Enterprises are buying AI computing power for speed, but most have no idea what it costs
A survey of 170 companies finds two-thirds are running live AI systems, yet fewer than half can rigorously track what that computing infrastructure actually costs them.
Key points
- 66% of the 170 surveyed enterprises already run AI workloads in live production, as of July 2026.
- Fewer than half (47%) rigorously track what their AI computing costs or returns.
- Among companies that own their own GPUs (the specialised chips that do the heavy number-crunching AI needs), 69% report those chips run at 50% capacity or less.
- Microsoft Azure leads as the primary AI platform for 26% of enterprises surveyed, with Google Cloud second at 19%.
- 62% of enterprises plan to switch or add an infrastructure provider within 12 months.
Most companies that bet big on AI have already moved past the experiment stage. Two-thirds of the 170 enterprises surveyed in new research first reported by VentureBeat are running AI on live, real-world systems. Nearly a third say they are doing it at meaningful scale. Only 4% have not started at all.
That is a striking number. It means the decisions these companies are making now, about what hardware to buy and which platforms to use, are being made under real pressure, with real bills arriving every month.
So why can't most of them say what it costs?
The answer is uncomfortable: speed took priority, and the financial controls did not keep up.
When companies were asked what mattered most when choosing AI computing infrastructure, integration with their existing software stack came first (40% of respondents). Performance, meaning how fast the system responds and how much it can handle at once, came second at 35%. Access to GPUs came third at 24%. Total cost of ownership, the full long-term price of running a system, came fourth at just 22%.
That ordering makes sense for a team under deadline pressure. What it leaves behind is accounting.
Fewer than half of the enterprises surveyed (47%) rigorously track what their AI compute costs and what returns it generates. Even among companies running AI at full production scale, that figure only reaches 56%.
| Metric | Share of enterprises |
|---|---|
| AI in live production | 66% |
| AI at production scale | 29% |
| Rigorously track AI compute costs | 47% |
| GPU utilisation at 50% or below | 69% |
| Plan to switch or add a provider in 12 months | 62% |
Meanwhile, among companies that own and operate their own GPU clusters, 69% report those chips running at half capacity or less. Twelve percent do not measure GPU usage at all. These are expensive pieces of hardware sitting largely idle.
What platforms are companies actually using?
The big incumbents dominate. The average enterprise runs three AI platforms at once. OpenAI appears in 49% of stacks, Google's Gemini in 48%, Microsoft Azure in 47%, and Google Cloud in 42%.
When asked to name a single primary platform, Azure leads at 26% and Google Cloud sits second at 19%. The model providers, meaning companies that sell direct access to AI models rather than cloud computing broadly, together account for 35% of primary status: OpenAI at 14%, Gemini at 14%, and Anthropic at 8%.
The so-called neoclouds, specialist AI infrastructure providers that have attracted enormous investor attention, barely feature in current use. CoreWeave and Lambda each appear in just 3.5% of company stacks. The entire specialist-cloud category is named as a primary platform by just 1% of respondents.
What happens next?
Here is where the numbers get interesting. Despite barely using specialist AI clouds today, 44% of enterprises name them as the top area they plan to evaluate next. That category also carries the strongest forward momentum of any infrastructure type in the survey.
Whether that evaluation converts to real spending is a different question. Sixty-two percent of enterprises say they plan to switch or add a provider within 12 months, but when asked which providers they are considering, they name the same large incumbents they already use.
For ordinary employees and customers, the practical effect of all this is simple: the AI tools your company runs are probably live and growing. Whether anyone inside that company has a clear handle on what they cost is, right now, genuinely uncertain.
Common questions
Does this survey speak for all businesses?
No. The 170 respondents are self-selected organisations actively building AI infrastructure, and 57% have more than 1,000 employees. Smaller businesses, and companies not yet investing in AI, are not represented.
What does GPU utilisation at 50% mean in plain terms?
It means companies are paying for computing power they are not using. A chip running at 50% capacity is roughly like renting a warehouse and leaving half of it empty, while still paying full rent.
Should employees be worried about their company's AI spending?
Not on safety grounds from this data. The concern here is financial efficiency, not risk. Companies spending heavily on AI without tracking returns will eventually have to account for that gap, which could affect future investment decisions in those organisations.
