The Business of Building God

The big AI labs are no longer just selling magic. They’re acting like companies that need durable revenue, and that shift changes the whole argument. OpenAI is talking up an ad platform tied to a huge consumer base, Anthropic is chasing wet labs and cancer work, and both are heading toward IPO territory. The old story was that better models would justify anything. The new one is whether the business model can survive contact with accounting.

The problem, as this view sees it, is that the frontier labs may only be one or two models ahead of everyone else, including open source. That gap is the moat. It helps that they have more talent, more compute, and access to the traces people leave behind while coding, working, and chatting. But that data is not rare forever, and if competitors can train on enough of the same material, the edge starts to look less like a fortress and more like a head start.

So the labs may have to expand into other businesses just to keep the revenue story alive. The source points to ad tech, robotics, and wet labs as examples of this push. But that route is brutally hard. Most industries do not throw off $100 billion in revenue, and when companies build their own model-based services, the market fragments instead of concentrating. Meanwhile, other model makers can see the same prize and spend toward it too. The source cites $1 billion training runs, plus another $1 billion each for data and talent, and says plenty of firms can still play that game.

The pressure is also coming from customers. AI token spending is already at more than $200 billion, and CFOs are starting to ask what they are actually buying. That’s why companies are already limiting top models to top users and pushing the rest to cheaper options. The fancy version can win the demo. The cheaper one wins the budget.

Underneath all of that is the real wildcard: recursive self-improvement. If the labs can build AI researchers that improve the next generation of models, the business gets stranger and the moat gets stronger. If they can’t, then the frontier may keep advancing in bursts without ever turning into the kind of all-purpose machine people keep pricing in. Either way, the labs may build some very powerful systems. Turning that into a lasting business is the harder trick.