83% of enterprises report GPU utilization at 50% or less. Fewer than 44% can rigorously track what their AI compute costs. Yet across 107 enterprises surveyed by VentureBeat Pulse Research in June 2026, spending is accelerating anyway. Only 21% run AI in production at scale. The rest are still building, and their infrastructure footprint is about to grow fast.
The current stack is hyperscalers and model APIs. Google Cloud leads at 48%, followed by other general-purpose clouds and providers like OpenAI and Anthropic. Specialized GPU neoclouds, CoreWeave, Lambda, Crusoe, register at or near zero in current deployments. That makes this finding sharp: 45% of enterprises plan to evaluate AI-specialized clouds in the next 12 months, the single largest planned evaluation category. 64% plan to switch or add an infrastructure provider within twelve months. 38% plan to do it within the next quarter. Buying decisions turn on stack integration at 41% and total cost of ownership at 35%. Cost per million tokens drives just 8% of decisions.
The piece is worth reading in full for two reasons. First, the compute gap framing is specific and data-supported, not a vague warning about AI costs. Second, the tension between Finding 2 and Finding 3, near-zero current use of specialized clouds versus 45% planning to evaluate them, is the clearest signal in the data that a re-platforming cycle is starting among mid-market enterprises. The report also flags that the shift from GPU compute to memory bandwidth as the frontier inference constraint is barely on the radar, with roughly one in five enterprises either unaware of it or not yet addressing it. That detail alone is worth the read.
[READ ORIGINAL →]