The Information: Apple this past summer canceled the development of a high-performance Mac chip to let some engineers in Israel work on its first AI server chip
Wayne Ma and Qianer Liu, in a piece today for The Information (paywalled up the wazoo, sadly), “Apple Is Working on AI Chip With Broadcom”:
Context & Ripple Effects
Apple’s reported server-silicon work had already surfaced as Project ACDC, a chip effort for data-center AI software, followed by reporting that it was being developed with Broadcom. The reassignment indicates that the effort was not merely exploratory: it was receiving engineers previously allocated to high-performance Mac silicon.
The move also sharpens Apple’s emerging division of labor between device-side AI and data-center capacity, after reports that [[a:864933|M2 Ultra would handle the first server-side workloads while simpler tasks ran on Apple devices]].
First-order effects
- Apple trades progress on a high-performance Mac-chip program for additional engineering capacity on its first AI server chip.
- The Israel-based engineers become part of a higher-priority server-silicon effort, while the affected Mac roadmap loses the resources attached to the canceled project.
Second-order effects
- A server chip developed with Broadcom becomes more central to Apple’s AI rollout, increasing the importance of execution across Apple’s silicon and infrastructure partners.
- The decision reinforces a split in Apple’s compute plans: on-device chips retain simpler AI tasks, while heavier workloads require dedicated server hardware.
Third-order effects
- If Apple continues redirecting client-chip talent to server silicon, AI infrastructure could become a lasting constraint on how quickly its product teams can introduce more capable AI features.
- The broader model is an integrated AI stack in which product differentiation depends on coordinating device chips with proprietary server capacity, rather than treating cloud compute as a separate utility.
The trend: Apple is reallocating custom-silicon talent from traditional device performance toward a heterogeneous AI compute stack spanning local devices and dedicated servers.