Apple is designing a future M7 Ultra chip to support as much as 1.5TB of unified memory and to push local AI performance toward Nvidia Blackwell accelerator territory, according to Bloomberg reporter Mark Gurman. The catch is not subtle: the top configuration depends on whether the memory market stops being a tire fire by the time Apple wants to ship it.
Gurman reported that the M7 Ultra is expected in 2028, after a shortened Mac chip cycle that would move Apple quickly from M6 to M7. Apple has not confirmed the roadmap, the specifications, or the launch windows.
The memory figure is the part that jumps off the page. Gurman said the 1.5TB target is roughly double what Apple has planned for the M5 Ultra. That would be unified memory, meaning the CPU, GPU, and neural processing blocks pull from the same pool rather than shuttling data between separate system and graphics memory.
That model is one reason Apple’s high-end Macs can run large local workloads cleanly. It also means Apple has to buy a lot of expensive DRAM for the biggest configurations. Gurman tied the 1.5TB option directly to memory availability, which is a polite way of saying the spec sheet may meet the procurement department.
The reported roadmap
According to Gurman, Apple’s Mac silicon schedule now looks faster than previously expected. The company reportedly began taping out the M7 about six months after starting the same process for the M6. Tapeout is the point where a chip design is sent for manufacturing, so it is a meaningful milestone, though not a shipping date.
- Base M6: expected in fall 2026 for entry-level Macs, with no Pro, Max, or Ultra versions in that generation, according to Gurman.
- Base M7: expected in the first half of 2027, after the compressed design schedule.
- M7 Pro and M7 Max: expected near the end of 2027.
- M7 Ultra: expected in 2028, with up to 1.5TB of unified memory if supply allows.
- M8, code-named Soko: reported for the 2028 period on a 1.4nm-class process.
- Cardinal: described by Gurman as a high-end Mac chip for the 2028 generation.
The Ultra tier is where Apple usually stops pretending these are ordinary desktop parts. Apple’s current M3 Ultra reaches 819GB/s of memory bandwidth by combining two Max-class dies, and the Ultra chips carry the heaviest local AI and workstation loads in the Mac line.
Gurman did not report that M7 Ultra will match Nvidia’s Blackwell data-center GPUs. He wrote that the processor would “dramatically” improve AI performance and bring it closer to dedicated AI accelerators such as Nvidia’s Blackwell. That distinction matters. Blackwell is built for data centers; an Ultra-class Mac chip has different power, packaging, and product constraints.
Apple is also working on AI server hardware, according to the same Bloomberg report. Gurman said a server based on the M5 Ultra, code-named J246, is being prepared for deployment soon. A later server chip based on the M7 Ultra is planned for 2029, he reported.
The reported 2028 chips would move to a 1.4nm process, which lines up with TSMC’s A14 node schedule. TSMC has said A14 mass production is planned for the second half of 2028. That timing is tight enough that delays in either silicon manufacturing or DRAM supply could turn the most expensive M7 Ultra configurations into a roadmap footnote rather than a product customers can actually buy.
This story draws on original reporting from Tom's Hardware.
