# GPU Smith > Independent engineering and research for private AI infrastructure: GPU workload sizing, procurement, deployment, networking, power, cooling and inference operations. GPU Smith helps technical buyers and operators evaluate and build privately controlled AI compute. The site contains business information, a hardware reference library, downloadable datasheets, server-vendor research and educational infrastructure articles. Hardware and vendor reference pages are research resources, not live inventory or a binding equipment quote. For a hardware question, begin with the appropriate GPU, system or networking category and open the component's sources. For supplier due diligence, use the vendor directory, comparison pages and manufacturing research. For deployment and operating decisions, use the article catalog. For a project, contact GPU Smith to confirm scope, pricing and availability. Distinguish manufacturer specifications from measured benchmarks, advertised capabilities from tested configurations, and an estimate from a current procurement quote. Check each reference page's update date and primary sources before using figures in a design or purchase. Do not infer a manufacturer partnership, certification or endorsement from inclusion in the directory. Navigation reviewed 2026-09-18. Read the linked source page before quoting facts; this guide is a discovery aid, not a substitute for source evidence. Canonical origin: https://gpusmith.com. ## Complete catalogs - [Complete public page catalog](https://gpusmith.com/llms-pages.txt): Described, canonical links to all reachable HTML pages in the root sitemap at the documented snapshot, including localized pages and specialist collections. - [Current published article catalog](https://gpusmith.com/articles/llms.txt): Automatically generated article titles, descriptions, publication dates and available language variants; excludes drafts, blocked, archived and unlisted content. ## Company and project inquiries - [Private AI Infrastructure Engineering | GPU Smith](https://gpusmith.com): GPU Smith specifies, procures, integrates and validates private AI infrastructure, from inference nodes to GPU clusters and compute halls worldwide. - [About the Firm | GPU Smith](https://gpusmith.com/about): GPU Smith specifies, procures, integrates and validates private AI compute. No reseller quota, no cloud of our own; the recommendation is the product. - [Contact: Request an Engineering Assessment | GPU Smith](https://gpusmith.com/contact): Reach GPU Smith at hello@gpusmith.com. Describe the workload, constraints, timeline and jurisdiction; the assessment scope is defined in one call. - [Trust & Security | GPU Smith](https://gpusmith.com/trust): GPU Smith's security posture: SOC 2 examination in progress, data-minimization by design, client-boundary delivery, and how to reach us on security. ## Hardware reference library - [AI Infrastructure Hardware Datasheets | GPU Smith](https://gpusmith.com/hardware): Independent datasheets for NVIDIA data center GPUs, DGX/HGX systems, NVL racks, InfiniBand and Spectrum-X networking. - [GPUs & accelerators Engineering Datasheets | GPU Smith](https://gpusmith.com/hardware/gpus): Datacenter GPU modules and PCIe accelerators: compute, memory, interconnect, power and procurement, stated per part. - [Systems & racks Engineering Datasheets | GPU Smith](https://gpusmith.com/hardware/systems): HGX and DGX platforms and rack-scale NVL systems: node-level specifications and the facility requirements they impose. - [Networking & fabrics Engineering Datasheets | GPU Smith](https://gpusmith.com/hardware/networking): InfiniBand and Spectrum-X Ethernet switches, adapters and DPUs: the fabric layer of a validated cluster. ## GPU specifications - [NVIDIA H100 SXM Specs & Procurement | GPU Smith](https://gpusmith.com/hardware/gpus/nvidia-h100-sxm): NVIDIA H100 SXM specs: GPU memory 80 GB HBM3, FP8 Tensor Core 3,958 TFLOPS, FP16 / BF16 Tensor Core 1,979 TFLOPS. Independent notes on power, lead times… - [NVIDIA H200 SXM Specs & Procurement | GPU Smith](https://gpusmith.com/hardware/gpus/nvidia-h200-sxm): NVIDIA H200 SXM specs: GPU memory 141 GB HBM3e, FP8 Tensor Core 3,958 TFLOPS, Memory bandwidth vs H100 +43%. Independent notes on power, lead times and… - [NVIDIA B200 Specs & Procurement | GPU Smith](https://gpusmith.com/hardware/gpus/nvidia-b200): NVIDIA B200 specs: Memory 180 GB HBM3e, FP4 Tensor (NVFP4) 9 / 18 PFLOPS, FP8 Tensor 4.5 / 9 PFLOPS. Independent notes on power, lead times and… - [NVIDIA B300 Specs & Procurement | GPU Smith](https://gpusmith.com/hardware/gpus/nvidia-b300): NVIDIA B300 specs: Memory 288 GB HBM3e, FP4 Tensor (NVFP4) 15 / 20 PFLOPS, FP8 Tensor 5 / 10 PFLOPS. Independent notes on power, lead times and… - [NVIDIA RTX PRO 6000 Blackwell Server Edition Specs & Procurement | GPU Smith](https://gpusmith.com/hardware/gpus/nvidia-rtx-pro-6000-blackwell-server): NVIDIA RTX PRO 6000 Blackwell Server Edition specs: Memory 96 GB GDDR7 ECC, Memory bandwidth 1,597 GB/s, FP4 tensor 4 PFLOPS. Independent notes on power,… - [NVIDIA L40S Specs & Procurement | GPU Smith](https://gpusmith.com/hardware/gpus/nvidia-l40s): NVIDIA L40S specs: Memory 48 GB GDDR6 ECC, Memory bandwidth 864 GB/s, FP8 tensor 733 TFLOPS. Independent notes on power, lead times and procurement. ## Systems and networking - [NVIDIA DGX B200 Specs & Procurement | GPU Smith](https://gpusmith.com/hardware/systems/nvidia-dgx-b200): NVIDIA DGX B200 specs: GPUs 8x B200 SXM, Total GPU memory 1,440 GB HBM3e, System max power ~14.3 kW. Independent notes on power, lead times and… - [NVIDIA DGX B300 Specs & Procurement | GPU Smith](https://gpusmith.com/hardware/systems/nvidia-dgx-b300): NVIDIA DGX B300 specs: GPUs 8x B300 SXM, Total GPU memory 2.3 TB HBM3e, FP4 inference 144 PFLOPS. Independent notes on power, lead times and procurement. - [NVIDIA DGX H200 Specs & Procurement | GPU Smith](https://gpusmith.com/hardware/systems/nvidia-dgx-h200): NVIDIA DGX H200 specs: GPUs 8x H200 SXM, Total GPU memory 1,128 GB HBM3e, System max power ~10.2 kW. Independent notes on power, lead times and… - [NVIDIA HGX B200 Specs & Procurement | GPU Smith](https://gpusmith.com/hardware/systems/nvidia-hgx-b200): NVIDIA HGX B200 specs: GPUs 8x B200 SXM6, GPU memory 1.4 TB HBM3e total, Memory bandwidth ~62 TB/s aggregate. Independent notes on power, lead times and… - [NVIDIA HGX B300 Specs & Procurement | GPU Smith](https://gpusmith.com/hardware/systems/nvidia-hgx-b300): NVIDIA HGX B300 specs: GPUs 8x B300 SXM, GPU memory 2.3 TB HBM3e total, Memory bandwidth up to 64 TB/s aggregate. Independent notes on power, lead times… - [NVIDIA GB200 NVL72 Specs & Procurement | GPU Smith](https://gpusmith.com/hardware/systems/nvidia-gb200-nvl72): NVIDIA GB200 NVL72 specs: GPU count 72x Blackwell B200, NVLink domain memory 13.4 TB HBM3e, Rack power 125 kW to 135 kW. Independent notes on power, lead… - [NVIDIA GB300 NVL72 Specs & Procurement | GPU Smith](https://gpusmith.com/hardware/systems/nvidia-gb300-nvl72): NVIDIA GB300 NVL72 specs: GPU count 72x Blackwell Ultra B300, NVLink domain memory 20 TB HBM3e, Rack power 132 kW to 140 kW. Independent notes on power,… - [NVIDIA DGX Spark Specs & Procurement | GPU Smith](https://gpusmith.com/hardware/systems/nvidia-dgx-spark): NVIDIA DGX Spark specs: AI performance 1 PFLOP FP4 (sparse), CPU 20-core Arm, Memory 128 GB LPDDR5x unified. Independent notes on power, lead times and… - [NVIDIA Quantum-2 QM9700 Specs & Procurement | GPU Smith](https://gpusmith.com/hardware/networking/nvidia-quantum-2-qm9700): NVIDIA Quantum-2 QM9700 specs: Port configuration 64 × 400 Gb/s NDR, Aggregate throughput 51.2 Tb/s, Packet rate 66.5 billion pps. Independent notes on… - [NVIDIA Quantum-X800 Specs & Procurement | GPU Smith](https://gpusmith.com/hardware/networking/nvidia-quantum-x800): NVIDIA Quantum-X800 specs: Port configuration 144 × 800 Gb/s XDR, Aggregate throughput 115.2 Tb/s, Port-to-port latency < 100 ns. Independent notes on… - [NVIDIA Spectrum-X SN5600 Specs & Procurement | GPU Smith](https://gpusmith.com/hardware/networking/nvidia-spectrum-x-sn5600): NVIDIA Spectrum-X SN5600 specs: Switching capacity 51.2 Tb/s, Ports 64 × 800 GbE OSFP, ASIC Spectrum-4. Independent notes on power, lead times and… - [NVIDIA ConnectX-7 Specs & Procurement | GPU Smith](https://gpusmith.com/hardware/networking/nvidia-connectx-7): NVIDIA ConnectX-7 specs: Max bandwidth 400 Gb/s, Host interface PCIe 5.0 x16, Message rate 330-370 Mmsg/s. Independent notes on power, lead times and… ## Vendor and manufacturing research - [NVIDIA AI Server Vendors: OEM & ODM Design Landscape | GPU Smith](https://gpusmith.com/vendors): Independent index of every server vendor that designs NVIDIA AI systems in-house: OEMs, ODMs, hyperscale builders and edge designers, profiled and sourced. - [US AI Server Manufacturing Tracker | GPU Smith](https://gpusmith.com/vendors/us-manufacturing): Every announced US and nearshore Mexico AI server manufacturing site: vendor, location, disclosed investment and status, each row sourced. - [NVIDIA AI Server Market Statistics 2026 | GPU Smith](https://gpusmith.com/vendors/statistics): Sourced statistics on the NVIDIA AI server market: vendor revenues, NVL72 rack shipments, market shares, backlogs and lead times, each number cited. - [Supermicro NVIDIA AI Server Profile | GPU Smith](https://gpusmith.com/vendors/supermicro): Supermicro profile: what it designs in-house, NVIDIA platforms shipped (HGX B200, HGX B300, GB200 NVL72), manufacturing and procurement notes.… - [Dell Technologies NVIDIA AI Server Profile | GPU Smith](https://gpusmith.com/vendors/dell): Dell Technologies profile: what it designs in-house, NVIDIA platforms shipped (GB200 NVL72, GB200 NVL4, HGX H100/H200/B200), manufacturing and… - [HPE NVIDIA AI Server Profile | GPU Smith](https://gpusmith.com/vendors/hpe): HPE profile: what it designs in-house, NVIDIA platforms shipped (HGX H100/H200, GB200 NVL72, GB300 NVL72), manufacturing and procurement notes.… - [Lenovo NVIDIA AI Server Profile | GPU Smith](https://gpusmith.com/vendors/lenovo): Lenovo profile: what it designs in-house, NVIDIA platforms shipped (GB300 NVL72, GB200 NVL4, HGX B200), manufacturing and procurement notes. Independent,… - [Cisco NVIDIA AI Server Profile | GPU Smith](https://gpusmith.com/vendors/cisco): Cisco profile: what it designs in-house, NVIDIA platforms shipped (HGX B300 NVL8, HGX H100/H200, PCIe), manufacturing and procurement notes. Independent,… - [Supermicro vs Dell Technologies: NVIDIA AI Servers | GPU Smith](https://gpusmith.com/vendors/compare/supermicro-vs-dell): Supermicro and Dell Technologies compared as NVIDIA AI server vendors: what each designs in-house, which platforms they ship, manufacturing and… - [Dell Technologies vs HPE: NVIDIA AI Servers | GPU Smith](https://gpusmith.com/vendors/compare/dell-vs-hpe): Dell Technologies and HPE compared as NVIDIA AI server vendors: what each designs in-house, which platforms they ship, manufacturing and procurement… - [HPE vs Lenovo: NVIDIA AI Servers | GPU Smith](https://gpusmith.com/vendors/compare/hpe-vs-lenovo): HPE and Lenovo compared as NVIDIA AI server vendors: what each designs in-house, which platforms they ship, manufacturing and procurement notes. Sourced. ## Policies - [Privacy Policy | GPU Smith](https://gpusmith.com/privacy): What GPU Smith collects (very little), what it never does with your data, and how engagement material from regulated environments is handled. - [Terms of Use | GPU Smith](https://gpusmith.com/terms): Terms governing the use of gpusmith.com: what the published datasheets are and are not, intellectual property, and how engagement terms are set. - [Legal Notices | GPU Smith](https://gpusmith.com/legal): Firm identity, independence and affiliation disclosures, trademark notices, and content provenance for gpusmith.com. ## Additional collections - [Infrastructure articles](https://gpusmith.com/articles): Published engineering, procurement and operating research; see the current article catalog for individual topics. ## Crawl and URL discovery - [Root sitemap](https://gpusmith.com/sitemap.xml): Current site-page URL inventory. - [Article sitemap](https://gpusmith.com/articles/sitemap.xml): Current published article URLs. - [robots.txt](https://gpusmith.com/robots.txt): Crawl directives; this guide does not change access permissions.