Cloud infrastructure operators are expanding deployments of custom-designed application-specific integrated circuits to run machine learning workloads. The move seeks to reduce reliance on third-party GPU suppliers and improve power-per-watt metrics in data centers.
Operating Costs and Vendor Diversification
Proprietary inference accelerators offer up to 40% lower cost-per-query compared to general-purpose GPUs. Hyperscalers are using these cost efficiencies to protect operating margins as enterprise adoption of generative AI features scales rapidly.