Attachment 1_ERS_Enterprise GPU Servers for AI (17JUL26).pdf

Enterprise GPU Servers for AI

Equipment Requirements Specification (ERS)

17 JULY 2026

1. SCOPE

This Equipment Requirements Specification (ERS) establishes the minimum technical, functional, and performance requirements for commercially available artificial intelligence (AI) computing infrastructure, associated software licenses, and manufacturer warranty support to support enterprise AI, machine learning (ML), and high-performance computing (HPC) workloads.

The Government intends to acquire commercially available AI computing products capable of supporting large language model (LLM) training, inference, advanced analytics, computer vision, and other AI-enabled workloads within secure Government computing environments.

The Government is interested in commercially available products that meet or exceed the requirements identified herein.

The Government anticipates acquiring equipment in varying quantities and configurations through individual delivery orders issued under a resulting contract vehicle. The requirements identified herein represent minimum technical requirements and do not constitute a commitment to purchase any specific quantity or configuration.

2. GENERAL REQUIREMENTS

2.1 Commercial Products

All products furnished under this ERS shall be commercially available products customarily offered to the general public.

2.2 Secure Environment Operations

All products shall be capable of operating within secure, air-gapped environments without requiring Internet connectivity for normal operations.

2.3 Data Center Compatibility

Unless otherwise specified, all equipment shall be compatible with standard commercial data center environments.

Category I - AI Compute Servers shall:

• Support installation within Government-provided standard 19-inch racks;

• Operate utilizing Government-provided 110/220V power infrastructure; and

• Be suitable for deployment within secure Government facilities.

Category II Rack-Scale AI Systems and Category III AI Pod Architectures may include vendor-integrated racks, cabinets, cooling infrastructure, and associated support equipment customarily provided as part of the commercial solution.

Vendors shall identify all facility, power, cooling, floor loading, footprint, environmental, and space requirements associated with proposed rack-scale systems and AI pod architectures.

2.4 Warranty

Unless otherwise specified in an individual delivery order, all hardware shall include a minimum three-year OEM warranty and technical support commencing upon Government acceptance. The warranty shall be provided or fully backed by the OEM and shall include repair or replacement of defective hardware, access to OEM technical support, firmware updates, and customer retention of failed data-storage devices when commercially available.

If any component is subject to an OEM warranty period shorter than three years, the Contractor shall identify the component and provide separately priced coverage sufficient to ensure that the complete system is supported for at least three years. Warranty obligations shall survive expiration of the contract ordering period and the applicable delivery order and shall not require the contract or order period of performance to remain open for the duration of the warranty.

Warranty support shall include:

• Hardware replacement support;

• Firmware updates;

• Technical support access; and

• Customer-retained failed hardware components (e.g., keep-your-drive or equivalent), when commercially available.

2.5 Software Licensing

Software licenses required for operation of proposed hardware shall be separately identifiable and separately priced.

3. PRODUCT CATEGORIES

3.1 Category I – B300 GPU Servers (Estimated Quantity: 4)

The Contractor shall provide commercially available B300 GPU servers optimized for AI model training and inference workloads.

Each server shall meet or exceed the following minimum requirements:

Processing

• Intel Xeon 6767P processors or equivalent/better.

Memory

Minimum:

• 2 TB RAM, with available expansion to 4 TB.

• Vendors shall separately identify and price any configuration required to expand memory capacity to 4 TB.

Storage

Minimum:

• Two (2) 500 GB or larger SSD boot drives; and

• Two (2) 7.68 TB or larger SSD data drives.

GPU Architecture

• One (1) HGX B300 eight-GPU air-cooled assembly or equivalent.

Networking

Minimum networking capabilities shall include:

• Two (2) 25 Gb Ethernet interfaces; and

• Two (2) NVIDIA BlueField-3 single-port 400 GbE DPUs or equivalent.

Confidential Computing

Each proposed system shall support confidential computing capability.

The Contractor shall identify the hardware, firmware, and software features included to enable confidential computing.

3.2 Category II – GB300 NVL72 Rack-Scale Systems (Estimated Quantity: 3) The Contractor shall provide commercially available integrated rack-scale AI systems designed to support large-scale AI training and inference workloads.

The proposed solution shall meet or exceed the following minimum requirements:

• One integrated GB300 NVL72 rack-scale system or equivalent;

• Integrated high-speed accelerator interconnect fabric;

• Integrated Ethernet networking equivalent to Spectrum-X architecture;

• Integrated cooling solution utilizing an in-rack or sidecar coolant distribution unit (CDU); and

• Compatibility with Government-provided facility water cooling infrastructure.

Minimum storage per compute tray shall include:

Two (2) 500 GB or larger SSD boot drives; and

Two (2) 3.84 TB or larger SSD data drives.

3.3 Category III – AI Pod Architecture (Estimated Quantity: 1) The Contractor shall provide one commercially available, integrated, on-premises AI pod architecture optimized for enterprise artificial intelligence workloads.

The proposed solution shall:

• Consist of one (1) TPU v8 pod, if commercially available; otherwise, one (1) TPU v7 pod or equivalent;

• Support enterprise AI workloads, large language model training, and inference operations; and

• Include all associated software, models, licenses, and supporting components necessary for operation.

Vendors shall provide detailed technical specifications for the proposed AI pod architecture, including, at a minimum:

• Accelerator type, quantity, and memory capacity;

• Aggregate accelerator memory available to workloads;

• CPU architecture, quantity, and memory, if applicable;

• Networking architecture, topology, and bandwidth;

• Interconnect architecture and bandwidth;

• Storage architecture and capacity;

• Scalability characteristics;

• Physical footprint and rack configuration;

• Power requirements;

• Cooling requirements; and

• Any additional architectural features necessary to fully describe the proposed solution.

Offerors shall provide a comprehensive gap analysis identifying any areas where the proposed pod may fall short in performance when compared to the GB300 NVL72 baseline, including an explanation of the specific limitations, how those limitations manifest, and the potential operational impact to training, inference, scalability, or other mission workloads.

Vendors shall document all software, models, licenses, and supporting components included with the proposed solution, including, at a minimum:

• Operating systems;

• Drivers;

• AI frameworks;

• Orchestration platforms;

• Models and model libraries;

• Supporting middleware and utilities; and

• Any software licenses included as part of the proposed solution.

Vendors shall clearly identify whether software components are perpetual, subscription-based, consumption-based, or otherwise licensed.

Vendors shall provide detailed model performance metrics, including:

• Throughput for each supported modality (e.g., text, image, audio, video, embeddings, and multimodal workloads);

• Latency for each supported modality;

• Scaling characteristics;

• Benchmark results, where available; and

• Any known performance constraints or limitations.

Vendors shall perform a comparative gap analysis against a commercially available Google Vertex AI environment, identifying differences in:

• Compute capability;

• Throughput;

• Scalability;

• Software stack;

• Model availability and support; and

• Any limitations that could affect model training or inference performance.

Vendors shall specifically identify:

• Any limitations associated with tool calling;

• Any limitations associated with structured output;

• Any limitations associated with agentic workflows;

• Any limitations associated with external integrations or APIs;

• Any other model capability limitations when compared to commercially available Vertex AI services.

Vendors shall clearly identify all software, models, AI services, and capabilities included as part of the proposed solution and describe how those capabilities compare to commercially available Vertex AI offerings, including any bundled or integrated model capabilities.

Vendors shall clearly identify all assumptions, benchmarks, methodologies, and test conditions used in developing the comparative analyses.

3.4 Category IV – AI Inference Accelerators

The Contractor shall provide commercially available NVIDIA L40S or equivalent inference accelerators optimized for production AI inference workloads.

Each accelerator shall:

• Provide a minimum of 48 GB GPU memory;

• Support mixed precision inference including FP32, FP16, BF16, and INT8;

• Be compatible with industry-standard AI frameworks including CUDA, TensorRT, PyTorch, and TensorFlow;

• Support high-throughput inference operations;

• Support low-latency inference;

• Support concurrent model execution; and

• Be suitable for production AI workloads.

4. TECHNICAL EQUIVALENCY

Products identified by manufacturer, model number, architecture, or trade name are provided solely to establish the minimum required salient characteristics.

Vendors may propose equivalent products provided the proposed products meet or exceed all identified salient characteristics and performance requirements.

The Government reserves the right to evaluate equivalency based on technical capability, compatibility, interoperability, performance, scalability, and overall suitability for the intended application.

5. DELIVERABLES

The Contractor shall provide:

• Hardware;

• Software licenses;

• OEM warranty documentation;

• Technical documentation;

• Shipping documentation;

• User documentation; and

• Any other documentation customarily provided with commercial products.

6. DELIVERY

The anticipated delivery location for most orders is expected to be the FBI Criminal Justice Information Services (CJIS) Division facility located in Clarksburg, West Virginia.

The Government reserves the right to designate alternate delivery locations within the continental United States (CONUS) at the individual order level.

Delivery locations, quantities, and specific delivery requirements shall be identified in each order.

FOB Destination shall apply unless otherwise specified.