Apple’s M7 Ultra could bring server-class AI to Macs

Apple is reportedly designing its M7 Ultra chip to support up to 1.5TB of unified memory, a jump intended to make high-end Macs and future servers more capable of running large AI models locally. Bloomberg’s Mark Gurman says the chip would more than triple the 512GB ceiling of the current M3 Ultra and roughly double the M5 Ultra’s planned 768GB limit.

The roadmap signals a faster AI-driven shift in Apple Silicon. Apple is expected to ship only a base M6 later in 2026 before moving to the M7 family, with the M7 Ultra targeted for 2028. Gurman reports that Apple is also preparing an M5 Ultra-based server, codenamed J246, and a dedicated M7 Ultra AI server product aimed at a 2029 launch.

Large unified memory could let Apple hold more model weights in memory and reduce dependence on cloud infrastructure for certain workloads. Supply remains a major constraint, however, with high-bandwidth DRAM described as scarce and expensive, and Gurman warning that the top configuration may depend on market conditions.