The week's loudest AI story reads like a morality play, but the more useful way to view it is as a positioning problem. Dario Amodei published an essay this past weekend urging the industry to ease the pace of capability advances, roughly a month before Anthropic is expected to list on the Nasdaq in what insiders say could be a $2 trillion debut. The company confidentially filed for its IPO on June 1 at a private valuation of $965 billion. So we have the chief executive of the fastest-growing AI lab in history calling for restraint at the exact moment he is asking public markets to pay for unrestrained growth. That contradiction is worth sitting with, because both halves are probably sincere, and that is what makes it interesting.
The slowdown and the sales pitch are the same document
Amodei's proposal is not vague hand-wringing. He laid out a plan that includes giving third-party evaluators access to frontier models, agreeing shared safety standards across labs, and, where possible, coordinating with authoritarian governments. Read commercially rather than morally, that plan describes a moat. If frontier labs collectively agree to pace releases and route evaluation through a small set of approved institutions, the winners are the incumbents who already sit at the frontier. Sam Altman and Elon Musk reportedly nodded along, which is less a sign of shared conscience than of shared interest. A pacing agreement among three leaders is a cartel by another name, and cartels tend to protect the firms already inside them.
The human drama gave the argument its fuel. Jacob Coxon, a 27-year-old researcher who spent three years on pretraining at both OpenAI and Anthropic, resigned on September 8 and wrote that neither company is acting responsibly and that both are "gambling with our lives." His thread drew tens of millions of views within a day, and within 24 hours two more safety researchers went public. I take Coxon at his word; people who leave frontier jobs and torch bridges on the way out are usually not running a marketing play. But the same event can be authentic for the person living it and convenient for the company he left. Anthropic has built its entire brand on being the careful lab, and a viral departure that frames the whole industry as reckless does not obviously hurt the firm whose pitch is caution.
The real fragility is in the revenue, not the robots
Existential risk chatter is absorbing attention that the numbers deserve more. New data from Ramp shows the top 1% of customers generate roughly 80% of enterprise revenue at both OpenAI and Anthropic, a concentration Ramp's lead economist calls unlike anything in any other software category it tracks, and one that has barely moved in three years. At Anthropic the concentration has names attached: coding tools Cursor and GitHub Copilot together drove around $1.2 billion of the company's revenue, close to a quarter of the total, and GitHub is owned by Microsoft, which has also put $13 billion into OpenAI. So a meaningful slice of the AI economy is AI companies buying from other AI companies, with more than a trillion dollars of compute commitments stacked on top of a demand base you could fit in a mid-sized conference room.
That is the number that should worry an investor pricing a $2 trillion listing. The heaviest spenders are precisely the customers most able to switch to cheaper open-source models or build in-house, and reporting this summer has documented large enterprises already leaning that way. A frontier slowdown, whatever its safety merits, conveniently slows the pace at which cheaper open alternatives can close the gap.
The physical world is quietly voting no
While the labs debate the speed of intelligence, the grid is imposing its own limit. Texas, which its own governor called the epicenter of AI development a year ago, paused approvals of new data center connections on August 3 after interconnection requests hit roughly 474 gigawatts, more than five times the state's record peak demand. Somewhere between 250 and 300 projects now face an audit. This matters more to the near-term trajectory than any manifesto, because compute is already the binding constraint, and both OpenAI and Anthropic were compute-limited before Texas pumped the brakes. When the scarce input is power and silicon rather than ambition, a call to pace the frontier costs the leaders very little. They were going to be paced anyway.
So what
The honest read is that Coxon's warning and Amodei's essay are both real and neither settles the question a business audience actually cares about. Whether superintelligence arrives by 2030 is unknowable and, for now, unpriceable. What is priceable is a near-trillion-dollar company heading to market on revenue that leans on a handful of accounts, against compute commitments that dwarf that demand, into a power grid that has started saying no. If you are weighing the Anthropic IPO, the safety debate is the part you can safely set aside. The concentration risk is the part that will show up on a balance sheet.