According to market news, OpenAI is cooperating with Arm, a subsidiary of SoftBank, to develop a new CPU to match the customized AI accelerator jointly developed with Broadcom.
Foreign media "The Information" reported that Arm will design a server-grade CPU as the core of OpenAI's next-generation AI rack. This may be one of Arm's biggest steps into the data center market so far.
This chip is an AI accelerator developed by OpenAI. The company announced on the 13th that it will cooperate with Broadcom to deploy customized AI accelerators and rack systems. Targeted specifically at inference workloads, the SoC is expected to enter production in the second half of 2026 and scale to approximately 10 gigawatts of computing power between 2026 and 2029. Among them, Broadcom’s accelerator will be manufactured by TSMC and is said to have been developed for about 18 months.
According to reports, Arm’s new role is not only to provide architectural blueprints. The company has recently begun to design and produce CPUs on its own, rather than just licensing core IP to partners. According to people familiar with the matter, OpenAI may not only use Arm-designed CPUs on Broadcom chips, but may also be used with systems from NVIDIA and AMD.
Through cooperation with OpenAI, the potential revenue of this CPU project is expected to reach billions of dollars, which will be a major gain for SoftBank, which holds nearly 90% of Arm’s shares and uses large amounts of debt to borrow equity. At the same time, SoftBank has also committed to investing tens of billions of dollars in the construction of OpenAI data centers and purchasing its AI technology to accelerate Arm's own chip development cycle.
OpenAI stated that, combined with previous agreements with NVIDIA and AMD, its chip project currently has a total data center capacity of up to 26GW. Analysts expect that if the development is successful, OpenAI's custom chip deployment combined with purchases from NVIDIA and AMD could bring the company's total construction costs to more than $1 trillion.
Arm developing custom CPU for OpenAI’s in-house accelerator — core IP would underpin 10GW of installed AI capacity Further reading: 2025 OCP Summit》Broadcom and Celestica join hands to promote 1.6T network technology to solve the million XPU challenge 2025 OCP Summit》Ethernet is the key technology for Scale-Up! Broadcom launches industry's first 800G AI Ethernet NIC