incident reporting, an independent investigator with real power, and decades of written standards. Frontier labs have agreed to none of that. "Collaborating on safety" with no named artifact, no shared test, no disclosure schedule is not research. The useful question, what coordination actually reduces risk without handing incumbents an information edge, is exactly the one Lehane left blank. Until there's a document you can point to, treat this as positioning.

Where they split

The real disagreement is whether the slowdown even ships. The Skeptic and the Safety Lens read the whole thing as legal cover, which means the talk continues and no binding pact ever arrives. The Enterprise Buyer and the Builder are war-gaming a world where it does arrive and converges release calendars. The Compute Pragmatist breaks the tie: Google sits on both sides of the table, earning on the buildout it would have to stall, so the party best able to make a slowdown real has the strongest reason to kill it.

The second split is on the auditors. The Safety Lens says in-house evaluators are safety branding. Nobody at the table argues they're independent. Both sides agree on that point, which means the auditor question is already settled before the negotiation starts.

What it hinges on

One thing. Does any named, auditable commitment appear, a shared test, a disclosure rule, an outside body with teeth, or does this stay at "we talk to each other"? Everything real depends on that artifact existing. Ferguson already told them the coordination route runs straight into the FTC. Combine a regulator saying no with a hyperscaler-lab (Google) that loses money if the slowdown works, and the machinery for a binding deal isn't there.

Prediction: No formal agreement among OpenAI, Anthropic, and Google to coordinate or slow model deployment pace will be signed or announced by 2027-03-31, and any published "safety collaboration" from the three will contain no independent, outside-audited enforcement mechanism.

Confidence: Medium. FTC opposition plus Google's own conflicting compute incentive block a binding deal.

Why: Ferguson called a coordination pact "moat digging" within a day, so any formal slowdown agreement walks straight into an antitrust fight the labs have said they're only "exploring." Google is both a party to the safety talks and a hyperscaler that has committed capital to a compute buildout a slowdown would strand, so its incentives point against making the pact binding. Lehane's own description was talk, not artifacts, with no shared test, disclosure rule, or outside referee named, which is what you announce when the goal is positioning rather than a constraint you intend to be held to. The opposite outcome, a signed pact with real teeth, would require the labs to accept the exact enforcement they spent this window declining to specify.

Revisit by 2027-03-31: We're right if no binding deployment-coordination agreement exists among the three and any joint safety framework lacks an independent, outside-audited enforcement body. We're wrong if the three announce a formal deployment-pace agreement, or a safety framework with a genuinely independent auditor holding disclosure or investigative power.

The in-house auditor structure is the clearest evidence this stays soft. An evaluator who reports to the lab it audits is exactly the mechanism you build when you want the label without the constraint.