Open weights, open questions: how open-source models are reshaping the stack

The rise of genuinely capable open-source models has introduced a real alternative to the closed-API default, and it is quietly changing how teams assemble their software. When powerful models can be downloaded and run on one's own terms rather than only rented through someone else's interface, the entire calculus of building with AI shifts in ways the industry is still working through.

Openness dramatically lowers the barrier to experimentation. Teams can inspect a model's behaviour, fine-tune it on their own data, and deploy it without per-call fees or dependence on a provider's roadmap, pricing and availability. For many organisations, that combination of control and freedom is genuinely transformative, opening doors that a metered API kept firmly shut.

Open weights are free to download and genuinely expensive to operate responsibly.

The build-versus-buy reset

Open weights reframe a foundational decision that most teams thought was settled. Where they once defaulted reflexively to a managed API as the obvious choice, they now weigh the control and economics of self-hosting against the sheer convenience of a service that someone else operates and maintains. Neither answer is universally correct anymore, and the choice has become genuinely strategic rather than a foregone conclusion.

Cost is a significant part of the recalculation. At sufficient scale, running an open model on your own infrastructure can be dramatically cheaper than paying a provider per token for every request — but only if the organisation actually has the considerable expertise required to operate it well, which many badly underestimate going in. The savings are real but conditional.

The operational burden

Self-hosting simply trades one set of problems for another, less visible set. Running models in production raises real and ongoing questions about infrastructure, monitoring, scaling, updates and security that a managed service otherwise quietly handles on your behalf. Open weights are free to download and genuinely expensive to operate responsibly, and the gap between those two facts catches many teams off guard.

Security is a particular and growing concern that the enthusiasm sometimes obscures. Models pulled from open ecosystems must be carefully vetted, and the supply chain for both the weights themselves and the tooling around them represents a new attack surface that defenders are only beginning to map and understand. Convenience and openness can each become a liability if treated carelessly.

A strategic axis

Openness has become far more than a licensing detail to be skimmed in the fine print; it is now a genuine strategic dimension of how companies build with AI. The decision to lean open or closed shapes cost, control, capability, security and risk all at once, and getting it wrong in either direction carries real consequences for the business.

The likely long-run outcome is not a clean victory for either camp but a durable, pragmatic coexistence: closed services where their convenience and capability decisively win, open models where control and economics matter more, and a generation of teams that learn to be genuinely fluent in both and to choose deliberately between them.