OpenAI and Broadcom Team Up to Develop Next-Gen AI Infrastructure for 2026 Deployment
OpenAI and Broadcom have announced a strategic collaboration to design custom AI inference chips, marking a significant shift from traditional reliance on GPUs in artificial intelligence computing.
The custom chip, optimised for OpenAI’s “o1” series models and future generative architectures, is expected to be deployed in data centres in the second half of 2026 and aims to scale up to 10 gigawatts (GW) of computing power over five years.
This initiative combines Broadcom’s Ethernet and high-speed interconnect technologies with advanced 3-nanometre fabrication processes to enhance performance and energy efficiency.
Industry sources estimate that the new design could reduce energy usage per generated token by approximately 30%, significantly improving the cost and environmental profile of large-scale AI workloads.
This partnership underscores a major evolution in AI infrastructure strategy: vertically integrated custom hardware tailored to specific large language model architectures.
While Nvidia GPUs have long dominated AI compute, this collaboration represents a bold bet on specialised silicon to lower costs and boost performance.
As generative AI applications proliferate across industries, the demand for compute power continues to rise sharply. A custom chip ecosystem could reduce dependence on general-purpose GPU makers, expand competition, and drive down operating expenses for hyperscale AI services.
OpenAI’s move parallels broader industry trends toward bespoke AI hardware, reflecting pressures from rising model sizes, cloud infrastructure costs, and global competition in AI chip development.
