The file's text, extracted by GovTribe without its formatting.
Appendix B — Technical Requirements Empire AI RFP: Secure HPC Enclave for Regulated Research Data · Issued: 09/10/2026 This appendix forms part of the above RFP and is incorporated into it by reference. Requirements are stated as functional outcomes. Bidders may satisfy them by any means, including reallocation of existing assets where the design assessments in the main RFP support it. Where this appendix conflicts with the main RFP or with Appendix A, Appendix A (CR-2) prevails, followed by the main RFP. Bidders must address each requirement in the compliance matrix required under Proposal contents in the main RFP and must cite requirement identifiers (TR-1 through TR-7) in their response.
TR-1 Compute fabric
| Requirement |
| Specification |
| GPU backend interconnect |
| Non-blocking 400G-class fabric sized for 64 GPUs across 8 nodes (64 host-side ports), supporting GPUDirect RDMA and collective communication at full bisection. Physically separate from existing production fabrics. |
| Frontend Ethernet |
| Redundant high-speed switch pair carrying platform, tenant ingress, and inference traffic. |
| Storage fabric |
| Redundant switch pair sized to the storage platform selected under Assessment A. State whether a dedicated storage fabric is required or whether convergence with the frontend is appropriate at this scale. |
| Out-of-band management |
| Dedicated redundant management network serving BMC/IPMI only, not bridged to or reachable from any existing management network. |
| Cabling |
| All host, fabric, and inter-rack cabling included and specified. All nodes dual-homed to redundant switch pairs. |
TR-2 Perimeter and security services
| Requirement |
| Specification |
| Boundary enforcement |
| Stateful and application-layer policy at the enclave boundary, deny-by-default with an enumerated allow-list, per the option recommended under Assessment B. Enforcement must be validated by test at acceptance per Appendix A (CR-2). |
| Shared services allow-list |
| Dependencies on existing infrastructure (DNS, NTP, licensing, identity) must be enumerated individually and permitted explicitly. General reachability to the existing core is not acceptable. |
| Tenant API access |
| Mutual TLS with per-tenant certificates, rate limiting, and per-tenant request audit logging for any externally exposed API. |
| Key management |
| FIPS Validated equivalent key management supporting per-tenant key separation and automated rotation. State whether a dedicated appliance is required or whether an equivalent service satisfies the requirement. |
| Secrets management |
| No plaintext credentials anywhere in the enclave, including BMC credentials, TLS material, and service accounts. |
| Administrative access |
| A single dedicated administrative entry point, not shared with or reachable from existing jump hosts. |
TR-3 Storage
| Requirement |
| Specification |
| Sourcing |
| Per the recommendation under Assessment A. Price each viable option. |
| Multi-tenancy |
| Per-tenant logical namespace with independent key separation. Cross-tenant access must be structurally impossible, not policy-restricted. |
| Performance |
| State sustained aggregate read and write throughput and metadata operation rates, with benchmark methodology, at each priced capacity tier — measured with encryption and tenant isolation enabled. |
| Client access |
| Protocol and driver support appropriate to the orchestration model recommended under Assessment C, with mount scope enforced per tenant. |
| Data protection |
| State tolerable simultaneous device failures, rebuild time, and degraded-mode performance. |
| Media disposal |
| Sanitization or destruction of media that has held regulated data, to NIST SP 800-88 standards, with certificates of destruction. State what is included in the base proposal and what is chargeable. Supports 45 CFR §164.310(d)(2)(i), a required specification. |
| Warranty return and media re-use |
| State the end-to-end process for any drive or media leaving the enclave for warranty return, replacement, or re-use: whether the platform supports cryptographic or instant secure erase, and whether the vendor’s RMA terms permit drive retention or destruction in place. An RMA that returns unsanitized media to the manufacturer is not acceptable. Supports §164.310(d)(2)(ii), a required specification. |
| Asset movement record |
| A record of storage hardware and media movements into, within, and out of the enclave, identifying the person responsible, integrated with the asset inventory under TR-7 rather than maintained as a parallel manual log. |
TR-4 Orchestration and platform
| Requirement |
| Specification |
| Orchestration |
| Per the recommendation under Assessment C, with HA control plane distributed across separate racks and failure domains. |
| Tenant scheduling |
| Per-tenant quota, priority, and fair-share allocation with no cross-tenant queue or resource visibility. |
| GPU allocation |
| Support for the allocation tiers recommended under Assessment D, with the default for regulated workloads stated. |
| Driver and fabric lifecycle |
| Automated GPU driver, telemetry, and fabric device management, with a stated update and validation process that does not require ad hoc manual intervention. |
| Cluster management |
| Must not join the existing shared cluster-management domain. |
TR-5 Identity, access, and audit
| Requirement |
| Specification |
| External federation |
| SAML 2.0 / OIDC federation with tenant institution identity providers. Tenants must not be required to enroll in Empire AI’s existing provisioning pipeline or maintain secondary credentials. |
| Multi-factor authentication |
| Enforced for all administrative and tenant access, satisfied by the federated provider where the institution supports it. |
| Authorization scoping |
| Access rights derived from the tenant’s approved protocol scope and enforced technically at the platform and storage layers. |
| Tenant onboarding |
| A single automated workflow provisioning federation, workload scope, resource quota, storage namespace, and key separation. |
| Deprovisioning |
| Automated revocation on workforce or project termination, with a stated maximum time to effect and an auditable record. |
| Access review |
| Support for periodic recertification of tenant and administrative access, with campaign records retained as evidence. |
| Privileged session audit logging |
| Audit logging of privileged administrative access to enclave hosts and platform, subject to the data-path constraints in Appendix A (CR-3). |
| Audit pipeline |
| Platform, storage, gateway, and privileged session records forwarded to centralized, tamper-evident storage with configurable retention. |
| Tenant-accessible logs |
| Each tenant able to retrieve a complete access log for its own datasets without mediation by Empire AI operations staff. |
TR-6 Physical and facilities
| Requirement |
| Specification |
| Footprint |
| Provide a rack elevation for the complete build, including the GPU nodes already on order. |
| Power |
| Typical and maximum draw per rack and for the cluster; dual-feed PDU per rack; circuit type and amperage specified. Final design must adhere to ~100kW max power. |
| Cooling |
| Maximum heat output per rack and cluster, and the cooling method required for the GPU racks. |
| Floor loading |
| Fully populated weight per rack. |
| Physical separation |
| The enclave must not share a cooling loop, PDU chain, or physical enclosure with existing production compute. Identify any proposal element that would violate this. |
TR-7 Infrastructure as code The delivery pipeline is a required deliverable. The enclave must be provisioned from code at first power-on and must remain so.
| Requirement |
| Specification |
| Source of truth |
| Version-controlled declarative configuration covering network and perimeter configuration, node provisioning, platform deployment, and tenant definitions. |
| Drift detection |
| Automated detection and reporting of divergence between declared and actual state, with a defined remediation workflow. |
| Evidence generation |
| The pipeline should produce compliance evidence as a by-product — configuration state, change history, and approval records available on demand rather than assembled for an audit. |
| Handover |
| Repository, pipeline, and documentation transfer to Empire AI ownership at acceptance. An Empire AI operator must be able to rebuild any component from code without vendor assistance. |