Research

VB Pulse AI Infrastructure and Compute Report

Cost overruns stall AI projects for three times as many respondents as compute shortages do

October 2026

VentureBeat Intelligence surveyed 172 people in October 2026, and 156 qualified: people who work on AI at organizations with 100 or more employees, and consultants serving such organizations. We asked how their organizations buy and run AI compute, what their AI projects cost and how they judge whether the spending is worth it.

Executive summary

In the past 12 months, 38% of respondents say their organization delayed, scaled back or canceled a planned AI project because it cost more than planned. A shortage of computing capacity was named by 12%, and 5% named both. In all, 55% name at least one reason for delaying, scaling back or canceling a project. So the main brake on AI projects is cost, and compute supply comes a distant second.

Of respondents, 79% expect their organization's AI budget to increase over the next 12 months, and 19% expect it to increase by more than half. Yet only 26% say their organization judges whether AI spending is worth it by a measured financial return it could defend to its finance team. Among respondents expecting an increase, the figure is 27%.

Of respondents, 56% say at least one software vendor added an AI surcharge, raised a renewal price citing AI, or moved features their organization already had into a higher-priced tier in the past 12 months. Of those respondents, 40% say the changes raised their organization's annual software bill by 10% or more.

Among respondents whose organizations operate GPUs or accelerators, 57% put average utilization at half or less, so those chips sit idle at least half the time. Another 33% do not measure utilization at all, and of those who gave a figure, only 23% say it comes from continuous telemetry across the whole fleet.

Of respondents, 56% plan to adopt a new, additional or replacement AI compute platform in the next 12 months, after 67% adopted one in the past 12 months. Of respondents naming a listed platform, 76% use two or more.

Of respondents whose organizations run open-weight models in production, 46% use at least one model from a Chinese lab, which is 8% of all respondents.

Finding 1. Cost overruns stall AI projects for three times as many respondents as compute shortages do

38% of respondents say their organization delayed, scaled back or canceled a project because it cost more than planned; 12% cite a shortage of computing capacity.

Computing capacity here means access to the GPUs, accelerators and cloud capacity that AI workloads run on.

We asked whether, in the past 12 months, the organization had to delay, scale back or cancel a planned AI project for any of a list of reasons.

Finding 1 — Delayed, scaled back or canceled a planned AI project, past 12 months

38%
It cost more than planned
15%
Another reason
12%
We could not get enough computing capacity
9%
An infrastructure failure or outage
38%
No, none of these
8%
I do not know

Base: 156 qualified respondents. Several answers allowed, so shares sum past 100%.

Of respondents, 55% say their organization delayed, scaled back or canceled at least one project for one of these reasons, and 5% named both cost and a shortage of computing capacity. Cost was named by 38% and a shortage of computing capacity by 12%. Among in-house respondents alone, the figures are 39% and 10%.

Among respondents whose organizations run AI in production at scale, 45% named cost, against 40% of those with some workloads in production, a gap that is not statistically significant.

Finding 2. Most respondents expect bigger AI budgets, but only 26% judge AI spending by a measured financial return

79% expect their organization's AI budget to rise over the next 12 months. The most common yardstick, chosen by 44%, is an operational result such as time saved.

We asked how the organization's total AI budget over the next 12 months is changing, compared with the last 12 months.

Finding 2 — AI budget, next 12 months against the last 12

19%
Increasing by more than 50 percent
49%
Increasing by 10 to 50 percent
12%
Increasing by less than 10 percent
13%
Staying about the same
3%
Decreasing
4%
I do not know

Base: 156 qualified respondents. One answer each.

We also asked which best describes how the organization currently judges whether its AI spending is worth it.

Finding 2 — How the organization judges whether AI spending is worth it

44%
A measured operational result, such as time saved or volume handled, but not a financial return
26%
A measured financial return we could defend to our finance team
26%
Leadership judgment, without a formal measure
4%
We have not judged it yet
0%
We have judged it and concluded it is not currently worth it

Base: 156 qualified respondents. One answer each.

A measured operational result was chosen by 44% and a measured financial return by 26%. Among respondents expecting their organization's AI budget to increase, 27% chose a measured financial return, so most of those budgets are rising without a financial case behind them.

Base: 124 respondents expecting their organization's AI budget to increase.

We asked separately whether the organization can quantify the return on its AI infrastructure spend today. Of respondents, 21% say their organization tracks the return rigorously and 37% say partially. We asked respondents whose organizations track the return rigorously or partially what their return figure is based on, and 34% of them name a figure their finance organization has reviewed and signed off.

Bases: 156 qualified respondents; 91 respondents whose organizations track the return rigorously or partially, the group the second question was shown to.

Finding 3. Software vendors are charging extra for AI at 56% of respondents' organizations

Of those respondents, 40% say the changes raised their organization's annual software bill by 10% or more.

We asked how many of the organization's software vendors, in the past 12 months, added an AI surcharge, raised a renewal price citing AI, or moved features the organization already had into a higher-priced tier.

Finding 3 — Software vendors that added an AI surcharge, raised a renewal citing AI, or moved features to a higher tier

18%
None
19%
One
24%
Two to three
13%
Four or more
26%
I do not know

Base: 156 qualified respondents. One answer each.

We asked respondents who named at least one vendor roughly how much those changes added to their annual software bill.

Finding 3 — Added to the annual software bill

15%
Less than 5 percent
36%
5 to 10 percent
23%
10 to 20 percent
17%
More than 20 percent
9%
I do not know

Base: 88 respondents who named at least one vendor, the group the question was shown to. One answer each. The answer options "5 to 10 percent" and "10 to 20 percent" both include 10%, so a respondent whose bill rose by exactly 10% could have chosen either.

Finding 4. 57% of respondents running GPUs say those chips sit idle at least half the time

Another 33% do not measure utilization at all, and of those who gave a figure, only 23% say it comes from continuous telemetry across the whole fleet.

Utilization is the share of available GPU or accelerator capacity that workloads actually use.

We asked what the organization's average GPU or accelerator utilization in production is.

Finding 4 — Average GPU or accelerator utilization in production

9%
Under 10%
22%
10–25%
27%
26–50%
10%
Over 50%
33%
We don't measure it

Base: 128 respondents whose organizations operate GPUs or accelerators. Another 28 qualified respondents chose the answer for consuming compute through APIs instead and are left out. One answer each.

Of respondents whose organizations operate GPUs or accelerators, 57% put average utilization at half or less, and 10% above half. Among those at organizations with more than 10,000 employees, 21% do not measure utilization, against 38% at smaller organizations, a gap that is not statistically significant.

We asked those who gave a utilization figure how it is produced.

Finding 4 — How the utilization figure is produced

27%
Periodic sampling or spot checks
23%
Continuous telemetry across the whole fleet
19%
A figure reported by our cloud or platform vendor
15%
Telemetry on some clusters, extrapolated to the rest
8%
An estimate rather than a measurement
8%
Don't know

Base: 86 respondents who gave a utilization figure, the group the question was shown to. One answer each.

We also asked what share of AI compute spend goes to capacity that is reserved or provisioned but not used. Leaving out organizations that do not reserve capacity in advance, 31% of respondents whose organizations operate GPUs or accelerators put the unused share above 25%, and 22% do not know.

Base: 114 respondents whose organizations operate GPUs or accelerators, leaving out those that do not reserve capacity in advance. The base includes 25 that answered "Don't know" and 10 that pay for idle capacity but have not quantified it.

Finding 5. AI compute buyers keep shopping: 56% of respondents plan to adopt another platform within a year

56% plan to adopt a new, additional or replacement AI compute platform in the next 12 months, after 67% adopted one in the past 12 months. Of respondents naming a listed platform, 76% use two or more.

We asked whether the organization adopted a new, additional or replacement AI compute platform in the past 12 months.

Finding 5 — Adopted a new, additional or replacement platform, past 12 months

33%
Yes, we added one alongside what we already had
20%
Yes, our first one
13%
Yes, we replaced an existing one
15%
No, we considered it but did not move
12%
No, we did not consider it
7%
I do not know

Base: 156 qualified respondents. One answer each.

Over the next 12 months, 56% of respondents say their organization plans to adopt a new, additional or replacement platform: 11% within three months, 21% within three to six months and 24% within six to 12 months.

Base: 156 qualified respondents.

We asked which AI compute platforms the organization uses in production today, and which one is its primary platform.

Finding 5 — AI compute platforms in production

Microsoft Azure
52%
Use it
29%
Name it primary
OpenAI
51%
Use it
13%
Name it primary
Anthropic
48%
Use it
19%
Name it primary
AWS
41%
Use it
19%
Name it primary
Google (Gemini)
31%
Use it
4%
Name it primary
Google Cloud (GCP)
25%
Use it
5%
Name it primary
Oracle Cloud Infrastructure
9%
Use it
3%
Name it primary
Open-source self-managed stack
7%
Use it
2%
Name it primary
Another platform, written in
8%
Use it
6%
Name it primary

Bases: 153 respondents who named a listed platform in use, for the use column; three named only a platform outside the list and are left out of that column. 156 qualified respondents for the primary column. Several answers allowed for use, so that column sums past 100%; one answer for primary. Listed platforms used by under 5% are not shown.

Of respondents naming a listed platform, 76% use two or more. Microsoft Azure is named primary by 29%, and AWS and Anthropic by 19% each; Azure's lead over AWS is not statistically significant.

On average, respondents rated their primary platform 3.96 out of 5 for overall satisfaction, 3.81 for ease of implementation and 3.65 for value for money.

Base: 156 qualified respondents.

Finding 6. 46% of respondents running open-weight models in production use Chinese models

That is 8% of all respondents. Open-weight models run in production at 17% of respondents' organizations, and 33% use them only for internal or experimental work.

Open-weight models are models whose weights are publicly released, whether self-hosted or accessed through a provider.

We asked whether the organization runs any open-weight models in production today. The answer options separated production workloads from "only internal or experimental workloads," so some of the 33% may run internal tools in production.

Finding 6 — Runs open-weight models

17%
Yes, for production workloads
33%
Yes, but only for internal or experimental workloads
28%
No
8%
We could not say with confidence which models are open-weight
13%
Don't know

Base: 156 qualified respondents. One answer each.

We asked respondents whose organizations run open-weight models in production whose models they run.

Finding 6 — Open-weight models in production, by maker

58%
OpenAI (gpt-oss)
42%
DeepSeek
27%
Google (Gemma)
23%
Meta (Llama, Muse Glimmer)
19%
MiniMax
12%
Mistral
12%
Another maker, written in

Base: 26 respondents whose organizations run open-weight models in production, the only group asked. Several answers allowed. Makers named by fewer than three are not shown.

Of respondents whose organizations run open-weight models in production, 46% use at least one model from a Chinese lab: DeepSeek, Alibaba, Moonshot AI, Z.ai, MiniMax or Xiaomi. That is 8% of all respondents. The gap between OpenAI's gpt-oss and DeepSeek in that group is not statistically significant.

What changed since the September Pulse

September's figures cover only respondents who would pass October's screen, so they differ slightly from the published September report. We asked the questions below in the same words in September and October. October's respondents answered about larger organizations: 31% about companies with more than 10,000 employees, against 8% in September. Adjusting September's answers to October's mix of company sizes moves September's figures for each Finding by up to about four points. The seniority mix also changed: directors were 37% of October's qualified respondents, against 15% in September. Adjusting September's answers to October's seniority mix moves September's figures for each Finding by up to about four points. The respondent profile gives both months' seniority mix.

Most of October's respondents came through a research panel, and respondents' sources were not recorded in September. We cannot establish from these surveys why any of these figures differ between the two months.

Findings: statistically significant changes

Findings: statistically significant changes

14% → 33%
Run open-weight models only for internal or experimental workloads
6% → 26%
Not likely to evaluate any of the emerging compute options listed
4% → 33%
Do not measure GPU or accelerator utilization, among respondents whose organizations operate them
50% → 27%
Put utilization at 26–50%, among respondents whose organizations operate GPUs or accelerators
79% → 57%
Put utilization at half or less, among respondents whose organizations operate GPUs or accelerators

A Finding is a statistically significant change that also holds among in-house respondents alone and after September's answers are adjusted to October's company-size mix and, separately, its seniority mix. Bases, September then October: 86 and 156 for open-weight models; 99 and 156 for emerging options; 94 and 128 for utilization.

Directional changes

Directional changes

8% → 19%
Expect to use on-premises or colocation less
48% → 59%
Expect to use inference APIs more
26% → 15%
Most likely to evaluate non-NVIDIA accelerators
5% → 18%
Consume compute through APIs instead of operating GPUs
3% → 10%
Call the return on AI infrastructure not a priority
28% → 11%
Plan a new platform within three months
15% → 24%
Plan a new platform within six to 12 months
4.00 → 3.81
Average ease-of-implementation rating, out of 5
20% → 8%
Could not say which of their models are open-weight
44% → 62%
Expect to use specialized AI clouds about the same over the next 12 months
46% → 31%
Expect to use specialized AI clouds more

Directional means the change meets the looser test described in the methodology, or is statistically significant overall but not among in-house respondents alone. A change that is statistically significant but not after the company-size or seniority adjustment is also Directional. Bases, September then October: 89 and 151 for on-premises; 88 and 153 for inference APIs; 86 and 156 for open-weight models; 90 and 156 for specialized AI clouds; 99 and 156 otherwise.

Not statistically significant

Not statistically significant

23% → 17%
Run open-weight models for production workloads
35% → 44%
Expect to use hyperscalers more
8% → 3%
Expect to use hyperscalers less
17% → 10%
Put utilization above 50%, among respondents whose organizations operate GPUs or accelerators
22% → 21%
Track the return on AI infrastructure rigorously
64% → 56%
Plan a new platform within 12 months
48% → 54%
Some workloads in production
3.92 → 3.96
Average overall satisfaction with the primary platform, out of 5
3.63 → 3.65
Average value-for-money rating, out of 5

Not statistically significant means the change does not meet the threshold described in the methodology. Bases, September then October: 86 and 156 for open-weight models; 91 and 156 for hyperscalers; 94 and 128 for utilization; 99 and 156 otherwise.

Platform use, two readings

Platform used in production

15% → 48%
Anthropic (margin of error: September 7 points, October 8 points)
53% → 25%
Google Cloud (GCP) (margin of error: September 10 points, October 7 points)
38% → 52%
Microsoft Azure (margin of error: September 10 points, October 8 points)
41% → 51%
OpenAI (margin of error: September 10 points, October 8 points)
29% → 41%
AWS (margin of error: September 9 points, October 8 points)
44% → 31%
Google (Gemini) (margin of error: September 10 points, October 7 points)
24% → 9%
Oracle Cloud Infrastructure (margin of error: September 8 points, October 5 points)

Vendor usage is reported as two survey readings until a second month or a quarterly comparison confirms a change. Bases: 99 and 153. The methodology explains how margins of error are calculated.

What we did not compare

The primary-platform question now lists only the platforms a respondent uses, so its shares are not comparable with September's. The open-source platform answer was reworded in the platforms-in-use and platforms-considered questions, and a new answer was added to the selection-factor question, so neither is compared. The inference-memory question was reworded and its history restarts in October.

The question on whose open-weight models an organization runs replaced September's question about Chinese labs. The question on what a return figure is based on was answered by 62% of September's qualified respondents, and September's data does not show who was shown it, so we did not compare it. The questions on how a utilization figure is produced and on idle capacity were not in September's data. The questions on adopting a platform, delayed projects, vendor price changes, AI budgets, judging AI spending and AI risks are new this month.

The bottom line: Cost overruns stall AI projects for three times as many respondents as compute shortages do

Of respondents, 38% say their organization delayed, scaled back or canceled an AI project in the past 12 months because it cost more than planned, against 12% who could not get enough computing capacity. Over the same period, 56% say at least one software vendor added an AI surcharge, raised a renewal price citing AI or moved features to a higher-priced tier. Looking ahead, 79% expect their AI budget to increase.

Among respondents whose organizations operate GPUs or accelerators, 57% put utilization at half or less and 33% do not measure it. Among respondents expecting a budget increase, 27% say their organization judges AI spending by a measured financial return. Together, these point to AI spending that is rising faster than organizations can measure its return or its use.

In the next Pulse, VentureBeat Intelligence will measure whether the share of respondents whose organizations delay, scale back or cancel projects over cost changes. VentureBeat Intelligence will also measure whether the share of respondents running GPUs or accelerators whose organizations do not measure utilization changes.

Respondent profile

Respondent typeRespondents
Works on AI for the organization that employs them90% (140)
Consultant who works on AI for client organizations10% (16)

Base: 156 qualified respondents. One answer each.

Organization size (main client, for consultants)Respondents
100–250 employees16% (25)
251–500 employees20% (31)
501–1,000 employees10% (15)
1,001–5,000 employees14% (22)
5,001–10,000 employees10% (15)
More than 10,000 employees31% (48)

Base: 156 qualified respondents. One answer each.

Stage of AI deploymentRespondents
Some workloads in production54% (85)
AI in production at scale27% (42)
Experimenting or proofs of concept19% (29)

Base: 156 qualified respondents. One answer each.

Job levelSeptemberOctober
Director15% (15)37% (58)
Individual contributor25% (25)21% (33)
Manager26% (26)19% (30)
C-suite19% (19)11% (17)
VP7% (7)9% (14)
Partner or principal6% (6)3% (4)
Other1% (1)0% (0)

Bases: 99 September and 156 October qualified respondents. One answer each.

Methodology

Who answered. VentureBeat Intelligence fielded this VB Pulse survey in October 2026. The survey received 172 responses, of which 156 qualified. An early look at this survey, on the first 106 responses, went to VentureBeat editors in October; this report replaces it. Respondents qualified if they work on AI at an organization with 100 or more employees, or consult for one as their main client with day-to-day or partial access to its AI systems.

Who did not qualify. We excluded staff at technology vendors serving their own customers, analysts, academics and journalists, and organizations or clients with fewer than 100 employees. We also excluded in-house respondents with no direct role in their organization's AI work or who do not work directly on AI systems, and one respondent who did not answer the role question.

Consultants. Consultants answered about the client organization where they spent the most time on AI work.

Where respondents came from. Of qualified respondents, 87% came through a research panel, 7% through a VentureBeat email and 6% through other links.

Who each figure covers. Each percentage is based on the respondents a question applies to, named in the note under each table. Several questions were shown only to respondents whose earlier answers made them relevant, and their bases are those routed groups.

Several answers. The questions on platforms in use, open-weight model makers and delayed projects allowed several answers. Their figures count people, not selections.

Contradictory answers. On the delayed-projects question, one respondent chose "I do not know" alongside a reason, and another chose both "No, none of these" and "I do not know"; both answers are kept and flagged. The 55% who named at least one reason counts 86 respondents.

Small groups. Respondents whose organizations run open-weight models in production number 26, and respondents at organizations with more than 10,000 employees that operate GPUs or accelerators number 39. Margins of error here are conventional calculations that assume a random sample, which this is not. On that basis, a result for a group of 26 could differ by about 19 percentage points either way, and for a group of 39 by about 16.

Tests. We describe a difference as statistically significant 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 to a several-answer question use the exact McNemar test, and two answers to a one-answer question use the exact binomial test. Ratings use the Mann-Whitney test. The tests treat every response as independent, although responses were not deduplicated by organization, so more than one respondent may represent the same organization. A Directional change meets p below 0.10.

Changes from September. Each change is labelled a Finding, Directional or not statistically significant. A Finding must meet the 95% significance threshold overall and among in-house respondents alone. The difference must also stay statistically significant after September's answers are adjusted to October's company-size mix, and separately to its seniority mix. The in-house check repeats the test without consultants, whose share fell from 44% to 10%; it does not test other changes in who answered. September's figures cover the 99 of the published September report's 102 qualified respondents who would pass October's screen, which also excludes in-house respondents with no direct role in AI work. Compared with the published September report, the fall in expecting to use hyperscalers less is no longer statistically significant, and no other label changes.

September's open-weight question has a base of 86 because the September data holds no answer for 13 qualified respondents. On the question about expected use of each infrastructure approach, September's data includes a few blank, "N/A" or unreadable answers, which we left out of September's bases.

The inference-memory question was left blank by 53% of qualified respondents and is not reported.

Respondents chose to take this survey, and they come from a research panel and VentureBeat's readers. Because the sample is self-selected, the results describe these respondents and may not hold for all enterprises.

Published by VentureBeat Intelligence. © 2026 VentureBeat. All rights reserved.