OpenAI has made a landmark decision by finalizing its custom silicon roadmap in conjunction with Broadcom, marking the end of the “GPU-only” era for AI models. By late December 2025, the duo will unveil a specialized AI inference engine aimed at delivering 10 gigawatts (GW) of compute capacity over the next five years. This shift from general-purpose hardware to custom Application-Specific Integrated Circuits (ASICs) aims to alleviate soaring token costs and optimize power delivery. The engineered silicon will leverage TSMC’s 3-nanometer process and integrate advanced networking, promoting a scale-out capability for massive AI applications. OpenAI’s strategy creates a competitive edge against giants like Nvidia, Google, and Amazon, potentially generating $100 billion in revenue for Broadcom by 2029. With immense demands for efficiency, this move signals a transformative shift toward specialized AI infrastructure, driving forward the “transistors to tokens” philosophy essential for sustainable AI expansion.
Source link
