Apple Ships M6 and M5 Ultra: Local AI Compute Goes Desktop

Both Mac models are available to pre-order now and will ship in late September, marking two distinct chips with very different strategies. The M6 is Apple's entry play; the M5 Ultra is the statement.

Start with the M6, because the headline belongs to it: the M6 is Apple's first based on a 2-nanometer production process. That matters for manufacturing pedigree—it validates TSMC's N2 node and signals that Apple expects to build on this for years. But the M6 itself stays modest. It's a competent refresh, not a revolution.

The M5 Ultra is where the real ambition lives. It means connecting two dual-die M5 Max chips to create a quad-die architecture. This is Apple's first quad-die configuration ever. And it scales significantly toward a GPU approaching eighty cores.

The base M5 Ultra configuration offers a substantial CPU, a mid-range GPU, and unified memory starting at thousands of dollars. Step up to the full configuration, and you're looking at five-figure pricing before memory upgrades. The maximum memory configuration reportedly lands substantially higher once fully specced.

The performance story, at least according to GPU benchmarks, is solid. The M5 Ultra has achieved meaningfully higher Metal scores compared to the M3 Ultra, and the chip's performance is roughly in line with NVIDIA's GeForce RTX 4080. Those numbers surfaced days before launch, offering the first evidence that Apple's claim to compete with discrete workstation GPUs isn't pure marketing. The RTX 4080 comparison matters because it's the kind of card AI developers actually rent cloud instances for.

But here's what Apple is really saying: stop renting. A fully configured M5 Ultra with maximum unified memory is a genuine alternative to leasing cloud GPU time for certain workloads—particularly large language model development and inference. Mac Studio scales up significantly in CPU and GPU cores and memory, enabling users to run enormous LLMs entirely on device.

The catch is memory pricing. Memory upgrades for the M5 Ultra are expensive—adding memory costs thousands, which is a substantial multiplier on the base price. That's real money, and it narrows the audience. You're not buying an M5 Ultra for general professional work—you're buying it because you need unified memory capacity and bandwidth that simply don't exist elsewhere in a single box at any price Apple will sell you.

The process node story cuts the other way. While the M6 gets the headline as Apple's first 2nm chip, the M5 Ultra—the flagship—remains on TSMC's third-generation 3-nanometer process. That's not outdated; it's pragmatic. Apple is reserving the 2nm advantage for the next generation of Ultra chips. For now, the M5 Ultra's power comes from architecture—the quad-die fusion, the per-core neural accelerators, the massive unified memory—not pure process geometry. The M6 in the Mac mini gets the manufacturing win to establish production maturity; the M5 Ultra uses the process it already knows.

Both ship in weeks. Pre-orders opened in late August, and they're due in stores by late September. This is real inventory, not roadmap speculation. That matters because Apple's bet on local AI compute only works if developers can actually hold the hardware.