Root Cause

Procurement processes that evaluate hardware on acquisition cost per server, without accounting for energy cost, maintenance cost, or the performance delivered per dollar spent.

Consequence

Hardware that is cheaper per unit but less efficient per workload produces higher total cost over 5 years, and may not support the workloads it was purchased to run.

Prevention

Evaluate compute on total cost of ownership per workload over 5 years. Include energy cost (watts per unit of compute), maintenance cost, and the performance delivered for the specific workloads.

Root Cause

Memory is the most commonly under-provisioned compute resource. Organizations size memory for average utilization rather than peak working set, or apply a standard memory configuration to all servers regardless of workload.

Consequence

Servers that run out of physical memory use swap space on storage, which is 100–1000x slower than RAM. Applications that swap under load deliver dramatically degraded performance that is difficult to diagnose and expensive to remediate.

Prevention

Size memory for the peak working set of the workloads the server will run, plus 20% headroom. Do not apply a standard memory configuration to all servers, workload requirements vary significantly.

Root Cause

Organizations order GPU hardware to meet AI deployment timelines without verifying that the facility has adequate power capacity, cooling capability, and network bandwidth.

Consequence

GPU servers that arrive before the facility is ready create storage and handling costs. GPU servers deployed in facilities that cannot support their power and cooling requirements throttle performance, delivering a fraction of their rated capability.

Prevention

Verify facility readiness before ordering GPU hardware. Confirm power capacity, cooling capability (liquid cooling for 10+ kW/rack), and network bandwidth. Allow time for facility upgrades before hardware delivery.

Root Cause

Organizations track hardware age but not manufacturer support status. Hardware that is within its useful life may have passed its end-of-support date, particularly for servers purchased with short support terms.

Consequence

End-of-support hardware cannot receive security patches or firmware updates. Vulnerabilities discovered after end-of-support cannot be remediated without hardware replacement. This is a documented attack surface that threat actors actively exploit.

Prevention

Track manufacturer support status for all compute hardware: not just age. Plan refresh programs based on end-of-support dates, not just hardware age. Negotiate extended support terms at procurement for hardware that will be retained beyond the standard support period.

Root Cause

Gray market GPU hardware is available at lower prices than authorized OEM channels. Organizations under budget pressure purchase from unauthorized resellers without understanding the warranty implications.

Consequence

Hardware purchased from unauthorized resellers does not carry manufacturer warranty. Failures are not covered by OEM support. The cost of replacing failed hardware without warranty coverage exceeds the savings from the lower purchase price.

Prevention

Purchase hardware only from OEM-authorized resellers. Verify authorization status before purchase. Request authorization letters from the reseller: current, not expired.

Root Cause

Organizations deploy new compute hardware directly to production without integration testing, to meet deployment timelines or because testing is perceived as overhead.

Consequence

Compatibility issues between new hardware, existing software, and the production environment are discovered after deployment, when remediation is more expensive and more disruptive than pre-deployment testing.

Prevention

Require integration testing as a deployment step. Define the tests that must pass before new hardware is placed in production. Include performance benchmarks for the specific workloads the hardware will run.