Apple is developing data-center AI chips (NPUs) according to The Wall Street Journal. If the article is paywalled for you, don’t worry. The title contains all the content. The WSJ speculates that Apple’s Project ACDC is only for inference. We agree because the software infrastructure required for inference is less than that for training, and an inference-only chip can be cheaper to implement. The company may also find it advantageous to employ similar inference technology in its data centers as it does in its PCs and smartphones.
The WSJ also references Apple’s chip-design expertise relative to other AI giants. However, this ignores the expertise of Google’s TPU (NPU) team, much less that of Nvidia and rival chipmakers, and it ignores the service businesses of Broadcom, Marvell, and others that complete hyperscalers’ chip designs.
The big question is: why is Apple developing a data-center NPU? Google embarked on its TPU endeavors to create a better chip than it could buy and has continued developing new TPUs to align its data-center architecture, hardware, and software. Meta’s MTIA efforts target recommenders, a type of model that the company has expertise in and for which the company has a specific need. If Apple wants to lessen its dependence on Nvidia, it could source from other companies, including established suppliers AMD and Intel and startups such as Tenstorrent. The iPhone and Mac maker certainly has the resources and skills to develop a data-center NPU, but success also requires a clear vision.