GPU Compute Servies for Network of Optimal Dynamic Energy Signatures

The Defense Advanced Research Projects Agency seeks a dedicated, managed high-performance computing solution providing 50 NVIDIA HGX H100 80GB GPUs (or equivalent/better, including NVIDIA Blackwell B200/GB200) with network-attached storage, high-speed networking, and comprehensive user support for the Network of Optimal Dynamic Energy Signatures (NODES) program, with proposals due May 8, 2026. The procurement is set aside for Women-Owned Small Businesses under NAICS 518210, with a size standard of $40 million.

Place of performance is Arlington, Virginia. The solution must be fully provisioned and available by May 15, 2026, with 100 additional GPU-baseboards held in reserve. Technical specifications include 550TB of NFS storage, one x86 compute virtual machine, dedicated 100G Internet access, isolated Virtual Private Cloud deployment, 24/7/365 Tier 1 customer support and Tier 3 scientific software support (encompassing CUDA, NCCL, drivers, clusterware, PyTorch, TensorFlow, and JAX) for up to 100 users, cluster testing with NCCL protocols, and monthly usage and performance reporting. Security requirements mandate SOC 2 Type II certification and CMMC Level 1 (Self) or higher; ISO 27001 and FedRAMP High are not acceptable substitutes. The current requirement does not involve CUI, ITAR, or EAR data, though future NODES continuations may include such requirements.

The contract is Firm Fixed Price under FAR Part 12 for a five-month base period (May 15–October 14, 2026) and five-month option period (October 15, 2026–March 14, 2027), with pricing structured as a single all-inclusive per-baseboard-hour rate based on metered consumption: 183,600 estimated hours (base) and 181,200 estimated hours (option). Billing reflects actual consumption only, subject to a Not-to-Exceed ceiling. Line Item 0003 accommodates ancillary services not included in the hourly rate.

Quotations are three volumes: Volume I (Technical, 10 pages), Volume II (Price, no limit), and Volume III (Administrative, no limit). Evaluation follows a two-step best value process with Pass/Fail Go/No-Go technical screening, followed by comparative assessment weighing Technical Approach and Past Performance (combined, significantly more important) against Price. Past performance assessment accepts up to three relevant contracts from the prior three years managing at least 25 high-end GPUs in high-availability HPC or GPU-as-a-Service environments. The Government prefers US-hosted solutions and may award without discussions. Amendment 0002 (May 5, 2026) changed billing from per-GPU-hour to per-baseboard-hour methodology and incorporated FAR Clause 52.222-90 (Addressing DEI Discrimination by Federal Contractors, Deviation 2026-O0038).