Organizations are considering a broader range of AI accelerators instead of limiting evaluations to Nvidia, but that does not mean they are preparing to leave their current platforms immediately. According to the July wave of the VB Pulse survey, which included 170 AI infrastructure participants, 39.4% said they were likely to evaluate non-Nvidia accelerators over the next 12 months, compared with 25.3% who said the same about Nvidia Blackwell (GB300) or next-generation Nvidia processors—a gap of 14 percentage points.
The alternatives named in the survey include AWS Trainium, Google TPU, AMD Instinct and Intel Gaudi, along with ASIC chips developed internally by organizations. Nevertheless, Nvidia remains the default choice in most production environments, making the result an indication of a broader evaluation pool rather than evidence of an immediate shift in production share.
More Activity, but Less Appetite for Immediate Change
The chip results fit within a broader trend: organizations are running AI infrastructure more intensively and trying to improve what they already have before deciding to change a major platform. The share of participants expecting to change platforms within zero to three months fell from 38.3% in June to 28.8% in July. Meanwhile, the share expecting a change within three to six months rose by 4.1 percentage points, and the share expecting one within six to 12 months rose by 5.3 points. The share planning no change remained roughly stable at 40%.
Several platforms recorded increases in production use. The share reporting Microsoft Azure use in production rose from 29% to 47.1%, while Google Gemini use rose from 41.1% to 47.6%. The share for OpenAI also increased from 40.2% to 49.4%, and Anthropic use rose from 12.1% to 24.7%. The Azure jump should be interpreted cautiously, because the July sample was more weighted toward large organizations: 57% of participants worked at companies with more than 1,000 employees, compared with 37% in June.
Organizations also improved their utilization of self-managed GPUs. The share operating these units at half capacity or less fell from 83% in June, in a sample of 100 participants, to 69% in July, in a sample of 155 participants. The share operating above 50% rose from 13% to 23%.
The Standard Is Shifting from Platform Size to Workload Performance
Selection criteria show that purchasing decisions are becoming more closely tied to actual operations. Integration with the cloud environment and enterprise data remained the most important factor, at 40.0% in July versus 41.1% in June. But performance priority rose from 24.3% to 35.3%, while the priority assigned to cost per million tokens increased from 7.5% to 15.9%, and the priority assigned to GPU access rose from 18.7% to 23.5%.
By contrast, the share choosing total cost of ownership as the main factor fell from 34.6% to 21.8%. This does not mean cost has lost importance; rather, evaluation appears to be more closely tied to the workload: how inference performs in production, how reliable it is, and what each unit of useful work costs.
Why Does This Trend Matter?
Interest in non-Nvidia alternatives rose from 31.8% in June to 39.4% in July. The increase was clearer among those with strategic purchasing authority, despite the small sample of senior executive positions: the share in this group rose from 42.9%, or six of 14, to 57.1%, or 12 of 21. Among final decision-makers, it increased from 35.4% to 50%.
These figures indicate that accelerator diversification is shifting from an engineering issue to a strategic infrastructure decision. Having alternatives on evaluation lists gives organizations greater ability to compare performance, cost and hardware availability, but it does not prove that these alternatives have achieved production superiority or that organizations have decided to move to them.
Control Extends Beyond Chips
The same trend appeared in the tool layer that connects models to enterprise data and services, which the survey described as including tools, data, orchestration, evaluation, identity, security and observability. In July, 36.6% preferred retaining independent best-of-breed tools alongside models, while another 36.6% preferred a mix of providers’ native runtime environments and independent tools. Thus, 79.2% supported an approach that retains some degree of architectural control outside a single model provider, compared with about 65.3% in June.
By contrast, the share preferring consolidation onto a single model provider’s native context stack fell from 20.8% to 11.9%. This control has practical importance: 68.3% of July participants said they had encountered a confident but incorrect answer caused by missing or incorrect context, compared with 57.4% in June. In addition, 62.4% said a governed semantic or context layer was in production or under construction.
Specialized Clouds and Open-Source Components
The share expecting greater use of specialized AI cloud providers, known as neoclouds, rose from 33% to 38%, while the share expecting less use fell from 9.7% to 5.4%. However, current production use among named providers such as CoreWeave, Lambda, Crusoe and Nebius remained much lower, rising from 1.9% to 5.9%.
The gap between expansion intent and current use points to potential evaluation and adoption, not completed adoption. Reliability, security, support, networking and data management will remain decisive factors in these companies’ ability to turn accelerator access into sustainable enterprise use.
Use of open-source and self-managed production packages also rose from 3.7% to 12.9%, with survey examples including PyTorch, Triton, vLLM, Ray and Kubernetes. Use of open-source key-value caching tools, such as LMCache and vLLM prefix caching, increased from 6.5% to 11.8%. But consideration of open-source platforms remained nearly flat, from 5.6% to 6.5%, meaning that expansion is concentrated among organizations that have entered the implementation phase rather than across a broader base of new evaluators.
Methodological limitations: The article compares two independent survey waves and two time periods, covering 107 participants in June 2026 and 170 in July 2026, rather than a longitudinal study of the same organizations. The sample composition also changed, and no statistical significance tests were conducted; therefore, the monthly differences should be treated as directional signals rather than proof of causality. The context-layer data is based on a separate survey of 101 participants in each month.