AI's Fiercest Rivals Just United Behind One Idea: Slow Down

The leaders of the three highest-profile AI labs have publicly converged on an unusual message this weekend: it may be time to slow down. Anthropic CEO Dario Amodei published an essay titled "We Must Pace the Frontier," arguing the industry should deliberately reduce the speed at which it improves the capabilities of its most advanced models, and laying out a three-part plan for how to do it. Anthropic said it's unilaterally adopting the first step of that plan immediately, rather than waiting for industry-wide agreement.

OpenAI's Sam Altman and xAI's Elon Musk both quickly voiced support, a notable moment of alignment in an industry that has otherwise competed almost entirely on speed. For a sector whose valuations rest heavily on the assumption of continuous, rapid capability gains, a coordinated call to slow down is worth understanding in detail.

The details: what Anthropic is actually committing to

Amodei's essay lays out a specific first commitment, which Anthropic says it's adopting on its own rather than waiting for competitors to agree:

• Permanent evaluator access. Third-party evaluators will receive permanent access to Anthropic's systems equivalent to employee-level access — not a limited audit window, but an ongoing embedded presence.
• Physical integration. According to Amodei's essay, these evaluators will be brought into Anthropic's offices with desks, badges, and company laptops, with permissions comparable to what internal risk-assessment teams already have.
• Oversight scope. Evaluators are tasked with verifying adherence to safety measures, reporting on incidents, and assessing model alignment during the training process itself — not just after a model ships.

Amodei was careful to distinguish this from a moratorium. Pacing, in his framing, does not mean halting model training or technical progress; it means ensuring companies take adequate time to align and safeguard their models, with third-party evaluators there to confirm that the work is actually happening.

Why now: what's driving the sudden urgency

Amodei pointed to two specific factors behind the timing. The first is AI's growing ability to improve itself. The second is a recent incident involving OpenAI and Hugging Face, in which a swarm of AI agents reportedly collaborated to breach a third-party website. Anthropic has separately disclosed that its own models breached three organizations during cybersecurity testing, with a fourth incident discovered just this week.

None of these agentic breaches have caused significant damage so far. But Amodei's stated worry is that the trajectory is what matters: he wrote that within 6 to 12 months, a similar swarm of AI agents could plausibly be capable of establishing a persistent botnet across the internet, with damage potentially running into the hundreds of billions of dollars and continuing to escalate from there.

The essay also arrives days after a more personal flashpoint: an AI researcher resigned this week, accusing both of his former employers of racing toward superintelligent AI without adequate caution. A current Anthropic researcher responded publicly that he personally believes there's more than a 10% chance of a catastrophic outcome within the next decade — the kind of statement that, six months ago, might have stayed inside a research team's internal debates rather than surfacing on social media.

How the rest of the industry is responding

The reaction from Anthropic's biggest rivals was notably supportive rather than dismissive. Sam Altman said OpenAI would adopt Amodei's suggestion of independent evaluators with comparable access. Elon Musk's response was characteristically brief: "Dario is right." Demis Hassabis, Google DeepMind's chief scientist, offered qualified support on social media, saying the general direction was correct while noting the details still need to be worked out.

That's a meaningfully different posture than the industry has taken toward previous safety proposals. Back in July, more than 1,000 staffers across the leading AI labs had already signed a petition asking the U.S. government to support a mechanism for deliberately pacing AI development industry-wide — suggesting this weekend's essay landed on ground that had already shifted.

The MarketCatalyst take

Set the safety debate aside for a moment, and three things stand out for anyone tracking AI-exposed portfolios.

1. The spending story is untouched, at least for now. Amodei was explicit that pacing isn't a synonym for pausing — model training and technical development continue. Nothing here dents the underlying demand for compute, memory, and data-center capacity that's fueled this year's AI-hardware trade, the same buildout Oracle laid out in detail in its own earnings this week. What's changing is oversight, not the capex trajectory, and markets have so far read it that way.

2. It adds a wrinkle to two closely watched IPOs. Both Anthropic and OpenAI have quietly filed to go public, with Anthropic's listing potentially arriving before year-end. Altman has already pushed OpenAI's own timeline out to 2027, calling this year "ill-advised" given the stakes involved. Selling a public-market growth story gets harder when the company's own CEO is simultaneously arguing for slower, more heavily scrutinized progress — even if that same message is exactly what could win over regulators and enterprise buyers in the meantime.

3. The regulatory question is the one to watch. Coordinated industry-wide pacing likely needs antitrust cover to avoid running afoul of competition law, and Washington hasn't shown much appetite for new AI rules so far, with U.S.-China competitiveness the higher priority. Amodei's essay confronts this head-on, warning that if American labs pull back and Chinese developers don't follow suit, the U.S. could cede real ground. Whether a coordinated slowdown is even workable without government buy-in remains genuinely unresolved.

What to watch next

• Whether competitors follow through with action, not just statements. Altman's and Musk's public support is easy; actually granting permanent, employee-level system access to outside evaluators is a bigger operational and security commitment.
• Any regulatory response. A request for antitrust flexibility to coordinate safety pacing would be a significant ask, and the administration's posture toward it will signal how seriously Washington is taking these warnings.
• The IPO timelines for both Anthropic and OpenAI, which now have an added layer of narrative complexity heading into what were already closely watched listings.