IT leaders are consistently underestimating AI infrastructure costs due to production workloads differing significantly from pilot projects. The data layer, rather than model size, is identified as the primary cost driver.