GPU Procurement & Capex Benchmarking

The GPU compute market splits sharply: volume providers bulk at scale, boutique specialists command premium. Supermicro's HGX B300 8-GPU system costs $485,000 per unit as of March 2026 (indicative Q1 2026 pricing; hardware costs are rising and vary significantly by volume, configuration, and supplier relationship).

That's $60,625 per GPU; bulk orders above 100 units unlock 8-12% discounts. This configuration delivers 640 GPU-teraflops aggregate compute in a 4U form factor.

Current-generation Blackwell accelerators command premium pricing relative to secondary-market H100 stock, reflecting yield maturity and application-specific optimisations. Racks deploy either eight single-socket or four dual-socket topologies depending on thermal envelope and interconnect demands. The market now supports modular configurations. Four, six, or eight GPU variants within the same chassis family allow operators to match capex precisely to workload density without over-specifying hardware.

Spectrum-X800 800Gbps fabric switching costs $191,407 per unit as of Q1 2026 (indicative pricing; hardware costs are rising and vary significantly by volume, configuration, and supplier relationship). This underpins cluster interconnect for both intra-rack and inter-rack GPU communication, supporting Blackwell's native 1.8Tbps per-GPU I/O bandwidth.

Direct attach cabling adds $8,200-$12,400 per switch depending on port density. RDMA-capable networking is mandatory.

Standard Ethernet introduces unacceptable latency variance in distributed training. Cross-rack fabric requires spine-leaf topology with dual 400Gbps spine switches at $124,000 per unit minimum. A 32-rack GPU cluster typically deploys four spine switches and eight leaf switches, driving networking capex to approximately $1.2M for compute fabric alone. Storage fabric interconnect requires separate 100Gbps commodity switching at roughly $8,400 per 64-port unit.

PeakAIO 100TB NVME SSD storage arrays quote at $150,000 per unit for enterprise-grade configurations (March 2026). Distributed parallel filesystems like Lustre or WekaIO require licensing ($2,400-$6,800 per node) plus dedicated metadata servers at $18,000-$24,000 per triplet.

A 32-rack GPU cluster demands 400-600TB aggregate staging capacity. Model size and training iteration frequency drive the variance.

Object storage via S3-compatible interfaces (Ceph or commercial platforms) adds $120,000-$200,000 capex for production-grade availability. Bandwidth to external data lakes or cloud buckets requires dedicated 400Gbps circuits at $14,000 per month recurring cost. This overhead persists across the entire operational lifetime and often dominates total storage cost over multi-year cycles.

Disintermediate's benchmark spans 120+ vendors globally, capturing 12,800+ discrete pricing data points updated monthly. Single-rack (8-GPU Blackwell) configurations land at $650,000-$780,000 all-in.

A 16-rack cluster reaches $10.8M-$13.2M capex before power, cooling, or real estate. Multi-tenant configurations requiring isolation add 18-24% premium over bare-metal.

Small providers (< 10 racks) incur 12-18% cost disadvantage versus hyperscalers due to engineering overhead per deployment. Regional variance peaks at 31% between North American and APAC suppliers, driven by logistics and customs duty. Secondary-market GPU refurbishers trade older inventory at 35-50% discounts versus new units, creating arbitrage opportunities for cost-conscious operators.

Supermicro HGX B300 8-GPU systems: $500,000–$520,000 per unit (indicative Q1 2026 pricing; varies by volume and supplier relationship); per-GPU capex floors at $62,500–$65,000 with bulk discounts

Spectrum-X800 fabric switching requires $191,407 per 800Gbps switch; 32-rack clusters need ~$1.2M networking capex

Storage staging at 400-600TB per 32-rack cluster costs $150,000-$200,000 in hardware plus metadata licensing

Total cluster capex: 16-rack configurations cost $10.8M-$13.2M before power, cooling, or real estate