Katamari strategy: applied AI firms training their own models
The observation flips the usual assumption that startups should stay thin wrappers over frontier APIs. Founders deciding between fine-tuning in-house and riding foundation model progress can use the three-part test: domain clarity, training budget, and demand density. It also hints at where vertical AI moats may erode as base models improve.
Sara (saranormous) argues that many applied AI companies are pursuing a katamari strategy of growing until model training becomes viable. The threshold requires three things: enough scale to know which capabilities are worth improving, the budget to train and win in a specific domain, and immediate customer demand to amortize the training investment. The post frames vertical AI companies as increasingly competing with foundation model providers on their own turf.
a lot of applied AI companies were playing katamari - get big enough that you know which capabilities are actually worth improving, can afford to train (and win) win your domain, and have immediate demand against which to amortize that investment
