Banner reading XPU dot pub

Using It to Make AI Chips, Softbank Would Risk Arm

ChatGPT, should Arm make AI chips?

Arm will establish a division to design and sell AI chips, according to Nikkei Asia. Dominant shareholder SoftBank Group is pushing the design licensor’s business in this direction as part of a larger AI vision. The first AI chips (NPUs) are to sample in one year, with production to begin by the end of 2025. If successful, the division could separate from Arm and operate under SoftBank directly.


  • Details are light. Even if Nikkei’s reporting is correct, companies consider all sorts of ventures, completing only a fraction. The SoftBank/Arm plan could change or terminate before completing the AI-chip project.
  • Expansion is many companies’ goal, but forward-integrating from licensing intellectual property (IP) to selling finished chips could repel Arm’s customers and depress its stock price, crippling SoftBank Group, which secures margin loans with its Arm shares.
  • Masayoshi Son helms SoftBank and aims to create a big AI company, investing billions according to Bloomberg. Billionaire Son supposedly operates on a 300-year plan but has also backed prima facie stupid businesses like WeWork.
  • Timing is bad for entering the AI-chip market but will get worse. Numerous companies have entered the market. Few have yet to do well, and those that survive will become more entrenched.
  • Arm might only be a launch vehicle for Son’s ambitions, a locus of chip-design skill he and SoftBank use to ignite an AI-semiconductor business. In this case, Arm doesn’t alter its business model.
  • Alternatives for SoftBank include resurrecting Graphcore, a failing data-center NPU company. (Graphcore calls its chips intelligence processing units, IPUs, further justifying the XPU.pub name for this site.) Bloomberg has confirmed earlier Financial Times reports that SoftBank has talked to Graphcore about an acquisition. The Graphcore entity and its remaining talent would be a better nucleus than Arm for a new NPU effort, particularly if the focus is data centers.

Bottom Line

SoftBank faces two questions:

  • Should it develop AI chips?
  • If so, should it involve Arm?

For edge AI, the variety of semiconductor products, supplier fragmentation, and Arm’s incumbency points to IP licensing’s superiority over chip development. Arm is already a competitive supplier, offering designs such as its Ethos-85. Many companies develop chips but lack the interest, expertise, and resources to forge a proprietary NPU and surrounding infrastructure, preferring to license the technology instead.

Data-center dynamics differ. The market is inhospitable to new suppliers, and key customers, such as Google and Meta, can roll their own NPUs if they want an alternative to Nvidia’s GPUs. Thus, a new AI chip supplier is ill conceived. Moreover, associating such a venture with Arm risks that company’s business just as it is breaking through in server and PC processors. AI may shape the next few centuries, but that doesn’t mean the world needs another chip company.




error: Unable to select