By admin 
The landscape of enterprise AI infrastructure is undergoing a significant shift, marked by a growing inclination to explore alternatives to Nvidia’s dominant position in the market. A recent VB Pulse survey of 170 AI infrastructure respondents reveals a notable trend: nearly 40% of enterprises are likely to evaluate non-Nvidia accelerators—such as AWS Trainium, Google TPU, AMD Instinct, Intel Gaudi, or in-house ASICs—within the next twelve months. This figure significantly outpaces the 25.3% of respondents who plan to consider Nvidia’s next-generation Blackwell (GB300) GPUs or other upcoming Nvidia offerings, indicating a 14-point gap favoring diversification. While Nvidia continues to be the de facto standard in most production environments, organizations are actively building strategic optionality, moving beyond a sole reliance on the industry titan.
This burgeoning interest in alternative hardware is embedded within a broader enterprise strategy focused on optimizing existing AI infrastructure before embarking on major platform overhauls. The survey data indicates that increased infrastructure activity has not translated into an urgent desire for platform changes. In fact, the proportion of respondents anticipating a platform switch within the next three months saw a decline from 38.3% in June to 28.8% in July. This dip occurred even as the adoption of AI in production environments, accelerator utilization rates, and exploration of neoclouds and open-source infrastructure all registered increases. This suggests a maturation of enterprise AI adoption, where the focus has shifted from rapid platform acquisition to maximizing the efficiency and effectiveness of current deployments.
The July data from the VB Pulse survey paints a picture of enterprises operating their AI infrastructure with greater intensity and bringing more provider platforms into production. Microsoft Azure demonstrated the most substantial growth in production adoption among major cloud platforms, with the percentage of respondents reporting its use in production soaring from 29% in June to 47.1% in July—an impressive 18.1 percentage-point increase. This surge can be partly attributed to shifts in the survey’s respondent pool, which in July skewed more towards larger organizations (57% from companies with over 1,000 employees, compared to 37% in June), where Azure adoption tends to be higher. Despite this, Google’s Gemini maintained its position as the most widely used platform across both survey waves, with its production adoption rising from 41.1% to 47.6% between June and July, narrowly surpassing Azure. OpenAI also saw a significant rise in production adoption, climbing from 40.2% to 49.4%, while Anthropic experienced even more dramatic growth, increasing from 12.1% to 24.7%.
Beyond cloud platforms, enterprises are also intensifying the utilization of their on-premises GPU infrastructure. The percentage of enterprises running their own GPUs at half capacity or less decreased from 83% in June (based on 100 respondents) to 69% in July (based on 155 respondents). Concurrently, the share of organizations reporting over 50% utilization rose from 13% to 23%. This improved efficiency reflects a growing emphasis on operational effectiveness. The importance placed on uptime and reliability as key measures of infrastructure effectiveness increased from 42.1% to 51.2%, while throughput saw a rise from 21.5% to 24.7%. Furthermore, the ease of implementation improved, with average ratings increasing from 3.84 to 4.04 on a five-point scale. Overall satisfaction saw a modest uptick from 4.07 to 4.14, and perceived value remained relatively stable around 3.9. This combination of factors suggests that enterprises are not necessarily experiencing a dramatic increase in perceived value solely due to deploying more infrastructure. Instead, they are becoming more adept operators with refined architectures, while simultaneously raising the bar for what that infrastructure must deliver, with reliability now taking center stage.
The declining urgency to make immediate platform changes is a significant counter-signal observed in the July findings. The proportion of respondents planning a platform change within zero to three months decreased by 9.5 percentage points. Conversely, the share anticipating a change within three to six months grew by 4.1 points, and the three-to-12-month window saw a 5.3-point increase. The segment with no planned changes remained relatively stable, hovering around 40%. This outward shift in urgency suggests that the ongoing discourse around open-weight models and open-source frameworks is influencing decisions, prompting enterprises to enhance existing components rather than undertaking wholesale platform replacements.
The evolving selection criteria for AI infrastructure further support this interpretation. Integration with existing cloud and data stacks remained the paramount factor in both survey waves, holding steady at 40.0% in July from 41.1% in June. However, the prioritization of performance saw a substantial increase, rising from 24.3% to 35.3%. Cost per million tokens also gained significant traction, jumping from 7.5% to 15.9%, and access to GPUs increased its importance from 18.7% to 23.5%. In contrast, the emphasis on broad total cost of ownership as a leading factor declined considerably, from 34.6% to 21.8%. This indicates a market progression from generalized infrastructure planning towards more granular, workload-level scrutiny. Buyers are increasingly focused on how platforms perform under production inference loads, their operational reliability, and the cost associated with each unit of valuable output.
The growing interest in Nvidia alternatives is particularly concentrated among strategic decision-makers and larger organizations. The 39.4% figure for evaluating non-Nvidia accelerators represents an increase from 31.8% in June, signaling a pre-existing momentum. This interest is even more pronounced among those with strategic purchasing authority. While 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. Similarly, among final decision-makers, this interest climbed from 35.4% to 50%. This trend is especially evident in small and medium-sized businesses. Among organizations with 251 to 1,000 employees, the proportion considering alternatives increased from 41.4% to 53.2%. For those with 101 to 250 employees, the rise was even more dramatic, from 33.3% to 57.7%. These findings underscore a strategic move by organizations to build optionality into their accelerator strategy, with increased attention from C-suite and final decision-makers indicating that accelerator diversity is evolving into a core strategic infrastructure question, rather than solely a technical concern for engineering teams.
The insights into AI infrastructure procurement align with findings from a separate VB Pulse survey on agentic context layers. This survey, which involved 101 substantive respondents in June and another 101 in July, explored the "AI harness"—the operational layer that connects AI models with enterprise data, tools, orchestration, evaluation, identity, security, observability, and business processes. In July, a significant 36.6% of context-layer respondents indicated plans to retain best-of-breed standalone tools alongside their chosen models. Another 36.6% expected to combine provider-native runtimes with standalone tools, while a mere 5.9% intended to build and own the context layer entirely in-house. Collectively, 79.2% of July respondents favored an approach that maintained architectural control outside of a single model provider, a notable increase from approximately 65.3% in June. Conversely, only 11.9% of July respondents preferred consolidating onto a single model provider’s native context stack, down from 20.8% in June. This widespread desire to preserve provider choice, independent governance, or control over critical components surrounding the AI model is becoming increasingly apparent.
The need for such control is further substantiated by the rising challenges associated with agent performance. In July, 62.4% of context-layer respondents reported that a governed semantic or context layer was either in production or under development, with production adoption alone increasing from 24.8% to 31.7%. Simultaneously, 68.3% of July respondents experienced at least one instance of a confident-but-wrong agent answer caused by missing or incorrect context, a significant rise from 57.4% in June. The share of respondents expecting to utilize multiple retrieval architectures based on specific use cases increased from 12.9% to 28.7%. The proportion anticipating a mix of provider-native and standalone context tools also grew substantially, from 20.8% to 36.6%. The emerging architectural paradigm points towards a controlled amalgamation of models, infrastructure, retrieval approaches, context systems, and operational tooling, meticulously selected to suit specific workloads.
The traction of neoclouds—specialized cloud providers focused heavily on AI infrastructure, particularly accelerator access and supporting services—is also on the rise. The July survey results suggest these providers are becoming a more credible component of enterprise multi-provider strategies. The share of respondents expecting to increase their engagement with neoclouds rose from 33% in June to 38% in July, while the proportion anticipating reduced engagement fell from 9.7% to 5.4%. This trend was particularly pronounced within the technology and software vertical, where 57.6% of July respondents expressed plans to do more with neoclouds, compared to 44.4% in June. While current production adoption remains modest compared to stated expansion intent, the production use of named providers like CoreWeave, Lambda, Crusoe, and Nebius increased from 1.9% of June respondents to 5.9% of July respondents. The substantial difference between the 38% expansion intent and the 5.9% current production use suggests a significant evaluation and adoption pipeline.
The demand pipeline for neoclouds is not merely theoretical. CoreWeave reported a revenue backlog of approximately $104 billion at the end of June, with an additional $25 billion in customer commitments secured in early Q3. Nebius, while not disclosing a directly comparable backlog metric, indicated that it could sell its entire 2027 capacity under current terms and reported four second-quarter AI cloud agreements, each averaging over $1 billion in total contract value. The broader implication is that neoclouds are emerging as a viable source of strategic leverage, offering enterprises additional options for accelerator availability, software stacks, workload placement, and valuable ammunition for negotiations with hyperscale cloud providers. However, neoclouds still face the critical task of demonstrating enterprise-grade reliability, security, support, networking, and data management capabilities. While specialized compute access may open the door, sustained enterprise adoption will hinge on the robustness of their surrounding operational stack.
The usage of open-source AI infrastructure is also expanding, particularly in production environments, though broad platform consideration remains relatively stable. The share of respondents reporting a custom, self-managed open-source production stack—encompassing technologies like PyTorch, Triton, vLLM, Ray, and Kubernetes—jumped from 3.7% in June to 12.9% in July. This movement was also evident in the adoption of open-source key-value cache tooling, such as LMCache and vLLM prefix caching, which increased from 6.5% to 11.8%. Within the technology and software segment, usage surged from virtually 0% to 13.3%. In contrast, open-source platform consideration saw only a slight uptick, moving from 5.6% to 6.5%, indicating that growth is concentrated among organizations actively implementing open-source solutions rather than a significantly broader base of evaluators. Open source appears to be deepening its penetration within an active segment of the market. Enterprises may be turning to open-source components for enhanced portability, greater model choice, and finer control over inference optimization. However, this increased ownership also transfers significant responsibility for managing upgrades, security, observability, integration, and production support.
The overarching theme emerging from these findings is a confluence of increased AI infrastructure activity, a tempered sense of urgency for immediate platform changes, and a deliberate pursuit of greater optionality. Enterprises are not only deploying more AI infrastructure but are also strategically keeping multiple avenues open across hardware (chips), cloud providers, and the critical layer that connects AI models to their proprietary data. The next wave of platform evolution, when it arrives, will likely be a carefully considered choice made from a position of strengthened strategic advantage and diversified technological capabilities.
Methodological Notes:
The analysis presented in this article is based on a comparison of two distinct, cross-sectional survey waves conducted in June and July 2026. The June survey involved 107 respondents, while the July survey encompassed 170 respondents. These waves are independent and do not constitute a longitudinal panel; therefore, the reported changes reflect shifts between different respondent populations rather than modifications within the same organizations over time. When analyzing platform-change timing, the sum of shares may slightly exceed 100% due to a small number of respondents in June (5) and July (9) selecting more than one time window.
Sample composition varied between the two survey waves. Respondents belonging to the 1-100 employee organization size category were excluded from the calculations to ensure a more focused comparison on larger enterprises. Despite this adjustment, the remaining wave compositions still differed, notably with a larger proportion of respondents from organizations exceeding 10,000 employees in the July survey. Consequently, month-to-month movements should be interpreted as directional indicators rather than definitive causal relationships. No statistical significance testing was applied to the comparisons presented herein.
The findings related to context layers are derived from a separate survey with 101 substantive respondents in June and 101 in July. This respondent base differs from the infrastructure survey, and these results are presented as supporting evidence rather than being integrated with the primary infrastructure survey data.
