Frontier vs Local Models

Cloud API access to frontier models vs local deployment of open-weight models.

Dimensions of Comparison

Dimension	Frontier (Cloud API)	Local Deployment
Capability	Highest (hundreds of billions of parameters)	Reduced (typically 7B-70B)
Data sovereignty	Data leaves site	Data stays on-site
Safety filters	Provider-controlled, may suppress nuclear content	Deployer-controlled
Latency	Network + inference	Inference only
Cost	Per-token pricing	Hardware investment
Auditability	Limited (provider-side processing)	Full (prompts, outputs visible)
Air-gap compatible	No	Yes
Availability	Depends on provider uptime	Local control

Nuclear Considerations

Data sovereignty is critical: plant operational data (parameters, procedures, maintenance records) must not leave the site boundary in most regulatory frameworks.

Safety filter suppression: Provider-side filters trained for consumer use may block discussion of radiological hazards, criticality accidents, or radiation exposure calculations. No indication is given that content was removed — the operator receives an apparently complete but silently edited response (Report 1 §7.2).

Auditability: For incident investigation, understanding why an AI system produced a specific recommendation requires access to the full prompt and reasoning chain. Cloud API providers may not expose this.

Recommendation

Local deployment is preferable for nuclear applications due to data sovereignty, auditability, and filter control. Quantisation (4-bit) enables 70B models on single-GPU servers with acceptable accuracy for many tasks. Frontier models remain useful for initial evaluation (Level 0 of capability-gradient) and for tasks where maximum capability is needed and data sensitivity permits.