- In a VentureBeat Intelligence survey, 64% of final AI purchasing decision makers say return is tracked rigorously, against 34% of other respondents.
- The survey cannot say why. Decision makers may see figures others do not, or may be scoring their own decision.
- Ask who produces your AI return figures, by what method, and whether anyone who did not approve the spending can reproduce them.
When a company says it tracks the return on its AI spending, who is saying it? In a new VentureBeat Intelligence survey, the claim comes far more often from people who control the purchase. Respondents with the final say on buying AI were much likelier to call their organization's return tracking rigorous: 64%, against 34% of everyone else. The survey cannot explain the gap. It does raise a question worth asking about any AI return figure.
What was measured
In August, VentureBeat Intelligence polled staff at companies with at least 100 employees about how their organizations pick, run and judge AI infrastructure. In all, 109 people answered, and 68 said their organizations run AI in production. The answers are self-reported.
The survey's headline concerns platforms. Google came first on the question of which platform a company relies on first. Within a group of 54 production respondents, 48% named a Google service, split between Gemini and Google Cloud. OpenAI held that spot for 22% and Microsoft Azure for 13%. VentureBeat says Google's share combines two products. On plain use, Gemini and OpenAI are too close to call.
Who says the return is tracked
Across all respondents, 48% claim they track the return on AI infrastructure rigorously. Another 38% say they do so partially, and 15% cannot yet quantify it.
The split by role is wide but not exclusive. Decision makers are 46% of respondents yet account for 62% of the rigorous claims. That leaves 38% of those claims coming from everyone else. Also, 34% of non-decision makers make the claim, and 36% of decision makers do not.
The pattern was not a one-month result. July's respondents showed a similar gap, 62% against 35%. It also appeared at organizations not yet running AI at scale, where 61% of decision makers made the claim against 25% of others.
VentureBeat says it cannot tell why. Buyers could have access to numbers that their colleagues never see. Or buyers could be scoring their own decision, which gives them a stake in the answer. The survey does not separate the two readings.
Satisfaction tracks the claim. On a five-point value-for-money scale, production respondents claiming rigorous tracking averaged 4.24. The rest averaged 3.50. VentureBeat cannot say which way the link runs. Measuring may help a team see what it buys. The cause may also run the other way: contented buyers may describe their tracking in kinder terms.
What the survey does not connect
The next figure is independent of the tracking gap, and the survey does not link them. Of 65 production respondents who gave an average accelerator utilization, four in five put it at 50% or lower. Accelerators are the GPUs and other chips that run AI. Owned or reserved ones cost money whether they work or sit idle.
Part of that idle time may be deliberate. VentureBeat notes that holding spare capacity for busy periods can pull the average down. It also points to a separate telemetry study. A 2026 Cast AI report on Kubernetes optimization reached 5% average GPU use in the clusters it examined. That covers different organizations and a different measure, so the figures cannot be compared.
Few respondents cite unit price. Just 11% listed cost per 1M tokens, the fee for processing a million units of text, among the factors behind their choice. Some 40% listed fit with the cloud or data stack already in place. Performance drew 37%, security or compliance 34% and total cost of ownership 26%. Respondents could give several answers.
The write-up does not tie chip use or buying factors to the tracking claim. Among respondents who track token cost, 44% claim rigorous tracking. Among those who do not, 49% do. VentureBeat calls that too close to call. Whether platform choices were carefully examined is a question these figures leave open.
Teams also lean on general yardsticks. The most common are not specific to AI: 56% of respondents watch uptime and 46% watch developer productivity. Latency, the most-used AI-specific measure, is used by 28%. VentureBeat warns that a team watching only general measures could overlook an AI workload that works but is slow or costly.
One way to read it
One hypothesis: a system is bought because it fits what a company already runs, and the buyer later vouches for it. The survey cannot establish that sequence. Integration was one factor among several, and the data do not link buying factors to who vouches. Nobody needs to be dishonest for the issue to matter. The person with the most context may also have the most at stake.
The limits are plain. This is one survey with small subgroups, and it reports what people say, not audited results. It does not show that firms are overspending.
Questions to put to your team
Ask who produces the AI return figures, and by what method. Ask whether someone who did not approve the spending can reproduce the number. Ask for average accelerator utilization on owned or reserved capacity, and how much idle time is planned headroom for peak demand. Ask whether workload cost is tracked at all, beyond uptime.
Finally, ask whether the platform was chosen mainly for price or for fit with existing systems, and whether that comparison was written down. Both can be sound reasons.
A return figure earns trust when someone who did not sign the purchase can reproduce it.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: VentureBeat.
