On Saturday, Dario Amodei published thirty eight hundred words on his personal website arguing that the AI industry must slow the rate at which it improves model capabilities. It is called We Must Pace the Frontier. Within hours Sam Altman agreed on X. Elon Musk and Demis Hassabis fell in behind. By Sunday, Donald Trump had rejected the whole thing from Ireland, which is roughly where we are as a civilisation.
Let me be precise about my objection, because I am not calling safety concerns fake. Some of them are serious, and an Anthropic researcher resigned last week warning that the people building this believe it might kill us all by the end of the decade, which is not a sentence anyone should read casually.
My objection is narrower and, I think, harder to dismiss.
You cannot credibly ask the world to slow down a thing you have not yet demonstrated works and delivers real ROI.
The receipts
Start with what actually happened, rather than what was announced.
Not one of the four labs has reduced its model release cadence. That is the gap between the rhetoric and the behaviour, and it is the only measurement that matters here.
Now look at what they did instead, in the same period they discovered their conscience. Anthropic closed sixty five billion dollars in May at a nine hundred and sixty five billion post money valuation, filed for IPO on the first of June, and by the eleventh of September was reported to be seeking up to a hundred billion more at close to two trillion. OpenAI closed a hundred and twenty two billion in March at eight hundred and fifty two billion and is targeting a trillion.
So the sequence, for anyone keeping score. Raise a hundred billion. Seek a two trillion dollar valuation. Publish an essay about restraint. Receive enormous credit for moral seriousness. Do not change the release schedule.
I have watched a great many commercial narratives in twenty years and that is among the most elegantly constructed I have encountered.
And we have seen this film. In 2023 the Future of Life letter asked for a six month pause and the Center for AI Safety compared the risk to pandemics and nuclear war. Altman, Amodei and Hassabis all signed. Nothing stopped. Same cast, same concern, three years later, and the only material change is that they are all now worth considerably more.
Where is the thing you sold me in 2023?
Here is the part nobody in that conversation wants to address, and it is the commercial question rather than the safety one.
The 2023 promise was specific. This would transform enterprise productivity, collapse cost structures, and create value at a scale that justified any capital expenditure. On that basis, roughly seven hundred and twenty five billion dollars of capex was committed this year alone by four companies, and a hundred and eighty-five thousand people in leading technology firms lost their jobs in six months while analysts openly modelled the freed cash flow as GPU funding.
And the return? Massachusetts Institute of Technology looked at three hundred enterprise AI initiatives and found ninety five percent achieving zero measurable return. Morgan Stanley found only twenty one percent of S&P 500 companies could cite any measurable benefit at all. Sixty one percent of enterprise AI projects were approved on projected value that was never measured after deployment.
I want to be fair here, because the lazy version of this argument is wrong. The models genuinely have improved. Capability has moved enormously since 2023, and anyone claiming otherwise has not used them.
However capability is not value, and the industry has spent three years conflating the two. What has not arrived is the economic return. Not delayed. Not partially realised. Substantially unmeasured and, where measured, largely absent.
So when the same men now say the technology is advancing so fast it may outrun our control, I notice that this claim and the missing ROI cannot both be comfortably true at once. Something advancing faster than our ability to comprehend it ought, somewhere, to show up in a profit and loss statement.
The shiny new model as a covering manoeuvre
There is a pattern here worth naming, because it is a commercial mechanic rather than a technical one.
Every model launches as a step change. Within months the limitations surface, the enterprise deployments stall, the ROI does not materialise. And precisely then, reliably, a new model appears with a new benchmark and a new set of promises, and the conversation resets before anyone finishes auditing the last one.
The next release is not merely a product launch. It is a narrative reset, and it arrives at exactly the moment the previous narrative was about to be tested.
Meanwhile, down here where the invoices land, the experience is rather more modest. The thing cannot reliably tell me the time. It apologises with great sincerity for having just done the opposite of what I asked, then bills me tokens for the apology. It hallucinates a citation, I correct it, it thanks me warmly for the correction and does it again four messages later.
I am not saying these tools are useless. I use them daily and they have genuinely changed how I work. I am saying the gap between what is being described at a two trillion dollar valuation and what arrives in the window is wide enough to drive a data centre through.
Could I have the value from the last one before I am required to be excited about the next one?
The opacity is the tell
Regular readers will recognise where this lands.
The monetisation economics of these companies are as obfuscated as their pricing models, and that is not a coincidence. It is the same failure viewed from two sides.
A company that could measure the value it delivers would price on that value, because outcome pricing is commercially superior wherever it is achievable. A company pricing on outcomes would necessarily know its cost per outcome. So when a vendor cannot tell you either, they are not protecting a trade secret. They are telling you the instrument does not exist.
Nobody obfuscates a number they are proud of.
Which is why I take the safety essays as commercial documents as well as moral ones. A narrative about imminent superintelligence is, among other things, an excellent reason not to discuss unit economics. It is very difficult to ask a man about his gross margin while he is warning you about the end of the world.
The actual catastrophe
So let me put the contrarian position plainly, because I think the entire risk conversation is pointed in the wrong direction.
The thing most likely to cause serious economic damage in the next three years is not a recursively self improving system escaping human control. It is the correction that arrives when several trillion dollars of capital investment meets three years of unmeasured returns.
Consider what is actually exposed. Pension funds hold this equity. Sovereign wealth is committed across the Gulf and beyond. Utility-scale power infrastructure has been built on demand forecasts underwritten by these companies. Employment has been permanently restructured to fund the capital expenditure. The debt market has already started pricing the risk, with a thirty basis point spread penalty now attached to AI adopters versus enablers.
That is not a technology risk. It is a capital allocation risk, and it is the one nobody is publishing essays about.
If this deflates, it will not be Terminator. It will be pension statements, redundancy programmes at companies that already made them, stranded data centre assets in jurisdictions that borrowed to build them, and a decade of underinvestment in whatever comes next because a generation of capital got incinerated proving a business model nobody bothered to measure.
The dark ages, if they arrive, will be delivered by a spreadsheet.
The position
Pace the frontier, by all means. I have no objection to embedded evaluators, common standards or international coordination, and Amodei's three steps are considerably more concrete than anything in 2023.
However while you are pacing, pace something else as well.
Show me the value from the last three model generations before you ask me to be frightened of the next one. Publish your cost per outcome. Price on what you deliver rather than what you consume. Let a customer audit whether the thing worked.
Because right now four men who have never once slowed down are asking for credit for wanting to, at valuations that require them not to, while the product they sold the world in 2023 remains, by the industry's own measurement, unproven.
Slow down if you must. Deliver something first.
First published in the Agentic Commercial Model newsletter on LinkedIn, September 14, 2026. Read the original on LinkedIn.