Apple is designing a chip carrying up to 1.5TB of unified memory, roughly double the capacity planned for this year's M5 Ultra, according to Bloomberg's Mark Gurman.
Rumour has it the M7 Ultra is due in 2028 and is being engineered to handle workloads closer to dedicated AI accelerators such as Nvidia's Blackwell than to a typical desktop processor.
It will underpin a server platform Apple is developing, targeted for roughly 2029, following an earlier M5 Ultra-based server generation expected to arrive first, per a report from MacRumors.
However, Apple has no track record selling compute infrastructure at scale.
Will it be competing in server hardware requires networking, cooling systems and an orchestration layer that Nvidia, AMD and the major cloud operators have spent a decade refining?
Apple's Neural Engine lineage originated in the abandoned Project Titan self-driving program and later migrated into consumer silicon, including the A11 Bionic chip in 2017.
Memory bottleneck
Whether Apple ships the full 1.5TB configuration depends on the state of the memory market, where a persistent supply shortage has driven component costs higher.
The same shortage is reportedly complicating Apple's broader component sourcing elsewhere in its supply chain.
At a common enterprise benchmark of roughly US$25 per gigabyte, a fully configured M7 Ultra machine would carry a price tag north of $35,000.
By comparison, the current top-tier M3 Ultra Mac Studio configuration tops out at 512GB, after Apple discontinued the 1TB option.
Strategy pivot?
Bloomberg reports that AI performance is now the primary factor shaping Apple's chip design decisions, rather than a secondary feature layered onto existing hardware.
Apple has traditionally built chips for general computing first and adapted them to new workloads afterwards, whereas the M7 family is reportedly being built around AI requirements from the outset.
Larger memory pools allow bigger AI models to run directly on the device rather than relying on cloud processing, which fits Apple's long-standing pitch that on-device inference protects user privacy and cuts latency.
Thing is, building a competitive enterprise AI offering involves far more than memory capacity, it also includes the software stack, developer tools and pricing that rivals have spent years establishing.
