The Hidden Risk of Picking a Single AI Lab

Stan Hulu, ex-Stripe and OpenAI, built Dust as a model-agnostic platform, betting no single AI lab will dominate. He compares buying AI from one lab to factory machines that only run on that supplier’s energy—a dangerous lock-in. With model leadership shifting rapidly, agnostic architecture prevents vendor risk and forced migration for engineering leaders.

The product landscape is converging toward horizontal platforms. Vertical AI scaffolding that once propped up weak models loses value as intelligence commoditizes. Defensibility now relies on network effects, not model-specific tuning. Dust’s move from flat to credit-based pricing mirrors industry pressure: unpredictable token consumption kills fixed-price models.

Tradeoffs are sharp: labs enjoy 70-80% margins but open source will recalibrate. Dust must rely on labs’ APIs and thin margins. The interview warns against overfunding—the “coffin corner” where companies can’t realize raised capital’s value. Leaders should evaluate tooling that decouples model choice from product and budget for variable usage, not fixed licenses.

What most people miss: the real signal isn’t just model agnosticism, but the dying value of vertical AI scaffolding. The hype is the vague “work will feel like not work” prediction. The evidence is largely anecdotal but logically consistent with market dynamics.