Why domain expertise alone is no longer a moat
A field report from three weeks in Silicon Valley frames the market's condensation into frontier labs along two axes: Time Gap × Domain Gap. The time gap is closing fast, the observation goes, leaving only the domain gap. So far this is the familiar conclusion—everyone can use the models, so whoever knows the domain deeply holds the moat.
But the frontier labs have declared they will do those very domains themselves: coding, legal, bio. Even their criterion for picking domains is explicit—they go first for domains where RLVR (reinforcement learning with verifiable rewards) can run. The method is blunt: they buy experts' tacit knowledge as post-train data, paying to lift the judgment that lived only inside people's heads and pour it into the model.
At that point domain expertise becomes not a moat but an acquisition target. The moment twenty years of judgment is sold as a dataset and absorbed into a model, simply 'knowing the domain' stops being a defensive line.
How deep has AI already gone into domains?
Bio is where this shift shows most dramatically. Amid persistent skepticism from traditional researchers, OpenCRISPR is offered as proof that AI has already produced real innovation in the domain—AI designed a novel gene-editing tool.
The more striking scene is at the individual level: someone built a personal genome-analysis pipeline in three days with Codex. Work that once required years of accumulated domain knowledge and engineering compressed into three days for someone holding the right tool. It is a signal that the cost of entering a domain is itself collapsing.
The question flips—who closes the gap first?
So the report poses a provocative frame: persuade the domain expert, or beat them. This inverts the direction we have been asking about. Until now we asked, 'will the domain expert learn AI?' But the real contest may be on the other side.
Two speeds compete: how fast a domain expert grows fluent in AI, and how fast someone holding AI catches up on domain knowledge. In a world where domain knowledge is bought as data and reproduced in three days by a tool, the latter speed is far faster than expected. The moat is decided not by 'what you know' but by 'who closes this gap first.'
So what is the moat now?
Follow this view and the definition of a moat changes—from a static asset (domain knowledge, model access) to a dynamic speed (the ability to close a gap). What you know gets replicated and bought; the habit of closing your own gap quickly does not replicate so easily.
From an AX (AI transformation) standpoint this leads to concrete prescriptions. If you are a domain expert, put the speed of translating the flow you know into the language of AI first. Refuse the tools while knowing the flow, and your knowledge gets sold off as post-train data while you alone fall behind. If you are AI-native, humility toward the domain becomes your moat—you may build a pipeline in three days, but what that domain truly fears can still only be learned in the field.
I do not yet have a settled answer to this question. But the sense that the moat is moving from ownership to speed is unmistakable. Which side closes the gap first—that result is what I am curious about.