AI Pulse · Business Models & Winners

Garry Tan is arguing that US open-weight AI labs should focus more on distilling frontier models rather than training every model from scratch. Put simply, distillation is the process of using a stronger model’s outputs to help train a smaller or more focused model. The argument is about cost structure. If a capable model can be created with less training spend, it may be easier to offer useful AI capability at a price customers will actually pay. You can read the framing in the original report.

For agency owners, consultants, and software businesses, this matters because many are effectively reselling model capability. You may wrap an AI model in a workflow, a service, industry knowledge, or a customer-facing product. Your margin depends on what the underlying capability costs, how reliably it performs, and whether you need the most expensive model for every task.

The practical question is not whether you should build or train a model. Most businesses should not. It’s whether each part of your offer needs frontier-level intelligence, or whether a smaller, cheaper model can handle repeatable work once the process is clear. The firms that separate those jobs well should have more room to price competitively without giving away margin.

This is the kind of thing we build into an AI command centre, tracking which AI capabilities are becoming cheaper, where quality is good enough, and what that means for the economics of your offer.