OEM Evaluation
Enterprise Server OEM Comparison
| OEM | Strengths | Considerations |
|---|---|---|
| Dell Technologies (PowerEdge) | Broad portfolio, strong support, OpenManage management | Premium pricing at high volume |
| HPE (ProLiant/Synergy) | iLO management, strong ISV certifications, GreenLake as-a-service | Complex licensing for management software |
| Lenovo (ThinkSystem) | Competitive pricing, strong HPC portfolio, XClarity management | Smaller US support footprint than Dell/HPE |
| Supermicro | Competitive pricing, GPU server specialization, OCP designs | Smaller support organization, fewer ISV certifications |
| NVIDIA (DGX) | Purpose-built AI training systems, NVLink fabric, NVIDIA support | Premium pricing, limited to GPU workloads |
Configuration Requirements
Server configuration must be validated against workload requirements before procurement. The most common configuration mistakes are under-provisioning memory (causing swap usage and performance degradation), over-provisioning CPU cores (paying for cores that are never used), and selecting the wrong storage interface for the workload's I/O profile.
Memory configuration
Size memory for the peak working set of the workloads the server will run, not the average. Memory that is insufficient causes swap usage that degrades performance by 10–100x. Memory that is over-provisioned wastes capital but does not degrade performance.
CPU configuration
For virtualization hosts, size CPU based on the total vCPU count of the VMs the host will run, with a 4:1 overcommit ratio as a starting point. For bare-metal workloads, size based on the specific workload requirements.
Storage configuration
NVMe for databases and high-IOPS workloads. SAS for high-capacity spinning disk. SATA for cold storage. Do not use SATA for database workloads, the latency is inadequate.
Network configuration
25 GbE is the minimum for virtualization hosts. 100 GbE for storage-intensive workloads. 200–400 Gb/s for GPU servers. Dual-port NICs for redundancy.
Power supply configuration
Redundant power supplies (1+1 or 2+2) for all production servers. Titanium efficiency rating for energy cost optimization. Size power supplies for the maximum configuration, not the initial configuration.
GPU Hardware Procurement
GPU hardware procurement requires planning that is qualitatively different from standard server procurement. NVIDIA H100 and H200 GPUs have lead times of 12–26 weeks from authorized OEMs. Organizations that do not plan procurement timelines accordingly consistently miss their AI deployment targets.
GPU allocation and lead times
Verify facility readiness first
Confirm power capacity, cooling capability, and network bandwidth before ordering GPU hardware. Hardware that arrives before the facility is ready creates storage and handling costs.
Order through authorized channels only
NVIDIA GPU hardware purchased through unauthorized channels does not carry manufacturer warranty. The gray market for GPU hardware is active, verify OEM authorization before purchasing.
Plan for software licensing
NVIDIA AI Enterprise software licensing is required for production AI deployments. License cost should be included in the TCO analysis.
Include liquid cooling in the procurement
GPU servers at full density require liquid cooling. Procure the cooling infrastructure at the same time as the GPU hardware, not after.
Support Contracts
Support contract quality determines the operational cost of hardware over its lifecycle. The key terms to evaluate are response time (how quickly a technician arrives on-site after a failure is reported), parts availability (whether the OEM maintains parts inventory for the specific hardware), and the escalation path for critical failures.
Support Contract Tiers
| Tier | Response Time | Parts | Best For |
|---|---|---|---|
| Basic (NBD) | Next business day on-site | Shipped next business day | Development, test, non-critical workloads |
| 4-Hour | 4-hour on-site response | Parts on-site within 4 hours | Business-critical production workloads |
| Mission Critical | 2-hour on-site response | Parts pre-positioned on-site | Mission-critical, 24/7 production workloads |
Procurement Models
Direct OEM Purchase
Pros: Best pricing at volume, direct manufacturer relationship, full warranty coverage
Cons: Requires procurement capability, longer lead times for custom configurations
Best for: Large organizations with established OEM relationships
Authorized Reseller
Pros: Single vendor for multi-OEM deployments, deployment services, financing options
Cons: Markup over direct OEM pricing, reseller quality varies
Best for: Organizations that need deployment services or multi-OEM procurement
As-a-Service (HPE GreenLake, Dell APEX)
Pros: OpEx model, capacity flexibility, OEM-managed hardware
Cons: Higher total cost over 5 years, less control over hardware configuration
Best for: Organizations that prefer OpEx over CapEx, or need rapid capacity scaling