Bring Your Own Compute
Could an application use AI compute you choose, while its authorization broker stays outside the plaintext inference path?
The application supplies the workflow. Who supplies the compute?
Useful AI models can increasingly run on personal computers, workstations and rented infrastructure. Applications, however, commonly assume that inference comes from a provider selected by their developer.
This concept paper explores User-Delegated Bring Your Own Compute: a user grants an application scoped permission to use AI compute that the user controls or selects.
The switchboard analogy
A telephone switchboard establishes a connection; it does not need to participate in the conversation. This deliberately limited analogy captures the separation between coordination and inference.
Identity, consent, discovery and revocation can belong to a broker. Prompts, attachments and responses travel between the application and authorized compute. An encrypted relay may still be necessary. The important property is who can see plaintext, rather than whether every connection is direct.
Ask for capabilities
An application could request vision, tool use, structured output or a particular context capacity. A user-controlled layer could match those requirements to an approved environment.
That might be personal hardware, an organization’s infrastructure or rented capacity. The choice depends on the task, cost, speed and acceptable data location.
No silent cloud fallback
An unavailable local machine does not automatically authorize sending data to another provider. Local preference, local-only processing and approved cloud processing are distinct policies.
Local AI is not inherently secure. Maintenance, access controls, scoped authorization and revocation remain necessary parts of the design.
An interoperability question
The paper proposes neither a new networking stack nor a finished protocol. Many relevant building blocks and overlapping products already exist. Its question is whether they can support a broadly adopted, application-agnostic way to delegate compute.
Concept paper v0.1 is dated 30 September 2026. Its comparison with existing systems is a snapshot. The original discusses prior art, boundaries and unresolved technical questions in detail.