Most AI product teams treat model selection as an infrastructure decision. Lin Qiao, CEO of Fireworks AI, wants them to treat it as a governance decision.

The Mythos shutdown (already reported here) handed Qiao a pointed illustration. A company whose product ran on intelligence it did not control found itself subject to decisions made by others, with no practical recourse. Qiao posted the argument on X on June 16, framing the episode not as a procurement failure but as a preview of a structural exposure that applies broadly to any product built on a rented foundation.

The core claim is a substitution of terms. The AI industry has spent two years debating open versus closed models through the frame of cost. Qiao’s reframe shifts the axis: the question is not what a hosted model charges per token, but who makes the decisions that determine whether your product functions tomorrow. Closed, hosted frontier models sit inside vendor roadmaps, vendor safety policies, and vendor business viability. An open model that a team has fine-tuned and runs on its own compute sits inside the team’s roadmap. Those are different risk profiles, not just different price points.

Fireworks cites customers Ramp, Cursor, and Harvey as validation that a tuned open model can match frontier quality at meaningfully lower cost. That evidence deserves the label it carries: these are Fireworks references. The company sells open-model inference, which means every customer who moves from a closed API to a hosted-open or self-managed deployment is a potential Fireworks customer. Qiao’s argument is structurally correct in its framing and commercially convenient in its conclusion. Both things are true simultaneously.

That conflict does not make the underlying question wrong. The own-versus-rent divide in AI is real, and the asymmetry of control is real. A startup that builds its user-facing experience on a single closed API is exposed to that vendor’s pricing changes, model deprecations, safety interventions, and the possibility that the vendor competes directly in the same market. Teams at Ramp and Harvey presumably evaluated that exposure before choosing to tune and deploy open models; the Fireworks sales pitch did not invent the risk, it responded to it.

The decision frame for operators is narrower than Qiao’s argument implies. Full model ownership requires ML infrastructure, fine-tuning capability, and ongoing evaluation work that most product teams cannot staff or fund. The more actionable question is: what is the minimum model independence your product needs to survive a bad vendor decision? For some products, the answer is prompt portability and the ability to switch providers in days. For others, particularly those in regulated industries where model behavior must be auditable and stable, the answer is something closer to genuine ownership of the weights and the inference stack.

Qiao’s post is advocacy dressed as analysis. The advocacy is worth reading anyway, because the stakes it describes are not invented. Any team currently committed to a single closed API without a tested migration path is carrying exposure they have probably not formally priced.

Lin Qiao, CEO of Fireworks AI, posted the original argument on X on June 16, 2026. Source URL: https://x.com/lqiao/status/2066403957824688462.