Enterprises favor non-Nvidia chips over Nvidia Blackwell GPU evaluation lists

The shift toward non-Nvidia accelerators suggests that the streaming and AI infrastructure market is moving away from a single-vendor monoculture toward a multi-provider strategy. By prioritizing AWS Trainium, Google TPUs, and AMD Instinct, organizations are gaining leverage against hyperscalers while focusing on specific metrics like cost per million tokens and uptime. This diversification allows streaming platforms to optimize inference costs and maintain architectural control through open-source harnesses rather than being locked into a proprietary stack. As neoclouds like Lambda and Crusoe scale, the industry will likely see a more fragmented but resilient compute ecosystem. Watch for whether Nvidia's Blackwell production yields can close this evaluation gap in the coming two quarters.

AWS Trainium and Google TPU have emerged as the leading alternatives in enterprise AI accelerator evaluations, driven by hyperscaler investment in custom silicon. In August 2025, AWS announced that Trainium3 chips would deliver up to 4x performance improvement over Trainium2 for large-scale inference workloads, with the company positioning the hardware as a cost-efficient path for organizations running generative AI at scale. Google has similarly expanded its TPU roadmap, with Google Cloud confirming in April 2025 that its sixth-generation Trillium TPU achieved a 67% improvement in compute performance per chip over TPU v5e, targeting both training and inference for large language models. These announcements align with the survey finding that enterprises are placing non-Nvidia accelerators higher on evaluation lists.

The competitive dynamics around Nvidia's Blackwell GPU have intensified as supply constraints and pricing pressure push buyers toward alternatives. In July 2025, AMD reported that its Instinct MI350 series had been adopted by multiple hyperscalers and AI labs for inference workloads, with the company citing growing demand from organizations seeking multi-vendor GPU strategies. Meanwhile, CoreWeave raised $1.5 billion in its March 2025 IPO, valuing the GPU cloud provider at over $23 billion, underscoring investor appetite for Nvidia-dependent infrastructure even as enterprises diversify. Neoclouds like Lambda and Crusoe have also expanded capacity, with Crusoe announcing a 1.2 GW data center campus in Abilene, Texas, in partnership with OpenAI to support next-generation model training.

Independent benchmarking efforts have begun to quantify the performance trade-offs between Nvidia and alternative accelerators. In June 2025, SemiAnalysis published a comparative analysis showing that AWS Trainium2 delivered 30-40% lower cost per token for LLM inference compared to Nvidia H100 clusters, though the report noted that software ecosystem maturity remains a barrier for teams accustomed to CUDA. Intel's Gaudi 3 has faced a different trajectory, with Intel confirming in May 2025 that it would deprioritize standalone AI accelerator sales in favor of integrating AI capabilities into its Xeon roadmap, effectively ceding the discrete accelerator market to Nvidia, AMD, and custom silicon from hyperscalers. For streaming platforms evaluating inference infrastructure, these benchmarks suggest that to unit economics and operational reliability rather than brand loyalty will drive procurement decisions.