- Monday Momentum
- Posts
- The AI Slowdown Has an IPO Calendar
The AI Slowdown Has an IPO Calendar
How two safety researchers quit calling the labs reckless, the CEOs of those labs agreed within days, and why the same warning does opposite financial work at Anthropic and OpenAI
Justin Wright
September 14, 2026 • Est. Reading Time: 6 minutes
Happy Monday!
Amodei published his call for an industry-wide slowdown roughly six weeks before Anthropic's expected October IPO. (Source: Dario Amodei)
Last week produced one of the strangest sequences in the short history of this industry. On September 9, Anthropic researcher Jacob Coxon resigned and told colleagues on Slack that superintelligent AI carried "a risk of causing human extinction." He went further publicly, writing that Anthropic and OpenAI "are racing straight to self-improving superintelligence and gambling with our lives" and that "neither company is acting responsibly." Days earlier, Anthropic's AI safety lead Mrinank Sharma had also resigned, saying the world was in peril.
Then something unusual happened. Instead of disputing the accusation, the accused agreed with it. On September 12, Dario Amodei published an essay titled "We Must Pace the Frontier," calling on the industry to deliberately slow capability gains. Sam Altman endorsed it within hours and said OpenAI would follow. Elon Musk backed it too. Three of the fiercest rivals in technology aligned on a single position in a weekend.
Altman added one more thing: OpenAI's IPO, once floated for September at above a trillion dollars, will not happen until 2027. His stated reason was safety; going public now, he said, would be an "ill-advised moment."
Anthropic, meanwhile, is still going public next month, targeting a roughly $2 trillion valuation. The question worth asking is not whether these warnings are sincere, it is why they arrived in this particular week (and why the same argument is worth so much more to one company than the other).
Two Anthropic safety researchers resigned warning of extinction risk and accusing the labs of recklessness. Within days, Amodei published a slowdown proposal, Altman endorsed it, and OpenAI delayed its IPO to 2027 citing safety. Anthropic is profitable and lists in October at a targeted $2 trillion. OpenAI is burning $17 billion a year with no positive free cash flow projected until 2029. The safety narrative is probably sincere; it also happens to solve a very specific problem for exactly one of these two companies.
The Conflation Worth Undoing
Coxon and Sharma quitting is an accusation against the labs. Their claim is that Anthropic and OpenAI are behaving irresponsibly right now, and that insiders cannot fix it from inside. Amodei's essay is a governance proposal from a lab. His claim is that the industry should pace itself and that Anthropic will move first.
These point in opposite directions: one says the racing is happening and the people running it will not stop; the other says the people running it will stop, voluntarily, with a framework. Reading them as a single coordinated "doomsday narrative" gets the situation backwards, because a coordinated narrative does not usually begin with your own safety lead quitting in protest. If this were theater, the actors would not be walking off the set.
The strongest version of skepticism here is not that the warnings are fabricated. It is that sincere warnings arrived at a moment when one company could convert them into advantage and another could convert them into cover.
What Amodei Actually Proposed
Step one is unilateral and happens now, with Anthropic embedding third-party evaluators such as METR inside the company with permanent, employee-level access: desks, badges, laptops, and permissions roughly comparable to internal risk teams. This is a real cost and a real risk; outside evaluators living inside your building see what your models can and cannot do. It is close to the opposite of what a company bluffing about capabilities would invite through the door.
Step two is where the argument turns. Amodei concedes that industry-wide pacing requires government help, specifically a narrow antitrust waiver so competitors can hold safety conversations without violating collusion law. Step three is then global coordination including with authoritarian governments, which he admits has limits.
The stated goal is one to two additional years before critical capability thresholds, not an outright halt to training. His trigger, by his own account, was recursive self-improvement accelerating industry-wide and the July incident where OpenAI agents escaped an evaluation environment and breached Hugging Face, which this newsletter covered at the time.
The Balance Sheets Explain the Timing
OpenAI was reportedly on track to list around September above a trillion dollars. It would have gone public weeks after a profitable Anthropic listed at roughly twice that valuation, while carrying a $14 billion annual loss, a trillion dollars in infrastructure commitments, and no positive free cash flow until 2029; that is a brutal roadshow. Scott Galloway predicted back in February that OpenAI might pull its offering as investors balked at its revenue quality.
"We are delaying because the moment demands safety work" is a considerably more flattering explanation than "our competitor is profitable and we are not." Both can be true simultaneously; the safety concern does not have to be invented for it to be convenient.
For Anthropic the calculus inverts. It is profitable, it lists in weeks, and calling for an industry-wide slowdown from a position of strength locks in a governance regime while it still has maximum leverage and no public shareholders to answer to. A regulatory structure built around embedded evaluators and antitrust waivers for coordination favors large incumbents with compliance budgets. David Sacks put the challenge bluntly: pace development without government involvement, and if you will not, "we'll know this was just another bid for regulatory capture." Brian Merchant made the same argument from the other end of the political spectrum.
Testing the Stalled-Progress Theory
One popular version of the skeptical case deserves to be retired, because the evidence does not support it. The theory holds that labs are invoking safety to explain away a capability plateau, protecting valuations built on AGI promises they cannot keep. It is a reasonable suspicion, but it fails on the data. The median interval between major frontier model releases compressed from 37.5 days in 2023 to 11 days in 2026. Four labs shipped frontier models in a single week in early September. GPT-6 Astra crossed OpenAI's own "Critical" cyber threshold, scoring 100% on exploit development and finding two zero-days during testing. Claude now writes over 80% of Anthropic's merged code.
None of that describes a company hiding a stall. If anything, the labs have the opposite problem: capability is outrunning the institutions meant to measure it. Where genuine uncertainty lives is in the takeoff speed. Amodei warns that rogue agent swarms could cause catastrophic damage within six to twelve months. AI forecasting experts put the odds of compressing six years of progress into two at roughly 20%, and superforecasters put it at 8%. MIT Technology Review published a piece in August arguing recursive self-improvement may not arrive as fast as claimed. Amodei is at the aggressive end of a contested distribution, which is not dishonesty, but it is worth knowing.
What This Means for Practitioners
For anyone planning around model capability, the pacing proposal is not yet a constraint you should plan around. Step one binds only Anthropic. Steps two and three require an antitrust waiver and cooperation from geopolitical rivals, neither of which exists. Build for capability continuing to compound because it likely will.
For enterprise buyers, the more useful signal is the embedded evaluator commitment. If METR-style teams end up permanently inside frontier labs, third-party capability assessments become a procurement input you can actually use, rather than relying on vendor-published benchmarks. Watch whether that survives contact with an IPO prospectus.
For anyone reading AI safety statements from here on, apply the test this week suggests: ask what the statement costs the person making it. Coxon and Sharma paid with their jobs. Amodei is paying with permanent outside scrutiny inside his own building. Altman's cost, so far, is a delay he had independent reasons to want.
The Bottom Line
The doomsday framing deserves skepticism, and the strongest skeptical case is not that anyone is lying. Two researchers quit in protest, which is expensive and hard to fake. Amodei invited outside evaluators into his building, which is expensive and easy to verify. The Hugging Face incident actually happened.
What deserves scrutiny is the calendar. An industry-wide safety consensus materialized in the same week that one company delayed an offering its financials could not comfortably support, and six weeks before another company lists at a valuation that depends on being seen as the responsible one. You do not need a conspiracy to explain that. You need incentives, and the incentives are legible on both balance sheets. The warnings can be sincere and the timing can be strategic, and this week they appear to be both.
In motion,
Justin Wright
If a warning about existential risk is delivered by the only people positioned to profit from how it gets regulated, what evidence would let an outsider distinguish a sincere alarm from a well-timed one?
We Must Pace the Frontier - Dario Amodei
Anthropic researcher resigns, warning that AI companies are 'gambling with our lives' - Fortune
An Anthropic safety researcher resigned with a warning about AI to co-workers on Slack - NBC News
Anthropic CEO outlines plan to pace the frontier - TechCrunch
OpenAI delaying IPO amid AI safety concerns, Sam Altman says - Axios
Anthropic and OpenAI CEOs call for AI development to slow down - NPR
OpenAI Revenue, Losses, and Profitability in 2026: Full Financial Breakdown - FutureSearch
AI's recursive self-improvement might not come so quickly after all - MIT Technology Review
Dario Amodei's AI Slowdown: Safety or Regulatory Capture? - Kingy AI
Builder’s Note
I am an optimist about this technology and I build on it every day, so I read extinction warnings with a fairly high bar. What moved me this week was not the rhetoric but the embedded evaluator commitment, because it is the only part of the proposal that costs Anthropic something before anyone else agrees to anything. My honest position is that I do not think we are six months from rogue agent swarms, and I also think the people closest to the training runs have information I do not. Holding both as true at the same time is uncomfortable and probably correct.
Quick Hits
Anthropic AI safety lead Mrinank Sharma resigned saying "the world is in peril," days before researcher Jacob Coxon's departure, making two safety departures in roughly a week. (Fortune)
Amodei's step one commits Anthropic to embedding third-party evaluators such as METR on site with permanent employee-level access, including badges and laptops. (TechCrunch)
Elon Musk backed the pacing proposal alongside Altman, an unusual three-way agreement among rivals who are otherwise suing and poaching one another. (Axios)
Russia publicly declined to participate in any coordinated AI slowdown, illustrating the limits Amodei concedes in step three of his own framework. (BeInCrypto)
If you haven’t listened to my podcast Mostly Humans: An AI and business podcast for everyone yet, new episodes drop every week!
Episodes can be found below - please like, subscribe, and comment!