Analysis
A VentureBeat survey of 170 AI infrastructure buyers found that 39.4% plan to evaluate non-Nvidia accelerators over the next 12 months, compared with 25.3% who plan to evaluate Nvidia's next-generation Blackwell-class GPUs -- a 14-percentage-point gap in stated evaluation intent, VentureBeat reported, based on its VB Pulse survey series.
What the Numbers Actually Measure
This is evaluation intent, not deployed spend or market share -- the survey asks buyers what they plan to test over the next year, not what they've already switched to. The non-Nvidia field named specifically includes AWS Trainium, Google TPU, AMD Instinct, Intel Gaudi and in-house custom ASICs. Nvidia remains, by a wide margin, the dominant chip in actual production AI workloads today; what's shifting is the share of enterprises willing to spend evaluation budget checking whether an alternative works well enough to reduce dependence on a single, currently supply-constrained vendor.
“The non-Nvidia field named specifically includes AWS Trainium, Google TPU, AMD Instinct, Intel Gaudi and in-house custom ASICs.”
The survey methodology matters here: VentureBeat's July wave had 170 respondents versus 107 in June, and it's a repeated cross-sectional design rather than a panel following the same companies over time -- meaning some of the month-over-month shift could reflect a different respondent mix rather than the same buyers changing their minds. That's a real limitation on how confidently this gets read as a trend rather than a snapshot.
Why Integration, Not Performance, Is Now the Top Filter
The more durable finding in the data may be the shift in selection criteria rather than the vendor gap itself: integration with a buyer's existing cloud and data stack is now the top factor at roughly 40%, ahead of raw performance (35.3%) and cost per million tokens (15.9%). That ordering makes chip choice largely a downstream decision of cloud commitment -- an enterprise already running on AWS has an easier evaluation path to Trainium than to a chip requiring new infrastructure, regardless of relative benchmark performance. That's a structural advantage for the hyperscalers' own silicon (AWS Trainium, Google TPU) over third-party alternatives like AMD and Intel, which have to win on standalone merit without an existing-stack tailwind.
The Tension With What Labs Are Actually Doing
This survey data sits in real tension with what the largest AI labs are doing with their own capital. The same week this survey ran, Anthropic signed roughly $80 billion in new compute commitments with Nvidia-backed Lambda and Nscale, and Nvidia agreed to acquire Hugging Face for $12.9 billion -- both moves that deepen rather than reduce Nvidia's position at the center of frontier AI infrastructure. The gap between what survey respondents say they're evaluating and what the largest, most capital-intensive buyers are actually committing to suggests two different markets: frontier labs training the largest models remain overwhelmingly Nvidia-dependent because CUDA's software ecosystem has no substitute at that scale, while a broader base of enterprise AI buyers running smaller, more standardized inference workloads have genuine, evaluable alternatives.
What determines whether the 14-point gap becomes actual market share movement: whether AMD, Google and AWS can close the software ecosystem gap CUDA still holds, and whether enough of that 39.4% evaluating alternatives actually completes a production migration rather than running a pilot and staying on Nvidia anyway.
