RESEARCHED PROVIDER · PRIMARY SOURCES

Super Protocol

Super Swarm deployment software for local model inference on confidential CPU/GPU infrastructure, with distinct trusted and untrusted modes.

INPUT / OUTPUT PRIVACY

Stage 1

Partial proof

ASSESSED DEPLOYMENT

Trusted-mode GPU model deployments

RESEARCH REVIEW

2026-10-09 · primary-source review

Provider website

Verification labels describe capabilities established by the cited sources. This review did not perform live cryptographic attestation, reproduce production builds or complete an independent security audit.

Who actually runs the models?

Customers create GPU clusters and deploy vLLM with model weights through Super Swarm. The guide supplies actual MedGemma and Apertus deployment scripts and model-serving endpoints. Inclusion covers inference hosted on this infrastructure/software, not unrelated upstream model APIs.

Deployment and TEE coverage

Covers appropriately configured trusted Super Swarm deployments of local model engines. Untrusted development mode, arbitrary applications and ordinary public HTTPS clients without evidence verification are not covered by the technical proof. Ingress, model configuration and content retention still require application-specific review.

Trusted mode admits confidential TDX/SEV-SNP VMs against registered measurements and rejects GPUs in debug mode. Untrusted mode omits these checks. GPU architecture, firmware, confidential-compute mode and inter-GPU links must match a supported configuration; a confidential CPU alone is insufficient.

At least one content layer has independently checkable technical evidence, including partial proof within a layer. The full privacy boundary is incomplete. The stage tooltip distinguishes partial findings from fully met requirements for Stage 2.

STEPS FOR STAGE 2

0 of 3 requirements for Stage 2 met. Missing proof: Request path, Inference execution, Logs & storage.

  1. Request pathNot established

    Private inputs and outputs across the request path, with E2EE or an equivalent attested channel that protects against intermediaries and binds keys to the accepted workload.

    Documented only. Example uses ordinary HTTPS

  2. Inference executionPartial

    Protected inference with a verifiable workload identity and protection covering the CPU, GPU and every place content is processed.

    Partial technical proof. Root and workload evidence

  3. Logs & storageNot established

    No content logging, persistent storage, human review or training. Any temporary content-derived cache must have verifiable isolation and enforced removal.

    Unknown. No fleet-wide deletion proof

Identity and operational metadata are assessed separately below. They do not set the content stage; prompt or response content in telemetry remains part of the content-retention assessment.

Request path

Documented only

Ingress requires verification. Swarm's certificate design binds VM public keys into hardware reports and exposes the network mode in the public root certificate. External verifiers must anchor trust in that root. The model-deployment example uses ordinary HTTPS and an API key; it does not establish an automatically verifying end-user channel through the published ingress to the selected workload.

Inference execution

Partial technical proof

VM proof; GPU gap. The documented public root/VM certificates carry hardware reports and key bindings, and signed deployment evidence lets external verifiers check part of the VM/workload boundary. Trusted mode requires registered CPU measurements and rejects GPU debug mode. Hardware guidance additionally distinguishes PCIe protection and inter-GPU topology. The debug-mode check alone does not establish full independent GPU firmware/link attestation or end-user validation of every serving artifact. Bootstrap-root trust and shared CA keys remain explicit dependencies.

Logs & storage

Unknown

Application-dependent. The published vLLM examples let customers choose model-serving arguments and include a multimodal processor cache. Encrypted SwarmDB cluster state and downloaded model weights are not prompt-deletion controls. The reviewed sources do not establish content-log suppression, tenant cache isolation/removal, backups or training exclusion for arbitrary workloads. These properties must be reviewed and bound to the selected deployment.

Identity & metadata

Unknown

Deployment-dependent. The deployment guide uses dashboard authentication, domains, published ingress endpoints and API keys. The reviewed technical sources do not establish end-user IP, account, billing or connection-metadata collection and deletion across model deployments.

Remaining trust assumptions

  • Intel/AMD confidential VMs and the selected NVIDIA GPU mode/topology.
  • The bootstrap root certificate and registered trusted measurements; the bootstrap node does not verify itself against another node.
  • Swarm members share root/subroot private keys in the documented design, expanding the CA trust boundary.
  • Customer model-serving code, ingress configuration, update policy and content-retention controls.

Limits of this assessment

  • Trusted and untrusted modes have materially different protection; only trusted mode is assessed.
  • Public deployment examples do not themselves provide an attestation-verifying inference client.
  • The inspected sources establish verification capabilities, not a verified live deployment or a fleet-wide no-logging guarantee.

Primary-source record

3 sources · reviewed 2026-10-09

Back to providers

Sources reviewed 2026-10-09