In the field of artificial intelligence, OpenAI has reached a major collaboration with Broadcom, marking the end of the era that relied solely on GPUs. The two companies have completed the design phase of a custom AI inference engine, which is expected to be deployed in data centers for the first time in the second half of 2026. This partnership plans to deploy up to 10 gigawatts (GW) of computing power over the next five years, which will completely transform the economic landscape of artificial intelligence.
This new chip is not a simple replication of existing hardware, but rather specifically designed for the "o1" series inference models and future versions of GPT. Unlike general-purpose chips such as NVIDIA's H100 or Blackwell, OpenAI's new chip uses a "system array" design, which is particularly optimized for dense matrix multiplication calculations in the Transformer architecture.
According to industry sources, this chip will be manufactured using TSMC's most advanced 3-nanometer process. At the same time, Broadcom will integrate its industry-leading Ethernet architecture and high-speed PCIe interconnect directly into the chip design to achieve up to 10 GW of computing power. Industry experts predict that this joint hardware and software design will reduce energy consumption per generated token by 30%.
This strategic shift has far-reaching implications for the technology industry. For a long time, NVIDIA has dominated the high-end AI chip market, but OpenAI's move toward custom chips provides a blueprint for other AI labs. In addition, OpenAI's vertical integration allows it to no longer rely entirely on external hardware suppliers, enhancing its competitiveness against large tech companies such as Google and Amazon.
At the core of this plan, OpenAI's CEO Sam Altman proposed the concept of "from transistor to token." This idea aims to view the entire AI process as a unified pipeline, and through controlling chip design, OpenAI can maximize token output per watt to meet the massive deployment needs of 10 GW.
As 2026 approaches, the main challenge for OpenAI and Broadcom is execution and manufacturing capabilities. Although the design has been completed, bottlenecks in advanced packaging technologies may affect the production schedule.
Key Points:
🌟 OpenAI and Broadcom have collaborated to launch a custom AI inference chip, planned for its first deployment in the second half of 2026.
⚡ This chip uses a 3-nanometer process and optimizes calculations in the Transformer architecture, significantly improving energy efficiency.
📈 OpenAI enhances its competitiveness through vertical integration, driving a transition in the AI industry from traditional computing to custom hardware.
