VentureBeat Intelligence surveyed people at organizations with 100 or more employees, and consultants serving such organizations, in September 2026. We asked how enterprises measure the return on AI infrastructure, how hard their GPUs work, and which compute platforms they plan to adopt or evaluate next.
Executive summary
Most enterprises are planning new AI infrastructure without verified ROI on what they already run. Among enterprises planning to adopt a new, additional or replacement platform, 91% have no return-on-investment figure their finance team has reviewed.
Enterprises with no plans to change look about the same, at 89%, a statistically insignificant gap. Across all enterprises, only 10% have a verified ROI figure, yet 63% plan to adopt a new, additional or replacement AI infrastructure platform within 12 months.
Rigorous tracking does not guarantee verified ROI. Only 22% of enterprises say they track the return on AI infrastructure rigorously, and fewer than half of those have a figure their finance team has reviewed.
Among enterprises running AI in production, 81% say the GPUs they use sit idle at least half the time, about the same as 80% in August.
More enterprises now weigh token prices when they choose AI infrastructure. The share naming cost per 1M tokens as a factor rose from 11% in August to 23% in September, the one statistically significant change.
Among enterprises using AI in production, 68% use Google Cloud, Gemini or both, against 41% for OpenAI.
Many enterprises are unsure which of their models are open-weight, or whether any come from Chinese laboratories. Among enterprises that answered, 27% could not say with confidence which of their models are open-weight, or did not know. Most enterprises running open-weight models could not say whether any come from a Chinese laboratory, or did not know.
Finding 1. 91% of enterprises planning new AI infrastructure purchases have no verified ROI on what they already run, about the same as the 89% of enterprises with no plans
No enterprise that tracks return only partially has a figure that finance has reviewed and signed off on.
An ROI calculation ties what an enterprise spends on AI compute to what that spending produces, such as revenue, savings or faster work. Finance sign-off means the organization's finance team has reviewed the figure and accepted it. A figure a team measures itself, or estimates from assumptions, has not had that review.
We asked whether the organization can quantify the return on its AI infrastructure spending today. We then asked enterprises that said yes or partially what their return figure is based on.
Finding 1 — Can the organization quantify the return on AI infrastructure spend?
Base: 102 qualified respondents, since the question applies to every organization. One answer each.
Finding 1 — What the return figure is based on
Bases: 22 enterprises that track return rigorously and 38 that track it partially, the only groups the question applies to; group sizes in parentheses. One answer each. Columns are rounded, so they may not sum to 100%.
Across all enterprises, 10% have a return figure their finance organization has signed off on. Among the 22% that say they track return rigorously, 45% have a finance-reviewed figure, and 27% measure return themselves without finance validation.
Among enterprises that track return partially, 37% are not sure what their figure is based on, 32% rely on an estimate built from assumptions and 29% on team reports that nobody consolidates.
We then compared enterprises planning to adopt a platform within 12 months with those that are not.
Finding 1 — Have no finance-reviewed return figure
Base: all 102 qualified respondents. Of these, 64 plan to adopt a platform within 12 months and 38 do not.
Among enterprises planning to adopt a platform, 91% have no finance-reviewed return figure, against 89% of those with no plans, a statistically insignificant gap.
Final decision makers on AI purchasing and respondents with other purchasing roles claim rigorous tracking at 20% and 22%, a statistically insignificant gap.
Finding 2. 63% of enterprises plan to adopt a new, additional or replacement AI compute platform within 12 months, most of them within six months
AI-specialized clouds, non-Nvidia accelerators and Nvidia's next-generation GPUs are named by 24% to 29% of enterprises as the emerging option they would most likely evaluate.
Enterprises can add compute from the big clouds, from model providers through their APIs, from specialized AI clouds such as CoreWeave, Lambda, Crusoe and Nebius, or from their own data centers. Each choice trades price, supply and control differently.
We asked whether the organization plans to adopt a new, additional or replacement AI compute platform in the next 12 months, and which platforms the enterprises planning one are considering.
Finding 2 — Plan to adopt a new, additional or replacement platform
Base: 102 qualified respondents, since the question applies to every organization. One answer each.
Finding 2 — Platforms under consideration
Base: 62 enterprises planning to adopt a platform within 12 months that named a platform; one more is evaluating without a shortlist, and one named none. Several answers allowed, so shares sum past 100%. Specialized AI clouds share one card; other platforms under 5% are not shown.
Among enterprises planning to adopt a platform, 75% expect to do so within six months.
Of enterprises planning to adopt a platform that named one, Google Cloud and Oracle Cloud Infrastructure are each under consideration at 31%, and OpenAI at 29%, statistically insignificant gaps. Specialized AI clouds are under consideration at 18%.
We also asked which emerging compute option the organization is most likely to evaluate in the next 12 months.
Finding 2 — Emerging compute option most likely to be evaluated
Base: 102 qualified respondents. One answer each.
AI-specialized clouds are named by 29%, non-Nvidia accelerators by 26% and Nvidia's next-generation GPUs by 24%, statistically insignificant gaps.
Finding 3. Google leads AI in production: 68% of enterprises use Google Cloud or Gemini, against 41% for OpenAI
Enterprises with AI in production rate their primary platform 4.0 out of 5 for overall satisfaction, 3.9 for ease of implementation and 3.6 for value for money.
Enterprises can run AI on several platforms at once: a big cloud for infrastructure, a model provider for its models, and sometimes a specialized AI cloud for GPU capacity. Which provider an enterprise picks shapes what it pays and how easily it can switch.
We asked which AI compute platforms the organization uses in production and which one is primary. We report the answers of enterprises running AI in production, at scale or for some workloads.
Finding 3 — Platform in production use
Base: 74 enterprises running AI in production, the group the question applies to. Several answers allowed, so shares sum past 100%. Specialized AI clouds share one card; other platforms under 5% are not shown.
Finding 3 — Primary AI compute platform
Base: 72 enterprises running AI in production whose primary platform is one they said they use. One answer each. Shares are rounded, so they sum to 101%.
Finding 3 — Rating of primary AI compute platform, average out of 5
Base: 72 enterprises running AI in production whose primary platform is one they said they use. Scale of 1 (very poor) to 5 (excellent).
Google is the most widely used provider in production. Of enterprises running AI in production, 68% use Google Cloud, Gemini or both, against 41% that use OpenAI and 36% that use Microsoft Azure. Among enterprises running AI in production, Google Cloud or Gemini is the primary platform for 32%.
Specialized AI clouds are a small part of production use. Only 12% of enterprises running AI in production use CoreWeave, Lambda, Crusoe or Nebius.
Finding 4. 81% of enterprises running AI in production say the GPUs they use sit idle at least half the time
31% say those GPUs do useful work a quarter of the time or less.
Accelerator utilization is the share of time an enterprise's GPUs or other AI chips spend doing useful work. Capacity that is bought, reserved or rented by the hour costs money whether it runs or sits idle.
We asked enterprises for their average GPU or accelerator utilization in production.
Finding 4 — Average GPU or accelerator utilization in production
Base: 71 enterprises running AI in production. The 81% and 31% leave out the 4% that do not measure utilization. Three enterprises that said they do not operate GPUs are not counted. The survey did not ask whether enterprises own GPUs, rent them or reach them only through another company's API. Leaving out the 20% of enterprises that use only API providers does not change the 81%.
Leaving out the 4% that do not measure utilization, 81% put average utilization at 50% or less. Enterprises running AI at scale and those with some workloads in production differ by a statistically insignificant margin.
Finding 5. Cost per 1M tokens is now named by 23% of enterprises as a factor in choosing AI infrastructure, up from 11% in August
Integration with the existing cloud or data stack is named by 48% of enterprises, and performance by 42%.
Cost per 1M tokens is the price a provider charges to process a million tokens, the units of text a model reads and writes. The figure makes providers easy to compare, though the total bill also depends on how many tokens a workload uses. Integration with an existing cloud or data stack decides how much surrounding work a team must do.
We asked which factors most influenced the organization's choice of AI infrastructure.
Finding 5 — Factor that influenced the choice of AI infrastructure
Base: 102 qualified respondents. Several answers allowed, so shares sum past 100%.
Integration with the existing stack is named by 48% and performance by 42%, a statistically insignificant gap. Cost per 1M tokens is named by 23%, up from 11% in August, a statistically significant change.
Finding 6. 27% of enterprises that answered cannot say with confidence which of their models are open-weight or do not know
Among enterprises running open-weight models, 56% could not say whether any come from a Chinese laboratory, or did not know.
Open-weight models are models whose weights are publicly released, whether an enterprise hosts them itself or reaches them through a provider. Several widely used open-weight families, including DeepSeek, Alibaba's Qwen and Moonshot AI's Kimi, come from Chinese laboratories.
We asked whether the organization runs any open-weight models today, for production or for internal and experimental work. We then asked enterprises running them whether any originate from a Chinese laboratory.
Finding 6 — Runs open-weight models
Base: 88 respondents who answered. One answer each.
Finding 6 — Any open-weight models from a Chinese laboratory?
Base: 32 respondents who run open-weight models, the group the question applies to. One answer each.
Among enterprises that answered, 23% run open-weight models for production workloads and 14% only for internal or experimental work. Another 27% could not say with confidence which models are open-weight, or did not know.
Among enterprises running open-weight models, 31% said none come from a Chinese laboratory, and 13% said some came in through a vendor, platform or framework. No enterprise said it chose a Chinese-origin model deliberately.
Finding 7. Nvidia and Dell are each named by about 40% of enterprises for inference memory, while 16% have not addressed inference-memory limits or say they are not aware of them as a constraint
Open-source KV-cache tooling and model-level efficiency are each named by 21% of enterprises.
When a model answers a long prompt or conversation, it keeps intermediate results, called the KV cache, in memory so it does not recompute them for each new token. The cache grows with context length and the number of users, and it competes with the model for scarce GPU memory. Vendors now sell storage and software that move the cache off the GPU.
We asked which inference-memory or KV-cache approaches or vendors the organization is using or evaluating. Our question opened by stating that the binding constraint on inference is shifting from GPU compute to memory.
Finding 7 — Inference-memory approach or vendor used or evaluated
Base: 102 qualified respondents. Several answers allowed, so shares sum past 100%. No respondent chose both of the last two answers, so together they account for 16%.
Nvidia is named by 41% and Dell by 39%, a statistically insignificant gap.
None of the enterprises that have not addressed inference-memory limits, or say they are not aware of them as a constraint, named any vendor or approach.
What changed since the August Pulse
We asked the questions below in the same words in August and September. Only the change in the share naming cost per 1M tokens is statistically significant.
What changed: statistically significant
What changed: statistically significant
Statistically significant means the difference passes the test described in the methodology. Bases: August 109, September 102.
More September respondents named cost per 1M tokens as a factor in their choice of AI infrastructure.
What did not change: statistically insignificant
What did not change: statistically insignificant
Statistically insignificant means the difference does not pass the test described in the methodology. Bases, August then September: utilization 65 and 68; production use 68 and 74; primary platform 54 and 72; other rows 109 and 102.
The September figures in both groups are the same figures used in the findings above. Leaving out the flagged respondent described in the methodology changes no result in either group.
The bottom line: Most enterprises are planning new AI infrastructure without verified ROI on what they already run
Among enterprises planning new AI infrastructure purchases, 91% have no verified ROI on what they already run, about the same as the 89% of enterprises with no plans. Only 10% of all enterprises have a return figure their finance team has reviewed, yet 63% plan to adopt a new, additional or replacement platform within a year. Among enterprises running AI in production, 81% say the GPUs they use sit idle at least half the time.
We measured return tracking and platform plans side by side. We did not test whether a finance-reviewed return figure changes what enterprises buy.
In the next Pulse, VentureBeat Intelligence will measure whether more enterprises secure a finance-reviewed ROI measurement, and whether the share naming cost per 1M tokens as a buying factor keeps rising.
Respondent profile
| Respondent type | Respondents |
|---|---|
| Works on AI for the organization that employs them | 57% (58) |
| Consultant who works on AI for client organizations | 43% (44) |
Base: 102 qualified respondents. One answer each.
| AI deployment stage | Respondents |
|---|---|
| Some workloads in production | 47% (48) |
| Experimenting or proofs of concept | 26% (27) |
| AI in production at scale | 25% (26) |
| Not yet running AI workloads | 1% (1) |
Base: 102 qualified respondents. One answer each. Shares are rounded, so they sum to 99%.
| Organization size (main client, for consultants) | Respondents |
|---|---|
| 100–250 employees | 23% (23) |
| 251–1,000 employees | 35% (36) |
| 1,001–5,000 employees | 21% (21) |
| 5,001–10,000 employees | 13% (13) |
| 10,001 or more employees | 9% (9) |
Base: 102 qualified respondents. One answer each. Shares are rounded, so they sum to 101%.
| Role in AI purchasing | Respondents |
|---|---|
| Final decision maker | 43% (44) |
| Recommender or influencer | 42% (43) |
| User | 12% (12) |
| No involvement | 3% (3) |
Base: 102 qualified respondents. One answer each. Consultants answered for their main client's AI purchases.
| Job level | Respondents |
|---|---|
| Manager | 26% (27) |
| Individual contributor | 25% (26) |
| VP or director | 23% (23) |
| C-suite | 19% (19) |
| Partner or principal | 6% (6) |
| Other (teacher) | 1% (1) |
Base: 102 qualified respondents. One answer each.
| Industry (main client, for consultants) | Respondents |
|---|---|
| Technology or software | 28% (29) |
| Healthcare or life sciences | 17% (17) |
| Financial services, banking or insurance | 14% (14) |
| Manufacturing | 10% (10) |
| Education | 7% (7) |
| Telecommunications or media | 6% (6) |
| Retail or e-commerce | 5% (5) |
| Government or public sector | 4% (4) |
| Energy or utilities | 3% (3) |
| Professional services or consulting | 3% (3) |
| Transportation or logistics | 1% (1) |
| Other industries (printing, restaurants) | 2% (2) |
| Not stated (unreadable write-in) | 1% (1) |
Base: 102 qualified respondents. One answer each. Shares are rounded, so they sum to 101%.
Methodology
VentureBeat Intelligence fielded this VB Pulse survey in September 2026 and received 178 responses, of which 102 qualified. Respondents qualified if they work at an organization with 100 or more employees, or consult for one as their main client with direct access to its AI systems.
We excluded analysts, academics and journalists, staff at technology vendors serving their own customers, and respondents who chose "Other" on the role question. We also excluded consultants without direct access to their main client's AI systems, and organizations or clients under 100 employees.
The survey asked consultants to think of the client organization where they spent the most time on AI work in the past six months. We treat their answers to later questions about "your organization" as describing that client. Of the consultants, 70% have day-to-day access to that client's AI systems.
Each percentage is based on the respondents a question applies to, named in the note under each figure. Questions on platforms in use, the primary platform, ratings and utilization asked about production, so they use enterprises running AI in production. Utilization also leaves out three enterprises that said they do not operate GPUs.
Respondents experimenting or running proofs of concept also answered the utilization question. Of those who gave a figure, 88% put utilization at 50% or less. Their answers are not included in Finding 4.
One respondent said their organization is not yet running AI workloads, yet named a primary platform and reported GPU utilization. We flag this respondent rather than remove them. Sentences use the full base, and the table below shows the main figures both ways.
| Figure | Full base | Leaving out the flagged respondent |
|---|---|---|
| Track return rigorously | 22% (22 of 102) | 22% (22 of 101) |
| Finance-reviewed return figure | 10% (10 of 102) | 10% (10 of 101) |
| Plan to adopt a platform within 12 months | 63% (64 of 102) | 62% (63 of 101) |
| Most likely to evaluate AI-specialized clouds | 29% (30 of 102) | 30% (30 of 101) |
Production figures are unaffected, because the flagged respondent does not run AI in production. Every finding holds on both bases.
Two respondents who could not quantify a return, or called it not a priority, also answered what their return figure is based on. They are left out of that question, which applies only to enterprises that track return.
The questions on platforms in use, platforms under consideration, buying factors and inference memory allowed several answers. Their figures count people, not selections.
Some groups in this report have fewer than 40 respondents: enterprises tracking return rigorously, 22; tracking it partially, 38. Further groups: enterprises running open-weight models, 32; enterprises with no plans to change, 38. Among enterprises giving a utilization figure: at scale, 24; experimenting, 24; in-house production, 37. In-house production enterprises whose primary platform is one they use, 39.
Percentages for groups this small are less precise; for a group of 38, a result could differ by about 15 percentage points either way from the true figure. We describe a difference between groups as real only when a statistical test confirms it is unlikely to be due to chance. Differences that fail the test are shown and marked statistically insignificant.
A difference counts only when its test gives p below 0.05. Two percentages from separate groups use the two-proportion z-test, or Fisher's exact test when either group has fewer than 40 respondents. Two answers from the same respondents use the exact McNemar test.
A change from August counts only when the test confirms it twice: on all September respondents and on September's in-house respondents alone.
VentureBeat Intelligence surveys a self-selected group of VentureBeat readers and panel respondents. The results describe those respondents, not every enterprise, and apply to the wider market only as far as this mix of respondents allows.