VB Pulse survey: enterprises put non-Nvidia chips 14 points ahead of Nvidia's next GPU

A VB Pulse survey conducted in July among 170 AI infrastructure leaders reveals a figure that sums up the current moment in the accelerator market: 39.4% of respondents say they are likely to evaluate chips not made by Nvidia —AWS Trainium, Google TPU, AMD Instinct, Intel Gaudi or in-house ASICs— over the next twelve months, versus only 25.3% who will evaluate Nvidia's next-generation GPUs (Blackwell/GB300 or others). That is a 14-point gap in favor of the alternatives. Nvidia remains the default standard in most production environments, but organizations are building real optionality into their accelerator strategy rather than limiting themselves to evaluating Nvidia products alone.

That interest in alternatives was already growing: in the June wave it stood at 31.8%. And it is not spread evenly, but is concentrated especially at the highest decision-making levels, though the sample in those segments is small. Among C-suite respondents, willingness to evaluate non-Nvidia chips went from 42.9% (6 of 14) in June to 57.1% (12 of 21) in July. Among those with final purchasing authority, it rose from 35.4% to 50%. The phenomenon is also pronounced at midsize companies: at firms with 251 to 1,000 employees it went from 41.4% to 53.2%, and at those with 101 to 250 employees, from 33.3% to 57.7%. That interest is growing so sharply among executives and final decision-makers suggests accelerator diversity is becoming a strategic business question, not just a technical one for engineering teams.

Paradoxically, this appetite for exploring alternatives does not translate into any rush to switch platforms. The share of respondents expecting a platform change within three months fell from 38.3% in June to 28.8% in July, while production adoption, accelerator utilization and exploration of neoclouds and open-source infrastructure all rose simultaneously. Put another way: there is more infrastructure activity, but the urgency to replace platforms is shifting into more distant time windows. The "change in 0-3 months" band dropped 9.5 points, while "3-6 months" rose 4.1 points and "6-12 months" rose 5.3 points; the share with no change planned held steady at around 40%.

As for where enterprise AI infrastructure is actually growing, Microsoft Azure posted the biggest jump in production adoption among the major platforms: from 29% in June to 47.1% in July, an increase of 18.1 percentage points. The article itself cautions that part of that jump reflects a change in sample composition, more skewed toward large companies in July (57% of organizations with more than 1,000 employees, versus 37% in June), and Azure adoption rises with company size in both waves. Google's Gemini was the most-used platform in production in both waves, going from 41.1% to 47.6%, slightly ahead of Azure. OpenAI in production rose from 40.2% to 49.4%, and Anthropic practically doubled, from 12.1% to 24.7%.

Among companies running their own GPUs, underutilization is improving: the share operating at 50% capacity or less fell from 83% (100 respondents) in June to 69% (155 respondents) in July, while the share above 50% utilization rose from 13% to 23%. What counts as "effective infrastructure" is also changing: the importance placed on availability and reliability (uptime) rose from 42.1% to 51.2%, and on throughput from 21.5% to 24.7%. Ease of implementation improved from an average score of 3.84 to 4.04 out of five, while overall satisfaction barely moved (from 4.07 to 4.14) and perceived value stayed essentially flat at around 3.9. In other words: companies are running more infrastructure with better architectures, but they do not perceive a dramatic improvement in value, because they are simultaneously raising the bar on what they demand, with reliability as the leading criterion.

Infrastructure selection criteria are also shifting toward production performance and real cost, and away from generic total cost of ownership (TCO) assessments. Integration with existing cloud and data stacks remained the top factor in both waves (41.1% in June, 40.0% in July). Performance as a priority criterion rose from 24.3% to 35.3%, cost per million tokens from 7.5% to 15.9%, and GPU access from 18.7% to 23.5%. By contrast, the share prioritizing broad TCO fell from 34.6% to 21.8%. The market appears to be moving from general infrastructure planning to scrutiny at the level of the specific workload.

A parallel VB Pulse survey on "agentic context layers," with 101 substantive responses in June and another 101 in July, reinforces this same logic of distributed control. This layer —the "harness"— is the operational level that connects models with enterprise data, tools, orchestration, evaluation, identity, security, observability and business processes, and it determines what an agent can access and how its outputs are evaluated. In July, 36.6% of respondents planned to keep independent "best-of-breed" tools alongside their models, and another 36.6% expected to combine vendor-native runtimes with independent tools; only 5.9% intended to build and own the context layer in-house. All told, 79.2% of July respondents favored an approach that kept at least some architectural control outside a single model provider, versus roughly 65.3% in June. Only 11.9% preferred consolidating everything into a single vendor's native context stack, down from 20.8% in June.

The need for that control becomes clearer in the reliability data: 62.4% of July respondents said they already had a governed semantic or context layer in production or in development (production adoption alone rose from 24.8% to 31.7%), while 68.3% of July respondents reported having experienced at least one "confident but incorrect" agent response due to missing or wrong context, versus 57.4% in June. The share expecting to use multiple retrieval architectures depending on the use case rose from 12.9% to 28.7%, and the share expecting to mix native and external context tools rose from 20.8% to 36.6%.

Neoclouds —cloud providers specialized in AI infrastructure, above all in accelerator access— are also gaining traction as part of multi-vendor strategies. The share expecting to make greater use of them rose from 33% to 38%, while the share expecting to use them less fell from 9.7% to 5.4%; the move was especially pronounced in the technology and software sector, where that figure reached 57.6% in July versus 44.4% in June. However, actual production adoption remains far below that expansion intent: among providers consistently named in both waves —CoreWeave, Lambda, Crusoe and Nebius— production use rose from 1.9% to 5.9%. That distance between 38% expansion intent and 5.9% actual production use suggests a considerable evaluation and adoption pipeline. The article provides context figures on demand: CoreWeave reported around $104 billion in revenue backlog at the end of June, not counting more than $25 billion in additional customer commitments secured early in the third quarter; Nebius, although it does not disclose a directly comparable backlog metric, said it could sell out all of its 2027 capacity on current terms and reported four AI cloud deals in the second quarter, each with a total contract value averaging more than $1 billion.

As for open-source infrastructure, production use is growing while overall consideration of these platforms remains essentially flat. The share of respondents reporting a self-managed, custom open-source production stack (with technologies such as PyTorch, Triton, vLLM, Ray and Kubernetes) rose from 3.7% in June to 12.9% in July. That growth was visible across several segments with July bases above 20 respondents: 23.9% among individual contributors, 13.7% among "recommenders and influencers," and 12.8% at organizations with 251 to 1,000 employees. Use of open-source key-value caching tools (such as LMCache and vLLM's prefix caching) rose from 6.5% to 11.8%, with a jump from practically 0% to 13.3% among technology and software sector respondents. By contrast, overall consideration of open-source platforms barely moved, from 5.6% to 6.5%, suggesting the growth is concentrated in organizations already moving to implementation rather than in a much larger population of new evaluators. Adopting self-managed stacks also shifts responsibility for updates, security, observability, integration and support onto the teams themselves.

The common thread running through all this data is a combination of more activity, less urgency and more optionality: companies are running ever more AI infrastructure while deliberately keeping several paths open across chips, clouds and the layer that connects models to their own data. According to the article, the next major platform shift, when it comes, will be made from a stronger negotiating position.

It is worth noting, as the article itself specifies in its methodology note, that these are two independent cross-sectional survey waves —107 respondents in June 2026 and 170 in July 2026— and not a longitudinal panel, so the findings describe changes between different respondent populations, not changes within the same organizations. The platform-change timeline percentages add up to slightly more than 100% because a minority of respondents marked more than one time window. Sample composition also changed between waves —organizations with 1 to 100 employees were excluded and July had a higher proportion of companies with more than 10,000 employees— so month-to-month movements should be read as directional signals and not as proof of causality, since no statistical significance testing was applied. The agentic context layer data come from a different survey, with its own respondent base, and are presented as complementary evidence, not combined with the infrastructure survey results.