Why regulated firms go private
The trigger is usually a data classification policy: material marked strictly confidential may not leave the corporate network, whatever the vendor's contract says. Banks and wealth managers carry client confidentiality duties, pharmaceutical companies guard pre-launch data, defence and public sector bodies work under security classifications, and many firms have promised their own clients where data will sit. Financial entities in the EU also answer to the Digital Operational Resilience Act, applying since 17 January 2025, which requires them to manage ICT third-party risk, register ICT service contracts and plan exits from services supporting critical or important functions. Marketing content is in scope more often than expected: unreleased products, customer data used for personalisation and deal communications.
VPC or on-premise
In a VPC deployment the vendor's software runs inside the customer's own cloud account. The customer controls the network boundary, encryption keys and logs, while cloud GPUs provide capacity that scales with demand. On-premise means the customer's own data centre, possibly air-gapped from the internet. It gives the most control and the most work: hardware with long procurement lead times, capacity planning, patching and monitoring all sit with the customer or a managed service. Between the two sits single-tenant hosting, a dedicated environment run by the vendor. A hybrid often makes most sense, sending confidential jobs to private models and routine work to hosted models under zero-retention terms.
Self-hosted models: what you give up
Self-hosting means running open-weight models on your own infrastructure. For many text and image tasks they are strong, but the most capable models in some categories are offered only as hosted services, and video generation is the most GPU-hungry medium to run yourself. Read each model's licence before commercial use, because open weights do not always mean unrestricted use: some licences add usage thresholds or acceptable-use conditions. You also inherit work a vendor would otherwise do: evaluating new versions, deciding when to upgrade, and running the safety filters that stop a model producing content the brand would never publish.
Cost and performance trade-offs
Private capacity is paid for whether it is used or not. Marketing demand is spiky, with launches, seasonal peaks and Black Friday, so GPUs sized for the peak sit idle for much of the year, while hosted services spread that cost across many customers. Latency can favour local deployment for small jobs and hosted capacity at peaks. The sensible approach is to route by confidentiality rather than make everything private. Synthetic White's orchestration layer picks the model per job by format, fidelity, confidentiality and cost; confidential work can run on private or self-hosted models, and deployment is available in the cloud, a private VPC or on-premise.
Updated 25 September 2026