How do banks run air-gapped AI with Plugsky?

Key facts

Definition	No external network dependency; models and runtime operate fully inside the perimeter
Deployment	On-prem and air-gapped tiers alongside cloud and VPC options
Models	Open-weight models suitable for offline operation, including Llama, Qwen, Mistral and Nemotron
API contract	OpenAI-compatible endpoint in all tiers; application code does not change
Key control	Customer-managed keys through KMS or on-prem HSM for a revocable kill switch
Updates	Delivered through your controlled intake and verification process on your cadence
Audit	Identity, key and inference events exportable to internal SIEM
Status	Live platform; enterprise deployment terms per engagement

TL;DR

• Air-gapping removes egress entirely; models and runtime stay inside your network.
• Plan update intake like any other controlled software supply chain.
• Keep keys in your KMS or HSM and separate key duties from platform administration.
• The API stays OpenAI-compatible, so applications do not change between tiers.
• Budget for GPU capacity and redundancy, not just licences.

How it works, step by step

1. Confirm the workload genuinely requires an air gap; many banking controls are met by a VPC or on-prem deployment with no egress.
2. Size GPU capacity for the models and concurrency you need, including headroom for peaks and failover.
3. Install the runtime inside the perimeter and load verified open-weight models through your controlled media process.
4. Register keys in your KMS or on-prem HSM and separate key administration from platform administration.
5. Keep the application integration unchanged: point it at the internal OpenAI-compatible endpoint.
6. Define the update intake process — provenance, scanning, staging and rollback — before the first patch arrives.
7. Stream identity, key and inference events to internal SIEM, and rehearse recovery: model restore, key restore and node failure.

Try it yourself

Open the private LLM deployment estimator →

What air-gapped AI means in a bank

Air-gapped AI means the inference stack has no network path to the internet: model weights, runtime and updates arrive through controlled media, and prompts, completions, embeddings and logs never leave the perimeter. It is the strongest interpretation of data residency and control, and it suits trading analytics, fraud investigation, internal knowledge search over sensitive dossiers and workloads on isolated network segments.

The practical benefit of Plugsky in this mode is continuity: the internal endpoint is OpenAI-compatible, so the same application code, SDKs and prompt tooling that run against the cloud tier run inside the perimeter. The network topology changes; the integration does not.

Updates, keys and capacity

Air gaps fail operationally, not architecturally. The hard parts are mundane: getting patches in, proving what is running, and keeping keys recoverable.

• Update intake: treat model and runtime updates as controlled software releases — provenance, scanning, staging, validation and rollback.
• Keys: store customer-managed keys in your KMS or on-prem HSM, and require two roles for destructive key operations.
• Capacity: size GPUs for peak concurrency plus node-failure headroom; an air-gapped estate cannot burst into the cloud.
• Recovery: back up model artefacts, configuration and keys, and test restore on a schedule.

Audit and governance inside the perimeter

Regulators will ask who ran what, on which model, with what result and who approved the exception. Keep identity, key lifecycle and inference metadata flowing to internal SIEM so those questions resolve from evidence rather than memory. Model registries should record version, provenance and validation results for every artefact in production.

Start with one bounded workload — document search over an internal corpus is a common first step — then extend to workflows with more sensitive data. Measure answer quality, review volume and override rates before widening scope, and keep a named owner for every deployed model.

Honest comparison

Requirement	Plugsky air-gapped	Cloud-only AI platform	Building in-house
Network isolation	No internet path required	Public endpoints by design	You design the perimeter
API continuity	Same OpenAI-compatible endpoint	Vendor API	You define the interface
Key control	Customer KMS or on-prem HSM	Limited options	Your HSM estate
Updates	Controlled intake on your cadence	Provider-scheduled	You manage everything
Audit	Events export to internal SIEM	Vendor dashboards	Your pipelines
Effort	Deployment plus GPU capacity	Low	Highest — models, serving, tooling

Frequently asked questions

Do we really need an air gap?

Only if regulation, classification or network architecture demands no egress at all. Many banking controls are satisfied by a VPC or on-prem deployment with controlled egress at lower operational cost.

Which models run offline?

Open-weight models suitable for offline operation, including Llama, Qwen, Mistral and Nemotron families. Check the model catalogue and docs for the current supported list.

How do updates work?

Through your controlled intake process: verified media, scanning, staging, validation and a rollback plan. You choose the cadence rather than receiving provider-scheduled changes.

Does application code change?

No. The internal API is OpenAI-compatible, so SDKs, prompts and tooling carry over. Only the base URL and model mapping differ.

What about key management?

Customer-managed keys live in your KMS or on-prem HSM. Revoking a key stops access to protected data, which makes it a real control — and a real responsibility.

How much GPU capacity do we need?

Size for peak concurrency plus node-failure headroom; air-gapped estates cannot burst. Use the deployment estimator to model options before committing hardware.

How do we start?

Run a bounded workload such as internal document search, with a named owner, an evaluation set and audit events flowing to your SIEM. Enterprise engagements cover deployment review and support terms.