Apple keeps cooking up its own chips because the company hates relying on anyone else, and the next project dropping is Baltra. This custom AI server chip should land sometime around 2027, with Apple and Broadcom teaming up on the design work using TSMC's 3nm manufacturing process.
The chip will probably handle inference workloads instead of training massive models from scratch. Apple already cut a deal with Google to use a customized Gemini model for cloud-based Apple Intelligence features, which costs them about a billion bucks annually. Baltra's architecture will focus on speed and throughput rather than the heavy lifting needed for model training, likely using lower-precision math operations.
Apple's silicon empire keeps growing beyond the usual A-series and M-series processors. The company recently started using its homegrown C1 modem chips, and rumors suggest a variant of the Apple Watch S-series chip might show up in upcoming AI smart glasses.
The chip will probably handle inference workloads instead of training massive models from scratch. Apple already cut a deal with Google to use a customized Gemini model for cloud-based Apple Intelligence features, which costs them about a billion bucks annually. Baltra's architecture will focus on speed and throughput rather than the heavy lifting needed for model training, likely using lower-precision math operations.
Apple's silicon empire keeps growing beyond the usual A-series and M-series processors. The company recently started using its homegrown C1 modem chips, and rumors suggest a variant of the Apple Watch S-series chip might show up in upcoming AI smart glasses.
