Custom ASICs Gain Ground as AI Inference Race Accelerates

Artificial intelligence computing is shifting from training large models toward inference, the continuous process of serving responses to consumers and businesses. General-purpose GPUs remain flexible, but their cost, power consumption and latency can become constraints at massive scale. That is driving demand for custom application-specific integrated circuits, or ASICs, designed around narrower workloads and optimized through tighter coordination between chips, systems and software.

As of August 2026, competition for the inference market is intensifying. AMD acquired chip startup Taalas to expand its custom-computing capabilities, while Etched raised $700 million to accelerate development of chips purpose-built for Transformer models. Frontier AI laboratories including OpenAI and Anthropic are also advancing in-house silicon programs, seeking greater control over performance and costs while reducing their dependence on general-purpose GPUs.