Deep dive: Merging the Labs: Safety Mechanism or Moat?

Why today

Last week, The Information reported that Sam Altman walked OpenAI staff through three endgames for building advanced AI safely. The first is an international inspector modelled on the IAEA, the UN's nuclear watchdog. The second folds the leading labs into a single CERN-style entity. In the third, one firm reaches self-improving AI first and, in a phrase that goes back to the safety theorist Eliezer Yudkowsky, melts everyone else's GPUs. Altman reportedly called that one "terrible" and favoured the inspector. The timing matters. In September, Dario Amodei's essay We Must Pace the Frontier drew public support from Altman, Elon Musk and Demis Hassabis. Within a week, paying subscribers sued all four companies, calling it an illegal cartel to limit output. By mid-October, Google and Meta were shipping new frontier models again. So today's question is narrow. If labs cannot agree to slow down without legal trouble, is merging them a safety mechanism, or a moat dressed up as one?

Plainly

Start with what was actually said. The report is paywalled. As relayed by Forbes and others, it describes options the industry has discussed, not a plan OpenAI has adopted. The merger idea is labelled the Demis plan, after Google DeepMind's Demis Hassabis. Beyond a single entity supplying models to partner countries, its specifics are not stated. Why would anyone want to merge labs at all? Because of a coordination problem. Each lab might privately prefer to go slower, but holding back alone is costly. Luisa Rodriguez of 80,000 Hours, a group focused on AI risk, puts it this way: "every week a company holds back, its rivals can release a better model, winning over more customers, more investors, and the best researchers." There are three ways out. Appoint a referee, merge into one team, or let one player win. Here is a worked example. Imagine four airlines agree to fly fewer routes so they can fix safety problems. However sincere the reason, that is an agreement between competitors to cut output. The lawsuit against the labs uses this framing. Its words are "an agreement among competitors about how fast their competing products will improve." Now merge the four airlines into one. Cutting routes becomes one company's scheduling decision. Nothing about the restraint changes. What changes is the legal category. That is the core mechanism. Consolidation turns a coordination problem that antitrust law polices into an internal choice. It is also exactly why consolidation is a competition problem. The merged entity sets the pace and the price for everyone downstream. The CERN analogy carries some weight, but less than it seems. CERN works because a particle collider is one shared instrument, and nobody sells collisions to consumers. Frontier models are products sold to millions of people, with whole companies built on top of them. A CERN for AI would be a supplier, and the phrase partner countries suggests everyone outside it is a customer. One thread runs through all three plans, and that is compute. The inspector plan reportedly involves monitoring compute and energy use. The winner-takes-all plan is about starving rivals of chips. Whoever controls access to compute controls who counts as a frontier lab.

What the evidence shows

Here is what is established, and how we know it. The lawsuit is a court record. Bloomberg Law reports it as Buist versus Anthropic, filed on September 18 in the Northern District of California. It brings a Sherman Act claim against Anthropic, OpenAI, SpaceXAI and Google, and calls the conduct an "output-restricting cartel." The public endorsements of Amodei's essay are on the record. On September 29, the White House had the leaders of six companies sign a voluntary Joint Commitment on Frontier Responsibilities. It is not legally enforceable. Separately, The Information reported that Google, OpenAI and Anthropic plan a self-regulatory body called the Standards Authority for Frontier AI. It is modelled on FINRA, the private body that polices US brokerages. The Next Web notes that reports do not name Microsoft, Meta or xAI as members. Altman's preference for an IAEA-style body is not new. He has argued for one publicly since May 2023. The staff meeting itself is a different kind of evidence. It is an internal remark, reported by a paywalled outlet, then relayed by others. The relays even disagree on the report's date by a couple of days. There is no transcript, and OpenAI has not published anything on it. Now for what is claimed but not shown. Nobody has shown that a merged lab would be safer than several competing ones. Nobody has shown it would actually go slower, rather than simply face fewer rivals. Membership is not stated. How open-weight developers would be treated is not stated. Any compute threshold is not stated. One relay says a decision on which path to take is due within a year, but it is unclear who would be deciding. The voluntary slowdown also gives us a small natural experiment. Crypto Briefing reports it largely unravelled within about a month, once Google and Meta released new models. That suggests self-restraint without enforcement is fragile. It does not tell us whether merging would be any sturdier.

How people read it

The same facts read differently depending on how close you think dangerous capability is. The AGI-forward reading takes race dynamics as the main risk. If systems are close to dangerous, fewer actors means fewer places to inspect and less pressure to cut corners. Writing in Lawfare, Nicholas Felstead argues that "competitive pressures will drive a race to the bottom on safety." He proposes narrow legal safe harbours so labs can coordinate pauses. The strongest form of this view is practical. You cannot meaningfully inspect fifty labs. The AI-realist reading starts from the absence of evidence that runaway capability is imminent. On that view, a slowdown mostly restrains product quality for paying customers, which is what the plaintiffs allege. Consolidation then entrenches whoever is already ahead. Fred Ashton of the American Action Forum, a centre-right policy group, warns that incumbents "can use sweeping antitrust exemptions to create a moat that solidifies their market position." The market reading notes who benefits from the rules. The proposed standards body has three founding labs. The slowdown held only until rivals saw an opening. The policy reading points to a cleaner path. Rodriguez argues that "governments set the rules, so they apply to every company at once, and nobody has to be the one to go first." That favours the inspector over the merger. To be fair, it is also the option Altman reportedly says he prefers.

What it means for builders and investors

Our read is this. Of the three endgames, consolidation is the one where the safety benefit is only asserted while the competitive effect follows automatically. A merger removes rivals by definition. Whether it lowers risk depends on governance nobody has described. That does not make the underlying safety concern fake. It makes this particular remedy the one to read most sceptically. Altman deserves credit for favouring the inspector, but an inspector's design still decides who is inside the tent. For builders, the near-term risk is probably not a merger. It is thresholds and membership. Rules that bite above some level of compute or energy will define who counts as frontier. Challenger labs and open-weight releases will sit on one side of that line or the other. For companies built on frontier APIs, fewer suppliers would mean less pricing power and less control over the roadmap. One question worth exploring in diligence might be this. If frontier supply narrowed to one or two providers, or release pace were capped, what would happen to this company's margins? Could it switch to a second model source, including open weights, within a quarter? What would change our mind is a published design with open membership, independent inspectors, and evidence that slowing down actually reduced incidents.

What to watch

Three things would settle a lot of this. First, the Trump administration's superintelligence oversight framework, which reports say is now being drafted. The questions are whether it uses IAEA-style inspection or licensing, and where any compute line sits. Second, the antitrust case. A ruling on a motion to dismiss would tell us whether courts weigh safety pacts on their actual effects. If they do, labs could coordinate without merging. Third, the standards body's charter, expected at the end of 2026 or early 2027. The questions are who gets in, whether open-weight developers can be certified, and whether it has any real power. Within about a year, we should know whether this talk has turned into a structure.

Steven's take

Regulatory strategies for world dominance seems timely in the USA. A straight merger of the leading labs would struggle to clear antitrust review, but watch for a government-sanctioned consortium or joint venture instead, with an antitrust exemption written into law or attached to a defence program and framed as a national champion against China. That route could achieve the same consolidation without ever being called a merger.