We asked 50 of them. They discount the labs' motives, back more regulation almost unanimously — and still would not give regulators the throttle. What they want is independent evaluation, and 60% would pay for it.
Each conclusion names the questions carrying it. Where a conclusion rests on a subgroup, the base is stated.
46.0% read lab slowdown proposals as competitive positioning against 20.0% reading them as genuine safety, while 70.0% of open-text responses carry a reason to slow down and 44.0% cite safety and guardrails specifically. Both readings come from the same 50 people. The skepticism is about who is asking, not about whether the question is real.
72.0% want more regulation and no respondent would remove government from frontier development entirely, yet only 14.0% would have government regulators set the pace — the lowest of the four named options, against 40.0% for independent third-party evaluators. The sample wants government to set a floor and evaluators to judge clearance against it.
Among the 36 favoring more regulation, safety and misuse risk draws 86.1% and clearer liability allocation 52.8%, while leveling the competitive field draws zero and pace relief draws one respondent. The protectionist reading of enterprise regulatory demand is not supported here.
70.0% would be more willing to deploy with independent evaluation, 74.0% would be more confident in a lab accepting it, and 60.0% would pay a premium. The ceiling is the caveat — 40.0% of the full sample caps the premium at 10.0% and only 20.0% would exceed it.
Among the 23 respondents reading lab proposals as competitive positioning, 60.9% still report more confidence in a lab that publicly paces and accepts independent evaluation. The difference between the two questions is third-party verification.
If progress slowed, 62.0% would feel their deployments were more secure and 62.0% more knowledgeable about them, and 52.0% would grow more interested in open-source models. The compute statement is the only one in the battery where disagreement leads agreement, 30.0% to 26.0%.
40.0% are one or more generations behind or deliberately frozen, including 12.0% falling further behind, and only 16.0% adopt within weeks of release against 44.0% within one to two quarters. In a sample screened for AI familiarity and decision authority, that is the adoption reality against which any pacing debate takes place.
94.0% knew of the incident and 74.0% knew details; 72.0% took at least one action, 58.0% added a technical control, and 20.0% paused or slowed agent deployments. Containment outnumbers retreat nearly three to one.
Respondents cite 3.9 providers on average and 86.0% cite three or more, with Microsoft at 84.0% and Anthropic at 80.0% ahead of OpenAI at 58.0% and Google at 52.0%. Self-hosted open weights reach 26.0%, against the 52.0% who say a slowdown would raise their interest in open source.
Every question in the instrument, distribution first and a clearly labeled read second.
When major AI labs propose regulations or voluntary measures to slow down AI development, do you consider these proposals to be genuine safety measures, or more about competitive positioning?
Skepticism about motive outweighs credit by more than two to one, and it is broad rather than concentrated: 80.0% of the sample declines to call the proposals mostly or entirely genuine. Hold this reading against Q11, where the same sample responds favorably to labs that pace and accept outside testing. Motive and behavior are being judged separately.
In a sentence or two, should the frontier labs slow down the pace of AI development? Why or why not? (Coded open text)
The argument against a slowdown is almost never an argument that the technology is safe. Across the 19 responses carrying a no code, the reason given is who else fails to slow down — China by name in several, competitors generally in others. That distinction matters commercially: this sample is not disputing the risk, it is disputing the unilateralism. Two respondents state the condition explicitly, that a slowdown works only if it happens globally.
Which of the following best describes your view on government regulation of frontier AI model development?
The zero on “no government role” is the finding here. The deregulatory position has no constituency in this sample, and the 18-to-1 ratio between the more-regulation and less-regulation nets leaves little ambiguity about direction. Among the 23 respondents who read lab proposals as competitive positioning at Q1, 56.5% still want more regulation — distrust of the labs' motives does not translate into distrust of oversight.
What are the primary reasons you favor more regulation? Asked only of the 36 respondents selecting somewhat more or significantly more regulation at Q3.
Two readings the data carries directly. First, the regulation this sample wants is about harm and who owns it: safety at 86.1% and liability at 52.8% account for the demand, and liability allocation is the one enterprises cannot solve for themselves through procurement. Second, the self-interested motives score near zero — one respondent for pace relief and none for leveling the field. The argument that enterprise demand for AI regulation is protective rather than substantive finds no support in these 36 responses.
What are the primary reasons you favor less regulation? Asked only of the two respondents selecting somewhat less regulation at Q3.
In your opinion, who should set the pace of frontier AI model development?
Q3 and Q6 look contradictory and are not. 72.0% want more regulation while only 14.0% want regulators holding the throttle. The consistent reading across both is a division of labor: government sets the floor — the safety standards and liability rules Q4 identifies — and technically credible evaluators judge whether a given model clears it. Independent evaluation outpolls government by nearly three to one, which is the single most actionable preference in this study.
Over the past 12 months, how well has your organization kept pace with new model and AI capability releases from your primary vendors?
Two things are true at once. Four in 10 organizations in an informed-buyer sample are a generation or more behind, and 12.0% are losing ground rather than gaining it. But being behind does not explain the appetite for a slowdown: among those behind and closing, 58.3% carry only yes-to-slowdown codes at Q2, against 60.0% of those keeping current — no separation at all. Whatever is driving the slowdown sentiment in this sample, catch-up relief is not it.
Would independent third-party evaluation of frontier models — testing and certification of a model's safety and performance by an outside organization, rather than by the lab that built it — make your organization more willing or less willing to deploy AI in production?
Would your organization pay a premium for independently evaluated and certified models? If so, how much more, relative to an equivalent uncertified model?
Willingness converts to price, but not fully and not far. Among the 35 respondents more willing to deploy under Q8, 74.3% would pay a premium; among the 14 reporting no change, 28.6% would. That 46-point separation is the strongest relationship in the study and it runs the direction a certification business would need. The constraint is the level: two-thirds of the would-pay group sits at or below 10.0%, so the demonstrated appetite is for certification as a low-single-digit-to-10.0% line item, not a premium tier. Both readings rest on subgroup bases of 35 and 14 and are directional.
In this scenario, how much do you agree or disagree with each of the following statements about your organization? Figures shown are agree nets.
The benefit this sample attaches to a slowdown is comprehension and control, not consumption. Three of four statements carry agree nets at or above 52.0% with disagree nets at or below 18.0%; the compute statement inverts, and it is the only one that does. That single inversion is the finding worth carrying forward: technology leaders in this sample do not expect a frontier pause to redirect their budget into GPUs or cloud compute. The open-source reading at 52.0% is worth setting against S3, where 26.0% already host open weights themselves — stated interest runs at twice current practice.
Does a lab publicly committing to pace its development and accept independent evaluation make you more confident or less confident building on that lab's models?
Set this beside Q1 and the pair resolves. The same sample that reads lab safety proposals as competitive positioning by 46.0% to 20.0% responds to lab pacing commitments with 74.0% more confidence. Within the 23 respondents who called the proposals mostly or entirely about positioning, 60.9% still report more confidence in a lab that paces and submits to outside evaluation. The distinction this sample is drawing is between a lab's stated motive, which it discounts, and a lab's verifiable commitment, which it does not. The commitment carries because a third party checks it.
Before today, how familiar were you with the details of this incident?
What, if anything, did the incident change at your organization?
The response was containment rather than retreat. Technical controls at 58.0% outnumber pauses at 20.0% by nearly three to one, and five respondents did both. Familiarity tracks action: 70.3% of the 37 who knew details took at least one step, as did nine of the 10 who knew only that it happened. This is the one question in the study measuring behavior rather than opinion, and the behavior points the same direction as the opinions: the sample tightens controls and keeps deploying.
All 50 respondents qualified on three screens, so this is an informed-buyer sample rather than a general enterprise sample.
How familiar are you with your organization's tech stack in each of the following areas? Figures shown are very or extremely familiar.
This sample knows AI better than it knows its own security stack. Read the Q13 security responses with that asymmetry in mind: the agent-control actions reported there come from people closer to the model layer than to the security operations layer.
When your organization makes decisions about AI models — which models to use or deploy — what best describes your role?
Verbatim responses to Q2, reproduced as written.
Yes; these are being rushed without proper guardrails and integration with security stack in place. These controls are afterthoughts and risks are significant.
There's no reason to. It's for competitive advantage and access to premium hardware. The current leaders are only saying it because their costs are astronomical and they need to temper Wall Street and the broader economy.
Pacing development allows critical time to build safety guardrails and prevent alignment failures like autonomous rogue agents or systemic cyber threats. However, a unilateral slowdown by Western labs risks ceding crucial strategic and technological leadership to global adversaries who are unlikely to pause.
Fortune 500 / Global 2000 members (n = 28) against respondents in neither index (n = 22). Both cells are entirely large-enterprise, so this is a membership contrast rather than a size contrast.
The index cell is 100.0% public company against 31.8% of the non-index cell, and skews toward IT/TelCo, 35.7% versus 4.5%. Read every difference below as partly a public-company difference.
CxO titles sit disproportionately in the non-index cell, 36.4% versus 7.1%, as do final decision makers, 36.4% versus 10.7%. This is the only cross-group gap in the study to reach conventional significance (Fisher p = 0.01).
AWS Bedrock reaches 60.7% of index respondents against 31.8% of non-index (Fisher p = 0.05), while Microsoft runs the other way at 75.0% versus 95.5%. OpenAI, Anthropic and Google show no separation, and average provider count is flat at 3.8 versus 4.0.
Both “somewhat less regulation” responses come from the non-index cell, 9.1% versus 0.0%, which also means the Q5 base of two is entirely non-index. Directional support for more regulation is otherwise flat, 75.0% versus 68.2%.
The competitive-positioning lean is close, 50.0% versus 40.9%, but index respondents concentrate on “mostly about competitive positioning” (46.4% versus 18.2%) while non-index respondents take the absolute ends, 22.7% “entirely about competitive positioning” and 27.3% “mostly genuine safety measures.”
Net more confident is identical within noise, 75.0% versus 72.7%, but “much more confident” is 7.1% among index respondents against 27.3% non-index, with the index cell pooling at “somewhat more confident,” 67.9% versus 45.5%.
75.0% took at least one action against 54.5%, “no change” is 21.4% versus 36.4%, restricted agent permissions 42.9% versus 31.8%, and paused or slowed deployments 25.0% versus 13.6%. Prior familiarity with the incident is flat, 75.0% versus 72.7%.
In the coded open text, the governance-and-regulation-gap rationale appears in 28.6% of index responses against 9.1% non-index, and safety-and-guardrails in 50.0% versus 36.4%. Geopolitical competition and competitive market dynamics are flat across both cells.
Adoption pacing, willingness to deploy after independent evaluation, willingness to pay any premium and all four slowdown-scenario statements differ by less than the 8-point threshold this note treats as separation. The two respondents deliberately frozen on a model version are both non-index.
Which of the following AI model providers or platforms does your organization currently use in production or active evaluation?
Multi-model is the default posture in this sample, not a leading-edge one: 43 of 50 respondents run three or more providers. Microsoft's and Anthropic's leads are consistent with platform-mediated access, since Azure and Bedrock both present other labs' models. The 26.0% self-hosting open weights is the figure to hold against Q10, where 52.0% say a slowdown would raise their interest in open source.
The sample is concentrated by design: every respondent works at a large enterprise and 98.0% sit in North America.
Four in five respondents sit a level below the C-suite, in the VP through Director band where model selection and deployment gating actually happen. That composition matters for Q9: price authority is mixed across this sample, so the premium readings describe willingness rather than a committed budget line.
All 50 respondents qualified on three screens — very or extremely familiar with their organization's AI stack, involved in AI model decisions, and using at least one named model provider in production or active evaluation. The sample is therefore an informed-buyer sample, not a general enterprise sample, and it should be read that way.
N = 50 supports directional reading only. A 2.0-point difference is one respondent, and nothing here treats gaps under roughly 8 points as separation.
Q5 has a base of two and is reported for completeness; no conclusion rests on it. Chi-square tests on the subgroup cells reached conventional levels only for the Q8-by-Q9 relationship.
The sample is 98.0% North American and 100.0% large enterprise. The regulation and geopolitical readings in Q2 and Q4 would likely differ in EMEA or APAC samples and should not be generalized beyond this frame. Q12 and Q13 measure what organizations report having done, not what an audit would find. Q9 measures stated willingness to pay in a sample where 80.0% sit below the C-suite — treat it as demand signal, not committed budget.
Markets are moving faster by the minute. Futurum keeps intelligence moving with them through analyst-backed research, independent editorial, and partner-aligned media programs held to the same standard of objectivity.
