39.4% of enterprise AI infrastructure buyers plan to evaluate non-Nvidia accelerators in the next 12 months. Only 25.3% plan to evaluate Nvidia Blackwell or next-gen Nvidia GPUs. That 14-point gap, from a VentureBeat VB Pulse survey of 170 respondents published July 2025, is the headline number. The alternatives drawing attention: AWS Trainium, Google TPU, AMD Instinct, Intel Gaudi, and in-house ASICs. Among C-suite respondents, 57.1% favor non-Nvidia evaluation, up from 42.9% in June. Among final decision-makers, that figure rose from 35.4% to 50%. This is no longer an engineering conversation.
The same survey shows enterprises intensifying what they already operate rather than switching platforms. The share expecting a platform change within three months dropped from 38.3% to 28.8% between June and July, even as production adoption rose across every major platform. Azure jumped 18.1 points to 47.1% in production. Anthropic doubled from 12.1% to 24.7%. GPU utilization above 50% climbed from 13% to 23% among self-operated fleets. Uptime and reliability as effectiveness criteria rose from 42.1% to 51.2%. Cost per million tokens as a purchase priority doubled from 7.5% to 15.9%. Buyers are shifting from general infrastructure planning to workload-level scrutiny.
A companion survey on agentic context layers adds another layer worth reading in full. 68.3% of July respondents reported at least one confident-but-wrong agent answer caused by missing or incorrect context, up from 57.4% in June. In response, 79.2% of respondents favor architectures that preserve control outside a single model provider's stack, up from 65.3% in June. Only 11.9% want to consolidate onto a single provider's native context layer, down from 20.8%. The original report details how enterprises are governing that harness layer and which components they refuse to cede to any one vendor.
[READ ORIGINAL →]