Perplexity and Valarian Bring Computer to Private and Air-Gapped Infrastructure

Perplexity announced a partnership with Valarian on October 20, 2026 to run Computer inside institutions' own infrastructure, including air-gapped networks. Valarian's ACRA platform supplies access and network controls around the agent. The offer targets organizations that need AI workflows without sending sensitive work through a public cloud service.

For enterprise teams, hosting a model is only the starting point. An agent also needs documents, tools, credentials and somewhere to execute code. Each connection adds a separate question about what that agent is allowed to do.

What Perplexity and Valarian are offering

According to Perplexity's announcement, deployments are available now through its enterprise contact process. They can use customer-controlled cloud environments or servers owned by the institution, with open-weight models running inside that environment.

Computer provides task planning and multi-step work. Valarian organizes workloads into controlled environments called enclaves, with rules for access to data, tools, users and credentials. Audit records expose access decisions and workload activity to security teams.

The announcement does not publish a standard price, a minimum hardware configuration or independent performance comparisons. This is a deployment offering that needs technical scoping, rather than a self-serve download with a universal specification.

How Sovereign Computer controls access

Valarian calls the product Sovereign Computer. Its description separates Perplexity's harness, orchestrator and planner from the infrastructure enforcing what those components may reach.

External communication is denied by default, with exceptions explicitly declared. Files, APIs, data stores and MCP servers can be connected within the boundary, and individual routes can be revoked. A model can be changed without rebuilding the surrounding control system.

That separation is useful because a model's decision and an organization's permission are different things. An agent may decide that an external service would help finish a task. The infrastructure still needs to refuse that request if the workload has no authorized route.

What ACRA adds underneath the agent

Valarian's ACRA architecture builds on Kubernetes and applies controls at workload level. Its published stack includes SPIFFE/SPIRE for workload identity, Cilium and Istio for networking, Kyverno for policy, Vault for secrets, and OpenTelemetry for observability.

An enclave combines identity, isolation, permissions and audit around a workload. Valarian also describes runtime containment and revocation, so a problematic workload can be isolated without destroying the surrounding system.

These are vendor descriptions, not an independent security assessment. Buyers should ask to see the actual policy decisions, audit records and failure behavior in their proposed deployment, rather than treating the component list as proof that every risk is covered.

What air-gapped operation gives up

An air-gapped deployment has no live connection to public websites or external model APIs. Computer must use the models, tools and information available inside that network. A current-events research request therefore needs an approved process for bringing fresh information into the environment.

Keeping execution inside a boundary also makes the quality of internal data more important. If an agent has access to an outdated policy document, local hosting does not tell it that a newer version exists elsewhere.

A useful pilot would test both successful work and refused actions: can it produce a source-backed internal report, and can it reliably stop when asked to reach data outside its assigned scope? Those outcomes are more informative than a broad claim of enterprise readiness.

The partnership makes the agent's deployment boundary a product decision. The next question is whether institutions can operate that boundary clearly enough that useful automation and accountable access grow together.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help enterprise AI teams explain complex technical products clearly for developers, buyers and decision-makers.

Book a 30-minute call.