The most lucrative idea in AI right now is a pause. And it has remarkably little to do with safety.

In early June, Anthropic proposed a coordinated, verifiable pause on frontier-model training. Three days earlier, the company had confidentially filed for its IPO. The stated trigger: fear of recursive self-improvement, the point at which an AI system can design its own successor with little human input.

First mockery, then their own watered-down version

Sam Altman's first reaction was mockery. The gist: you build the bomb, then sell the shelter. A week later, OpenAI published its own version. In “Built to benefit everyone” it writes that it wants to make it possible for the world to take coordinated action, “including slowing frontier development when needed”. The catch: neither proposal commits anyone to slow down by a single day. The language is conditional and unenforceable by design.

Why it is brilliant: a look at the leaked numbers

How fitting that OpenAI's leaked 2025 financials landed in the middle of this debate, first reported by Ed Zitron and verified by the Financial Times. They tell two stories at once.

First: the product makes money. Around 13.1 billion dollars in revenue against roughly 7.5 billion dollars to serve it. That leaves about 5.6 billion dollars in gross profit, a margin of 43 percent. Second: what actually costs billions is staying ahead. Research and development alone consumed 19.18 billion dollars, more than the entire company's revenue. The result is an operating loss of around 21 billion dollars.

One detail softens that pretty margin: of the 19.18 billion in R&D, roughly 10.6 billion went to Microsoft, for training on Azure. Part of that runs through the equity arrangement rather than paid-out cash, so the real cash margin is softer than the 43 percent suggests. It barely changes the core point: serving customers pays for itself, the race is the loss driver.

And this is exactly where the pause debate becomes a business move. It tells Wall Street the most convenient story imaginable, just ahead of two of the largest IPOs in history: the engine underneath is profitable. We only burn billions because we choose to sprint. The race is a choice, not a constant. Whether a pause ever happens is then almost beside the point. Setting the narrative is already the win.

"That these models already pay for themselves says more about the value of AI than any hype: it is real and already being paid for. And the best part: we are only at the beginning.", Thorsten Vellmerk

What this means for you as a decision-maker

For all the skepticism about the timing, the numbers carry an honest message. The business model is profitable, and there is a reason for that: you can genuinely achieve a lot with artificial intelligence. Otherwise customers would not pay 13 billion dollars a year for it. The expensive part is the race for the next frontier model, not the productive use of what already works reliably today.

"The interesting question for companies is not who wins the race at the top. It is how much of what already works profitably today still sits unused in your organization.", Thorsten Vellmerk

And where do you start when the options are this varied? A simple map helps: the Vellmerk Matrix sorts AI initiatives by impact and effort, so you begin with the steps that pay off most.

If reading this leaves you with the feeling that more should be possible with AI in your organization, in your company too: that is exactly where we start. Talk to me, and in a no-obligation conversation we will find where AI offers the fastest and safest leverage for you.

Sources