Two-thirds of 170 enterprises surveyed now run AI workloads in production, and 29% run them at scale. Only 4% have not started. The average enterprise runs three infrastructure platforms simultaneously, with OpenAI (49%), Google Gemini (48%), Microsoft Azure (47%), and Google Cloud (42%) each present in roughly half of all stacks. When forced to name one primary platform, Azure leads at 26%. The specialized GPU neoclouds that dominate AI headlines, CoreWeave and Lambda, each appear in 3.5% of stacks and are named primary by 1% of enterprises combined.
Enterprises have reordered their buying criteria in ways that reveal production pressure over financial discipline. Integration with existing cloud infrastructure tops selection factors at 40%, but performance ranks second at 35% and GPU availability third at 24%, both ahead of total cost of ownership at 22%. That same logic governs success metrics: uptime and reliability leads at 51%, developer productivity at 39%, cost per million tokens at 31%. The problem is that the economics underneath this are largely invisible. Fewer than 47% of enterprises rigorously track AI compute costs and returns. Among the 155 enterprises operating their own GPUs, 69% report utilization at 50% or less, and 12% do not measure utilization at all. Value for money scores 3.87 out of 5, the weakest satisfaction dimension, against 4.14 for overall satisfaction.
The next round of spending is pointed at AI-specialized clouds, which top planned evaluation areas at 44% and carry the strongest net momentum of any infrastructure category at plus-36, yet current usage barely exists. Meanwhile, 62% of enterprises plan to switch or add a provider within 12 months, though the consideration set is dominated by the same incumbents already running their stacks. The full report is worth reading for what it reveals about the gap between where money is going next and what enterprises can actually see about where it went last. The utilization and cost-tracking numbers alone are a serious operational problem, not a planning one.
[READ ORIGINAL →]