Enterprises put non-Nvidia chips 14 points ahead of Nvidia’s next-gen GPUs on their evaluation lists

When enterprise buyers build out their next AI accelerator evaluation list this cycle, they’re more likely to put a non-Nvidia chip on it than Nvidia’s own next-generation GPU. According to VentureBeat’s July VB Pulse survey of 170 AI infrastructure respondents, 39.4% said they’re likely to evaluate non-Nvidia accelerators — AWS Trainium, Google TPU, AMD Instinct, Intel Gaudi or in-house ASICs — over the next 12 months, compared with 25.3% for Nvidia Blackwell (GB300) or other next-generation Nvidia GPUs, a 14-point gap.

Nvidia remains the default in most production environments. But organizations are building real optionality into their accelerator strategy rather than treating Nvidia as the only evaluation worth doing.

The finding sits inside a broader pattern: enterprises are expanding and optimizing the AI infrastructure they already operate before making another major platform change. Greater infrastructure activity did not produce greater urgency 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, even as production adoption, accelerator utilization, and exploration of neoclouds and open-source infrastructure all rose.

Where is enterprise AI infrastructure actually growing?

The July data shows organizations operating AI infrastructure more intensively and putting more provider platforms into production.

Microsoft Azure posted the largest production adoption growth among the major platforms measured, with the share of respondents reporting Azure in production increasing from 29% in June to 47.1% in July, an 18.1 percentage-point increase. Some of that jump reflects who was surveyed: July’s respondent base skewed more up-market than June’s (57% at organizations above 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 both waves, with the share of respondents reporting it in production rising from 41.1% in June to 47.6% in July, narrowly ahead of Azure.

The share of respondents reporting OpenAI in production rose from 40.2% to 49.4%. Anthropic production adoption increased from 12.1% to 24.7%. 

Among enterprises that operate their own GPUs, the share running at half capacity or less fell from 83% in June (100 respondents) to 69% in July (155 respondents), with the share above 50% utilization rising from 13% to 23%.

The definition of infrastructure effectiveness is also becoming more operational. The share of respondents who selected uptime and reliability as important effectiveness measures increased from 42.1% to 51.2%. The share selecting throughput rose from 21.5% to 24.7%.

Ease of implementation improved from an average rating of 3.84 to 4.04 on a five-point scale. Overall satisfaction moved only slightly, from 4.07 to 4.14, while perceived value was essentially unchanged at approximately 3.9.

That combination is telling. Enterprises are not reporting a dramatic improvement in value simply because they are deploying more infrastructure. They are becoming more capable operators with better architectures, but they are also setting a higher bar for what that infrastructure must deliver, with reliability leading the way.

Why is platform-change urgency shifting outward?

The strongest counter-signal in the July findings is the declining share of respondents who plan to make an immediate platform change.

The share expecting a change within zero to three months declined by 9.5 percentage points. The share expecting a change within three to six months rose by 4.1 points, while the six-to-12-month window rose by 5.3 points. The share with no planned change remained effectively flat at approximately 40%.

Urgency is shifting outward, with the open-weight-model and open-source-harness debate playing a role in which pieces get enhanced versus fully replaced.

The selection criteria support that interpretation. Integration with existing cloud and data stack was the top factor in both waves, holding steady at 41.1% in June and 40.0% in July. The share of respondents prioritizing performance increased from 24.3% to 35.3%. The share prioritizing cost per million tokens increased from 7.5% to 15.9%, while the share prioritizing access to GPUs rose from 18.7% to 23.5%.

By contrast, the share selecting broad total cost of ownership as a leading factor fell from 34.6% to 21.8%.

The market appears to be moving from general infrastructure planning toward workload-level scrutiny. Buyers increasingly want to know how a platform performs under production inference, how reliably it operates and what each unit of useful work costs.

Interest in Nvidia alternatives is concentrated at the top

That 39.4% figure was 31.8% in June, already climbing before this wave. The alternatives enterprises are weighing include AWS Trainium, Google TPU, AMD Instinct, Intel Gaudi and other in-house ASICs.

Interest was even stronger among respondents with strategic purchasing authority, though the C-suite sample is small: the share of C-suite respondents likely to evaluate non-Nvidia accelerators rose from 42.9% (6 of 14) in June to 57.1% (12 of 21) in July. Among final decision-makers, the same interest rose from 35.4% to 50%.

This was especially true for organizations in the small and medium-size business tiers. Among organizations with 251 to 1,000 employees, the share increased from 41.4% to 53.2%. Among organizations with 101 to 250 employees, it rose from 33.3% to 57.7%.

These findings show organizations building optionality into their accelerator strategy.

The increased attention from C-suite respondents and final decision-makers suggests that accelerator diversity is becoming a strategic infrastructure question, not just a technical one for engineering teams.

Enterprises want to own the harness

The infrastructure findings align with a separate VB Pulse survey of agentic context layers. That survey included 101 substantive respondents in June and 101 respondents in July.

The AI harness is the operational layer connecting models to enterprise data, tools, orchestration, evaluation, identity, security, observability and business processes. It determines what an agent can access, which actions it can take and how the organization evaluates its output.

In July, 36.6% of context-layer respondents said they planned to retain best-of-breed standalone tools alongside their models. Another 36.6% expected to mix provider-native runtimes with standalone tools, while only 5.9% intended to build and own the context layer in-house.

Combined, 79.2% of July respondents favored an approach that maintained at least some architectural control outside a single model provider, compared with approximately 65.3% in June. Only 11.9% of July respondents favored consolidating onto a single model provider’s native context stack, down from 20.8% in June.

Most want to preserve provider choice, independent governance or control over critical components around the model.

The need for that control is becoming clearer. In July, 62.4% of context-layer respondents reported that a governed semantic or context layer was either in production or being built. Production adoption alone increased from 24.8% to 31.7%.

At the same time, 68.3% of July respondents reported experiencing at least one confident-but-wrong agent answer caused by missing or incorrect context, compared with 57.4% of June respondents.

The share expecting to use multiple retrieval architectures by use case increased from 12.9% to 28.7%. The share expecting to mix provider-native and standalone context tools increased from 20.8% to 36.6%.

The emerging architecture is a controlled combination of models, infrastructure, retrieval approaches, context systems and operational tooling selected by workload.

Are neoclouds gaining enterprise traction?

Neoclouds are specialized cloud providers focused heavily on AI infrastructure, particularly access to accelerators and supporting services. The July results suggest that these providers are becoming a more credible part of enterprise multi-provider strategies.

The share of respondents expecting to do more with neoclouds increased from 33% in June to 38% in July. At the same time, the share expecting to do less with neoclouds fell from 9.7% to 5.4%.

The movement was especially pronounced among respondents in the technology and software vertical. The share of that July segment expecting to do more with neoclouds reached 57.6%, compared with 44.4% in June.

Current production adoption remains much smaller than broad expansion intent. Across the named providers measured consistently in both waves, such as CoreWeave, Lambda, Crusoe and Nebius, production use increased from 1.9% of June respondents to 5.9% of July respondents.

The difference between 38% expansion intent and 5.9% current named-provider production use may point to a sizable evaluation and adoption pipeline.

The neocloud demand pipeline is not theoretical. CoreWeave reported around $104 billion in revenue backlog at the end of June, excluding more than $25 billion in additional customer commitments secured during early Q3. Nebius does not disclose a directly comparable backlog metric, but said it could sell its entire 2027 capacity under current terms and reported four second-quarter AI cloud agreements, each averaging more than $1 billion in total contract value.

The larger implication is that neoclouds are becoming a viable source of strategic leverage. They give organizations additional options for accelerator availability, software stacks, workload placement and ammunition for negotiations with hyperscale providers.

Neoclouds will still have to demonstrate enterprise-grade reliability, security, support, networking, and data management capabilities. Specialized compute access may open the door, but durable enterprise adoption will depend on the surrounding operational stack. 

Is open-source AI infrastructure usage growing?

The most accurate answer is that open-source production usage is growing, while broad platform consideration remains relatively flat.

The share of respondents reporting a custom, self-managed open-source production stack increased from 3.7% in June to 12.9% in July. The stack definition included technologies such as PyTorch, Triton, vLLM, Ray and Kubernetes.

The movement was visible across several segments with July bases above 20 respondents:

• Among individual contributors, 23.9% reported production use in July.

• Among recommenders and influencers, 13.7% reported production use in July.

• Among organizations with 251 to 1,000 employees, 12.8% reported production use in July.

The share of respondents using open-source key-value cache tooling, including LMCache and vLLM prefix caching, increased from 6.5% to 11.8%. Among technology and software respondents, usage increased from effectively 0% to 13.3%.

Open-source platform consideration ticked up slightly but remained essentially unchanged, moving from 5.6% to 6.5%.

This combination suggests that growth is concentrated among organizations moving into implementation rather than across a dramatically larger population of evaluators. Open source appears to be deepening inside an active portion of the market.

Organizations may be turning to open-source components for greater portability, model choice and control over inference optimization. But ownership also transfers responsibility. Teams adopting self-managed stacks must operate upgrades, security, observability, integration and production support themselves.

That combination of more activity, less urgency and more optionality is the throughline across all of it. Enterprises are running more AI infrastructure while deliberately keeping multiple paths open on chips, clouds and the layer that connects models to their own data. The next platform change, when it comes, will be a choice made from a stronger position.

Notes on methodology

For this article, I compared two independent, cross-sectional infrastructure survey waves: 107 respondents in June 2026 and 170 respondents in July 2026. These waves are not a longitudinal panel, so the findings describe changes between respondent populations rather than changes made by the same organizations. Platform-change timing shares add to slightly more than 100% because a small number of respondents selected more than one window (5 in June, 9 in July).

Sample composition changed between the waves. Respondents selecting the 1–100 employee organization-size category were excluded before calculating results. The remaining wave composition still differed, including a larger July share from organizations with more than 10,000 employees. Month-to-month movements should therefore be treated as directional signals rather than proof of causation. No statistical-significance testing was applied to the comparisons reported here.

The context-layer findings come from a separate survey, with 101 substantive respondents in June and 101 in July. Those results use a different respondent base and are included as supporting evidence, not combined with the infrastructure-survey results.