Why AI Infrastructure Costs Keep Surprising IT Leaders
TechRadar Pro, Monday, July 13th, 2026
Aerospike's Srinivasan Seshadri says AI cost overruns live in the data tier, not model or token spend.
IDC projects AI infrastructure costs at Global 1000 companies will run 30% above current budgets by 2027, a gap the author attributes to the data layer rather than model size or token pricing.
Pilots run on narrow datasets with few users, while production agentic workflows expand one prompt into tens or hundreds of parallel lookups across CRM, session and vector systems under sub-100ms latency budgets.
The overrun accumulates from locally rational defensive choices: over-provisioning for spikes, duplicating data and layering services to insulate against slow dependencies.
Recommended fixes include matching each data tier to its access pattern, engineering for fan-out and tail latency, treating write consistency as correctness, and making KV cache reuse first-class infrastructure.