If by “air-gapped GPUs” you mean GPU infrastructure that can be deployed with no external network dependency, there are several credible routes. The GPU itself isn't usually the “air-gapped” component—the server, management plane, networking, storage, software supply chain, and operational model are what make the deployment air-gapped.
Vendors to look at
NVIDIA — NVIDIA has an explicit AI Factory for Government reference architecture. Its RTX PRO Server platform is designed for inference-heavy workloads and specifically documents air-gapped/classified-enclave deployments. NVIDIA's reference architectures are delivered through partners including Dell, HPE, Lenovo, Cisco and Supermicro.
Hewlett Packard Enterprise (HPE) — probably the most directly packaged option. HPE Private Cloud AI is available in an air-gapped configuration, with NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, and HPE explicitly markets sovereign/air-gapped AI factories for governments.
Dell Technologies — Dell's AI Factory with NVIDIA provides sovereign/on-prem infrastructure, and Dell Federal has NVIDIA reference architectures using RTX PRO 6000 Blackwell GPUs.
Supermicro — a major NVIDIA platform partner and useful if you want to build a more customized air-gapped stack rather than buying a turnkey private-cloud appliance. NVIDIA lists Supermicro among the system partners for its government reference architecture.
Lenovo — likewise participates in NVIDIA's government reference-architecture ecosystem and can provide GPU servers for an isolated deployment.
Google — if you want more than bare GPU infrastructure, Google Distributed Cloud air-gapped is specifically designed for completely disconnected public-sector environments and can run AI workloads/models on-premises.
For a sovereign government cloud
I'd separate the procurement into three architectures:
Architecture
Typical fit
RTX PRO 6000 + 2–4 GPU nodes
Classified/air-gapped inference, edge sites, departmental AI
HGX B200/B300
Centralized training/fine-tuning and large-model inference
Turnkey private AI cloud
Government wants a supported cloud-like platform rather than assembling Kubernetes/GPU/storage/networking
NVIDIA's government architecture explicitly covers both RTX PRO 6000 for distributed/inference-oriented air-gapped environments and HGX B200/B300 for centralized large-scale workloads.
If I were designing the RFP, I would not specify “air-gapped GPU” alone. I'd specify requirements for zero external connectivity, disconnected management/control plane, offline software/container updates, secure boot/firmware provenance, GPU virtualization/isolation, logging, key management, supply-chain controls, and accreditation boundary. HPE, for example, explicitly offers air-gapped management across its private-cloud stack, while NVIDIA has government-ready components for its GPU software stack.
If you tell me the country, approximate GPU count (e.g. 64/256/1,000+), and whether this is classified/defense or merely sovereign/non-internet, I can narrow this to the vendors and specific GPU/server configurations that are actually realistic for the procurement.
When building a fully air-gapped, sovereign AI cloud for a government or defense client , "air-gapping" is a function of physical infrastructure architecture and deployment , not a special hardware SKU from chip manufacturers.
NVIDIA, AMD, and Intel do not sell a physically distinct "air-gapped GPU." Instead, standard enterprise/data-center GPUs (like NVIDIA H100/H200/B200, AMD Instinct, or Intel Gaudi) are deployed inside a customer-owned or sovereign-operated classified data center where network cables are physically severed from the public internet, and local management planes replace remote cloud telemetry.
To deliver an air-gapped sovereign AI cloud, you typically source the hardware via Original Equipment Manufacturers (OEMs) or integrate it with turnkey sovereign stack providers who specialize in disconnected operations.
1. The Hardware Providers (The Silicon)
You procure the raw compute units through enterprise server vendors certified to build high-density AI nodes:
NVIDIA (via OEMs like Dell, HPE, Supermicro, Cisco, Lenovo): Dominates government AI infrastructure. Through programs like the NVIDIA AI Factory for Government , they provide reference architectures that integrate with local orchestration layers (like NVIDIA Run:ai and NIM) designed to run completely disconnected.
AMD (via OEMs and systems integrators): The AMD Instinct accelerator ecosystem is increasingly utilized as a hardware-agnostic alternative for sovereign data centers looking to avoid single-vendor lock-in.
Building the orchestration, multi-tenancy, and security tooling for an air-gapped GPU cloud from scratch takes 12–24 months. Most government sovereign clouds rely on pre-packaged, disconnected software/hardware platforms:
Google Distributed Cloud (GDC) air-gapped: Specifically engineered for disconnected, sovereign operations meeting strict NIST/FedRAMP High or defense-level security controls. It packages local AI/database services (including Gemini capabilities) that operate with zero external telemetry or cloud callbacks.
Microsoft Azure Local / Disconnected Operations: Allows Azure-consistent services, AI models, and container orchestration to run on local hardware completely disconnected from the public Azure backbone.
Mirantis (with NVIDIA): Specializes in full-stack lifecycle management and Kubernetes orchestration explicitly certified for classified, air-gapped federal and defense environments.
To help narrow down the best path, could you share:
What security classification or compliance framework (e.g., FedRAMP High, Top Secret, local national standard) does the cloud need to meet?
Are you procuring on-premises bare-metal servers to manage entirely yourself, or looking for a pre-packaged sovereign cloud software stack to run on top?
Yes. For a government sovereign-AI cloud, there are several credible procurement paths. One important distinction: the GPU itself isn't “air-gapped”; you want a GPU server/platform whose hardware, firmware, management plane, drivers, orchestration, licensing, and support model can operate without external connectivity.
Strongest vendors to shortlist
Vendor
GPU / platform
Air-gapped / sovereign evidence
My take
NVIDIA
Blackwell B200/B300, RTX PRO 6000 Blackwell, Hopper H200
NVIDIA's Government AI Factory reference architecture explicitly identifies RTX PRO Servers for air-gapped operations and classified enclaves.
Best GPU ecosystem for a large government AI cloud
Hewlett Packard Enterprise (HPE)
HPE Private Cloud AI, ProLiant, NVIDIA GPUs
HPE explicitly supports air-gapped Sovereign AI Factory deployments; its Private Cloud AI is available in an air-gapped configuration and supports RTX PRO 6000 Blackwell and Hopper GPUs.
Probably the strongest turnkey sovereign-cloud option
Supermicro
NVIDIA HGX B200/B300, H200, RTX PRO, etc.
Supermicro has a dedicated federal portfolio for classified/unclassified agencies, U.S. manufacturing, transparent supply-chain attestation, and complete on-prem AI infrastructure.
Excellent for custom/high-density GPU clusters
Cisco
Cisco AI PODs + NVIDIA/AMD accelerators
Cisco explicitly offers AI PODs for government/defense with air-gapped deployment and no external connectivity, plus a validated path to turn them into an internal GPU cloud.
Very interesting if you want GPU-as-a-Service architecture
Dell Technologies
PowerEdge GPU servers + NVIDIA accelerators
Dell is one of the server manufacturers participating in NVIDIA's Government AI Factory reference design.
Strong conventional government procurement option
Lenovo
ThinkSystem GPU servers + NVIDIA
Lenovo is part of NVIDIA's Government AI Factory ecosystem and is supporting NVIDIA's security architecture across on-prem and air-gapped environments.
Worth RFQ'ing, especially for standard rack-server deployments
AMD
Instinct MI300X/MI325X/MI350-series
AMD's ROCm stack explicitly supports secure air-gapped environments, including offline installation; AMD also cites Instinct deployments in air-gapped environments.
Best alternative to NVIDIA if avoiding CUDA/vendor concentration
Foxconn
NVIDIA HGX/MGX/large-scale AI systems
Foxconn is an NVIDIA infrastructure partner building rack-scale AI factories and sovereign-AI infrastructure.
More relevant as an ODM/integration route than normal government OEM
What I'd put on your RFQ
For a new sovereign AI cloud, I'd investigate these four configurations first:
1. HPE + NVIDIA — turnkey sovereign cloud
HPE is unusually explicit about the combination you're asking for: sovereign infrastructure, multi-tenancy, GPU allocation, compliance, and air-gapped operation. Its current Sovereign AI Factory supports on-premises, jurisdictionally controlled and air-gapped deployments.
2. Supermicro + NVIDIA HGX B200/B300 — maximum compute density
This is attractive if your team wants to own more of the architecture rather than buying a fully packaged private cloud. Supermicro is specifically targeting classified and federal environments and offers NVIDIA GPU systems with U.S.-based manufacturing and supply-chain attestation.
3. Cisco AI POD + NVIDIA + GPU orchestration — sovereign GPU cloud
This is particularly interesting given your wording “sovereign AI cloud.” Cisco and Rafay have a validated architecture specifically for turning on-prem GPU infrastructure into a self-service, multi-tenant GPU cloud, including air-gapped deployment, RBAC, audit logging and quota/chargeback capabilities.
4. NVIDIA RTX PRO Server / B200/B300 — build your own
NVIDIA's Government AI Factory reference design is probably the best starting point for defining the technical requirements. It distinguishes smaller/distributed RTX PRO servers—which NVIDIA specifically positions for air-gapped/classified environments—from HGX B200/B300 systems for centralized training and large inference pools.
A potentially important alternative: AMD
I'd absolutely include AMD in the competitive bid rather than making the project NVIDIA-only.
AMD's current ROCm stack is explicitly designed to operate in secure and air-gapped Kubernetes environments, and ROCm provides an offline/self-contained installation mechanism.
The trade-off is ecosystem maturity: NVIDIA's CUDA/NVIDIA AI Enterprise/NIM ecosystem remains considerably broader, which can matter a lot for government workloads and third-party software.
One thing I'd specify in the procurement requirements
Don't merely write “air-gapped GPU server.” Require the vendor to demonstrate:
No Internet dependency for boot, provisioning, GPU management or normal operation
Offline installation/update mechanism for GPU drivers and AI software
No mandatory cloud-based license validation
Local management plane and telemetry
Ability to disable vendor remote access
Firmware/BMC update process suitable for disconnected networks
SBOM and signed firmware/software artifacts
Supply-chain provenance and country-of-origin requirements
Secure boot / measured boot / TPM
GPU partitioning/MIG or equivalent multi-tenancy
Kubernetes/OpenShift support in disconnected mode
Offline container registry mirroring
Local model/container/package repositories
Support and break/fix procedures that don't require network access
Evidence that the entire stack, not merely the GPU, has been validated for the agency's security boundary
That last point is critical. For example, NVIDIA provides government-ready components for its GPU Operator, but explicitly notes that not every component is currently government-ready.
My initial shortlist would therefore be: HPE, Supermicro, Cisco, Dell, Lenovo, NVIDIA direct, and AMD/Instinct. For a serious government procurement, I'd solicit both a turnkey HPE/Cisco proposal and a more open Supermicro/NVIDIA or Supermicro/AMD proposal so you can compare lifecycle sovereignty and vendor lock-in rather than just GPU $/TFLOP.
If you tell me (a) U.S. or another government, (b) target GPU count, and (c) training vs. inference vs. mixed, I can turn this into a vendor-by-vendor RFQ shortlist with current GPU/server SKUs, estimated cluster architecture, supply-chain/security certifications, and procurement contacts.
Intel (Gaudi accelerators / Xeon-native stacks): Frequently used for secure, cost-effective, or CPU/accelerator hybrid inference workloads in regulated public-sector setups.
Nokia & Specialized Integrators: Offer custom carrier-grade or government-grade private cloud orchestration platforms designed to manage petabyte-scale GPU fabrics without reliance on hyperscaler connectivity.
