Baltra AI Server ASIC
- Details surfaced about Baltra, Apple's custom inference‑optimized server ASIC intended for private‑cloud AI workloads. - The chip reportedly uses Broadcom partnerships, a TSMC 3nm chiplet design, and targets mass production in H2 2026. - Apple hopes the ASIC will reduce dependence on NVIDIA for inference in controlled private clouds. (x.com)
Apple is reportedly building a custom server chip for artificial intelligence, a sign it wants more of its cloud AI work to run on Apple-designed hardware. (datacenterdynamics.com) The chip is internally called Baltra, and The Information reported in December 2024 that Apple was developing it with Broadcom and aiming for mass production in 2026. Reuters, in a pickup published the same day, said the project was intended to reduce Apple’s reliance on Nvidia processors. (finance.yahoo.com) A server chip is the processor inside a data-center machine, and an inference chip is tuned to answer requests from a model that has already been trained rather than build the model from scratch. Reports tied Baltra to that inference role, which fits Apple’s existing split between on-device features and larger cloud models. (apple.com) Apple already says some Apple Intelligence requests go to Private Cloud Compute, its system for sending harder tasks to server-based models running on Apple silicon. Apple’s security documentation says those servers are designed so user data is not stored and is not accessible to Apple. (apple.com) That makes the Baltra reports less of a jump into public cloud renting and more of an extension of infrastructure Apple has already described. Apple’s support pages say Private Cloud Compute handles complex requests with larger server-based models, while simpler tasks stay on the device. (apple.com) The manufacturing details are still reported, not confirmed by Apple. DatacenterDynamics, citing The Information, said the original plan pointed to Taiwan Semiconductor Manufacturing Co.’s N3P process, while newer April 2026 reports said TSMC’s N3E 3-nanometer process was under discussion instead. (datacenterdynamics.com; technode.com) Broadcom’s role also fits a wider pattern in the industry. Reuters reported on April 14, 2026 that Meta extended a custom-chip partnership with Broadcom for both training and inference accelerators through 2029, showing how large tech companies are leaning on Broadcom to build alternatives to standard Nvidia systems. (usnews.com) Apple has not publicly announced Baltra, named a launch customer, or published specifications such as memory, networking, or power draw. For now, the clearest read is that Apple’s cloud AI stack already runs on Apple silicon servers, and Baltra appears aimed at making that stack more specialized by 2026 or 2027 if the reported schedule holds. (apple.com; macworld.com)
