Open weights can justify an investment before they justify a saving. Depending on the licence, architecture and deployment route, they can widen an organisation's options over where a model runs, who operates it and how it might change course. Those options may create structural value. They do not demonstrate lower operating cost. Leaders should approve the control they need on its own merits, then test any cost hypothesis against one production workload and every material responsibility required to sustain it.
Control changes the operating boundary
Open weights are model parameters made available under a licence that permits specified forms of use. That is a set of model rights, not ownership of a service and not a hosting model. The same weights might be served through a managed endpoint, a specialist operator, a hybrid arrangement or infrastructure run by the enterprise. The licence defines permitted use; the deployment route determines where data, infrastructure and third-party dependencies sit; the operating model assigns reliability, security, evaluation, lifecycle, incident and support duties.
That separation matters because control and cost follow the actual arrangement. A managed provider may perform work that is embedded in its charge. If another route moves duties inside, external charges may change while platform, compute, monitoring, on-call, upgrade, fallback and specialist work appears elsewhere. Some arrangements also commit capacity before demand is clear or concentrate effort around incidents and model changes. Access price alone cannot reveal the result. Compare a business task completed at the required quality and service level, including all material human, platform and third-party inputs under each option.
Keep four forms of value separate. Realised cash would require an observed, attributable net movement in money paid or received after the new operating cost. Capacity is measured time released or consumed; it does not become savings through salary multiplication. Structural value may arise from a durable deployment choice or a tested switching option. Any future revenue or scale effect remains modeled upside until recognised and attributed. These categories support different decisions and must not be summed into a headline benefit.
Make two decisions, not one
The decision is to write a control case and a cost hypothesis separately. The control case should state which rights the licence grants, which deployment route will be used and who will own reliable operation. It should name the boundary or option that matters, assign security, evaluation, lifecycle, incident, fallback and support accountability, and explain why that control merits its full price. It must not borrow a speculative saving to make the rationale look stronger.
The cost hypothesis should hold one production workflow constant from request to accepted outcome. Define the comparison baseline, required quality and service level, demand assumptions, material inputs and observation period. Estimates can inform approval when assumptions are visible. Observation is required before the comparison becomes an outcome claim. “We can move later” is not a switching plan: identify interface, data, tooling, evaluation and process dependencies, then test the path at a depth proportionate to its importance.
What would count as proof?
Proof starts with a responsibility map for each actual option. Every material duty needs an owner and an input, whether it sits with an internal team or a third party. Apply the same workflow, accepted-completion definition, demand profile, quality threshold and service requirement to the alternatives. Record infrastructure use, operator effort, review and exception work, completed tasks, reliability, incidents, recovery and the work needed to evaluate and introduce changes.
Test control as well as service. Confirm that the deployment boundary meets the stated requirement. Where exit is part of the approval rationale, rehearse switching at least once, with depth matched to consequence. Keep estimates distinct from observations. Any cash conclusion needs a baseline, timing, attributable net money movement, all added operating inputs and protection against double counting. Capacity, structural value and modeled upside retain their own evidence and approvals.
What remains unclaimed?
Access to weights alone does not establish residency, independence, portability, lower concentration, reliable service, easy switching or lower cost. A technical deployment shows that a model can run in a chosen arrangement; it does not show that the arrangement can sustain an accepted business outcome. A projected unit cost is still an estimate, and released time remains capacity rather than cash. The defensible conclusion is narrower: open weights can expand an option set under specific licence and operating conditions. Whether that option creates structural value or a financial result must be learned from the selected workload.