Underwriting AI Startups Before the Moat

The hardest thing about investing in AI startups is that the moat you are underwriting usually does not exist on the day you write the cheque. The model is rented, the data is thin, the product is a thin layer over an API anyone can call. You are not buying defensibility. You are buying a credible path to it — and most paths are mirages.

Traditional venture underwriting leans on signals that AI startups frequently lack at the seed and early stages: proprietary technology, network effects, switching costs. In their place is a fog of fast growth that may be borrowed from a temporary capability gap that closes the moment the next model ships. Underwriting through that fog requires a different discipline.

Separate the capability from the company

The first cut is brutal and clarifying: how much of this company's value is the underlying model's capability, and how much is the company's own? If the product is impressive only because the model beneath it is impressive, you are underwriting a feature that the model provider — or the next open-weight release — will likely absorb. We have watched too many "AI companies" turn out to be temporary arbitrage on a capability gap.

The companies worth backing are doing something the model alone cannot: accumulating a proprietary feedback loop, owning a workflow deep enough that the model is a component rather than the product, or building distribution and trust in a domain where being right matters more than being clever. The model is necessary; it is rarely sufficient.

If a better model would kill the company rather than strengthen it, you are underwriting the model, not the company.

The data flywheel — real and imagined

Every AI pitch invokes a data flywheel. Most do not have one. A genuine flywheel requires that usage generates proprietary data, that the data measurably improves the product, and that the improvement drives more usage — a loop competitors cannot short-circuit by buying the same off-the-shelf capability. The diagnostic is whether the data the company accumulates is genuinely scarce and genuinely useful, or merely abundant and generic.

• 01Is the data proprietary, or could a competitor obtain equivalent data through ordinary means?
• 02Does more data actually improve the product, or does the model plateau long before the data does?
• 03Does the improvement compound into distribution, or does it stay trapped as an internal metric no customer feels?
• 04How long is the loop? A flywheel that takes years to turn is a thesis, not a moat.

Price the optionality, not the projection

Spreadsheet projections for early AI companies are close to fiction, because the technology, the cost curve, and the competitive set all move faster than any model can capture. We have moved toward underwriting these investments as portfolios of optionality: what are the distinct futures in which this team wins, how plausible is each, and how cheaply are we buying exposure to the ones that matter?

This reframes the team as the central asset, which in genuinely novel markets it usually is. We are betting on a group's ability to navigate a landscape no spreadsheet can describe — to recognise when their initial wedge is closing and to have built the muscle to find the next one before it does. In AI, adaptability is not a soft factor. At this stage, it is close to the only factor.

None of this is a reason to avoid the sector. It is a reason to underwrite it honestly. The returns will be real and large. They will accrue to investors who could tell the difference between a company and a capability — and who were willing to say no to the many ventures that were only ever the latter.