Alphabet, Amazon and Meta are ramping up development of proprietary application-specific integrated circuits to handle key AI workloads, signaling a move away from reliance on Nvidia GPUs for both training and inference tasks. This shift could transform hardware procurement strategies across major cloud providers.

Cloud providers are prioritizing energy efficiency metrics such as joules per token to manage escalating power costs as generative AI demand grows. Custom chips promise higher throughput and lower operating expenses compared with general-purpose GPUs by optimizing power consumption per output unit.

As enterprise AI workloads transition from model training to inference, demand is tilting toward processors tailored for low latency and high efficiency. Alphabet's specialized ASICs could gain an advantage in applications ranging from customer-facing chatbots to robotics and other physical AI systems.