The technical lead that OpenAI and Anthropic hold over the rest of the field might be smaller than their valuations suggest: roughly one or two model generations, according to an essay by independent researcher Rohit Krishnan on his newsletter Strange Loop Canon. That gap, he argues, is the entire moat, and it is shrinking as open-weight labs in China close the distance.
Krishnan’s case starts with economics rather than capability. Training runs, data acquisition and talent now each cost frontier labs on the order of a billion dollars, a bar high enough to keep out casual entrants but not high enough to keep out the ambitious ones. He points to Meta’s roughly $90 billion metaverse bet as evidence that plenty of capital is willing to chase a speculative frontier once the potential payoff looks large enough.
Meanwhile the buyers of that intelligence are getting pickier. Krishnan describes finance teams increasingly scrutinizing AI spend as a new operating expense line, pushing companies to route routine work to cheaper models and reserve frontier-tier access for a smaller set of high-value tasks. He writes that he personally stopped using the most expensive tier of any model once mid-tier options became “good enough,” a shift he treats as an early signal of a broader pattern: when the smartest available model is only marginally better, most spending migrates toward whichever option is cheapest per task.
That dynamic, Krishnan argues, is why OpenAI and Anthropic are pushing outside pure model-building. OpenAI is building an advertising platform on top of its consumer user base and has re-entered robotics. Anthropic has moved into wet-lab biology research aimed at drug discovery. Neither move is really about intelligence for its own sake; both are attempts to convert a temporary model lead into a durable business before rivals close the technical gap. He is skeptical this works cleanly: few industries generate the $100 billion-plus in annual revenue the frontier labs already command, and any new market a lab enters gets crowded fast, since rivals see the same opportunity.
The bigger prize in Krishnan’s telling is recursive self-improvement, an AI system capable of designing its successor well enough to compound gains far faster than human researchers can. He calls this the “true unknown unknown” and argues it would justify the labs’ spending even if their consumer products never became conventionally profitable. But he is explicit about his own doubts: models solving formal problems like the Navier-Stokes equations have not yet translated into measurable disruption of white-collar employment, which he takes as evidence that raw intellectual capability and real-world usefulness are not the same axis. He expects labs will build systems that are superhuman at narrow tasks well before, if ever, they build one that is superhuman at everything.
For operators, the practical read is not to bet on a permanent capability gap holding up pricing or product strategy. If the moat really is only one or two generations wide, the safer planning assumption for any team building on frontier APIs is that today’s performance edge from a single vendor will compress within twelve to eighteen months, and that a company’s own product differentiation, not the underlying model, should be doing the defensive work.
Rohit Krishnan detailed this analysis in an essay published on Strange Loop Canon on September 21, 2026.
