Forrester predicts 25% of AI infrastructure RFPs will demand non-NVIDIA accelerators in 2027. NVIDIA will remain dominant for frontier training and large inference. Enterprises will segment workloads across accelerators to cut costs and reduce risks. Hyperscalers are already adding proprietary silicon like AWS Trainium and Google Ironwood TPU. 30% of large enterprises will build private AI infrastructure for inference. Rising token costs and complex agentic workflows will drive this demand. 20% of infrastructure investments will shift to industry-specific edge devices. Manufacturing, healthcare and retail will run AI at the edge for latency and resilience. Small models and domain-specific AI will move onto operational tech devices. Leaders must plan for more distributed, specialized and adaptable infrastructure.
What it unlocks
Segmenting AI workloads across different accelerator types based on cost and risk. Running sensitive or latency-critical inference on dedicated private infrastructure. Moving AI workloads to edge devices for faster, on-site decisions.
- 25% of AI infrastructure RFPs will require non-NVIDIA options
- 30% of large enterprises will build private AI infrastructure for inference
- 20% shift in investments to industry-specific edge devices
Sources