AI Infrastructure Spend Jumps 36% as FinOps Teams Miss Half of Optimization Targets

AI spending is growing faster than organizations can control it

Enterprise AI infrastructure budgets reached an average of $85,521 per month in 2025, up 36% year-over-year, according to engineering consultancy MILL5. At the same time, only 51% of organizations report they can effectively manage and optimize their AI-related cloud costs. The gap between spending growth and control capability is widening, and it shows up in idle GPU hours, overprovisioned training clusters, and inference workloads that run at fixed capacity regardless of demand.

The problem is not abstract. MILL5's analysis of client deployments identifies four high-impact optimization categories with specific savings ranges: shutting down idle development resources recovers 15–25% of spend, implementing autoscaling for inference workloads saves 20–40%, rightsizing overprovisioned instances cuts 25–35%, and moving suitable training workloads to spot instances delivers 60–90% savings on those workloads. These are not future possibilities — MILL5 reports these as achieved results within an 8-week implementation window.

Cloud workload growth is accelerating waste, not eliminating it

Flexera's 2026 State of the Cloud report shows enterprise workloads in the cloud increased from 52% to 54% year-over-year, while SMBs saw a sharper jump from 55% to 63%. The shift matters because every percentage point of workload migration adds variable, usage-based cost to IT budgets. Flexera's historical survey data places wasted cloud spend in the 27–32% range, and the new report highlights AI, software licensing, and sustainability as emerging pressure points in cloud cost management.

The implication for buyers: a larger share of IT budget is now subject to cost volatility that traditional capital expenditure planning does not address. CFOs accustomed to predictable hardware depreciation schedules now face month-to-month swings in cloud bills driven by workload spikes, inefficient autoscaling policies, and untagged resources that escape governance. FinOps tooling moves from optional to required control system when 54% of enterprise workloads generate bills that can double in a quarter without architectural changes.

Flexera's focus on AI and sustainability also signals a competitive shift. Cloud cost platforms that cannot tag and allocate GPU hours, model training runs, or carbon-weighted region selection are losing ground to specialized AI cost observability tools and GreenOps platforms that tie workload placement to both cost and carbon metrics. Buyers should expect vendor pitches to emphasize AI workload controls, license-aware optimization (tying Oracle, Microsoft, and SaaS licenses into infrastructure decisions), and sustainability reporting as table-stakes features in 2026 RFPs.

Resilience is overriding cost as the primary buying criterion

IDC's 2026 IT buyer research reports a structural change in how infrastructure decisions are made: operational resilience has overtaken cost optimization as the primary driver. Concerns about hardware supply constraints increased by more than 15%, and IDC describes a "clear pivot toward resilience" with cybersecurity at the top of investment lists and multi-region cloud architectures accelerating. The buyer question is no longer "How do we optimize spend?" but "How do we keep the business running when things break?"

This reframes cost-optimization projects. Any initiative that reduces resilience — consolidating to a single region for lower egress costs, eliminating warm standby environments, tightening RTO/RPO windows to save on backup storage — is likely to be blocked. IDC's analysis gives IT leaders language to justify higher redundancy spend even when it raises cloud bills, because the alternative is business continuity failure during a supply chain disruption or regional outage.

For vendors, this changes competitive positioning. Cloud cost tools that model only steady-state efficiency are disadvantaged relative to platforms that can simulate cost versus resilience trade-offs. RFPs now require evidence of multi-region reference architectures and quantification of performance under disruption, not just cost per unit of compute. FinOps teams must demonstrate "cost per unit of resilience" as a new metric, which means tagging workloads by criticality, modeling failover costs, and showing how optimization decisions affect RTO and RPO.

What to watch: AI cost observability and resilience-aware optimization converge

The next 12 months will separate vendors who can integrate AI cost controls, license governance, and resilience modeling into a single platform from those who treat these as separate products. Buyers should ask three questions in every cost-optimization RFP: Can the platform tag and allocate GPU and model-level costs? Can it simulate multi-region failover scenarios and show the cost-resilience trade-off? Can it connect software license consumption to cloud resource decisions?

The 36% growth in AI infrastructure spend and the 15% increase in supply-constraint concerns are not independent trends. They converge in a single buyer requirement: infrastructure that costs less to run in normal conditions and survives disruptions without manual intervention. Vendors who answer only one side of that equation will lose deals to those who answer both.