Apple Launches M6 and M5 Ultra Chips: First 2nm Desktop Silicon and Local 100B-Parameter LLM Inference

Apple unveiled two new chips — the M6 and M5 Ultra — alongside updated Mac Mini and Mac Studio lineups at its latest event. Two headlines stand out: the M6 is Apple's first chip built on a 2-nanometer process, and the M5 Ultra is the most powerful M-series processor Apple has ever made, with the top-spec Mac Studio capable of running 100-billion-parameter LLMs entirely on-device. All four new machines are available to pre-order today, shipping September 22nd, with the highest-end configurations already subject to purchase limits.

M6: First 2nm Desktop Chip, Built for Efficiency and Local AI

In an intriguing strategic move, Apple didn't save its first 2nm chip for the next iPhone — it went to the desktop lineup instead. That decision alone signals how seriously Apple is treating local AI computing: desktop platforms offer far more thermal and power headroom, allowing the benefits of a cutting-edge process node to be fully realized.

Built on 2nm, the M6 achieves a substantially higher transistor density than its predecessor. The CPU scales up to 12 cores (4 performance + 6 efficiency, with an optional super-core configuration) — two more than the M5 — and Apple claims the fastest single-core performance in the world, with multi-core performance 1.2x faster than M5. The 12-core GPU (also two more than M5) ships with a neural network accelerator per core, delivering nearly 30% more peak AI compute than M5 and more than 8x the performance of M1. The Neural Engine has been upgraded to a dual 16-core (32-core total) design, doubling the peak throughput of the previous generation.

The M6 supports up to 32GB of unified memory at 170 GB/s bandwidth — roughly 10% faster than M5. Compared to the M4-based Mac Mini, the M6 delivers 4x the AI performance, 2x the graphics performance, and roughly 40% faster CPU performance. That kind of generational leap speaks to just how much headroom the 2nm process opens up.

Mac Mini with M6: An Always-On Agent Computing Device

The Mac Mini retains its compact 12.7cm square footprint and comes in two configurations: M6 and M5 Pro. The M6 model starts at $899 with 16GB of memory and 256GB of storage — $100 more than the M4 version — while the M5 Pro model starts at $1,699, topping out at an 18-core CPU, 20-core GPU, 64GB of memory, 307 GB/s bandwidth, and Thunderbolt 5. Both models include Wi-Fi 7, Bluetooth, and 2.5GbE Ethernet, and ship with macOS 27.

Notably, Apple specifically positioned the Mac Mini as a device suited for use as an always-on agent computing machine — a clear nudge toward local AI inference server use cases, rather than simply everyday desktop work.

M5 Ultra: Apple's First Four-Die Fusion Processor

The M5 Ultra is Apple's first chip to use four-die fusion. Since the M5 Max is itself a dual-die package, Apple connected two M5 Max chips via the next-generation UltraFusion interconnect — four dies in total — with over 4.4 TB/s of inter-die bandwidth. The system treats the entire assembly as a single unified processor.

On the specs side: up to 36 CPU cores (12 super cores + 24 performance cores), up to 80 GPU cores each with an integrated neural network accelerator, up to 512GB of unified memory, and 1.2 TB/s of memory bandwidth — 50% more than M3 Ultra and 2x the M5 Max. Peak AI performance is more than 4x that of M3 Ultra.

The significance of 512GB of unified memory is hard to overstate: a single Mac Studio can load an entire 100-billion-parameter model directly into memory and run inference fully offline, with no network dependency and no per-token billing. For developers and enterprises that prioritize data privacy or want predictable inference costs, this is a genuinely compelling proposition.

Mac Studio Pricing and Cluster Capabilities

Mac Studio comes in M5 Max and M5 Ultra variants. The M5 Max starts at $2,499 with up to 40-core GPU and 128GB of memory; the M5 Ultra starts at $5,499 with 30-core CPU, 64-core GPU, 96GB of memory, and 1TB of storage. Both include Wi-Fi 7, Bluetooth 6, and Thunderbolt 5.

Multiple Mac Studio units can also be connected via Thunderbolt 5 to form a cluster for distributed AI inference, with Apple claiming up to 3x the throughput. This signals that Apple is actively building a scalable hardware story around on-premise large model deployment.

Industry Reaction and Three Key Signals

Alex Cheema, founder of EXO Labs — a project focused on running large models locally on Apple devices — received an M6 Mac Mini on launch day. Tim Cook's announcement post described the new Mac Mini as "small in size, big in performance, ready for everything from everyday tasks to all kinds of AI workloads." The M5 Ultra announcement was even more pointed: a two-unit purchase limit per customer, with a note that stock would sell out fast.

Taken together, this launch sends three clear signals:

First, a deliberate process strategy. Apple chose to debut mass-produced 2nm silicon on desktop hardware rather than the next iPhone — a strategic bet on desktop AI as a high-priority computing arena.

Second, the frontier of local AI is expanding. M5 Ultra delivers more than 4x the AI performance of M3 Ultra. Apple's own demo scenario featured a local model interfacing with a simulation tool, running entirely offline. High-end personal workstations are now genuinely capable of running frontier AI models.

Third, cost and supply constraints are real. Memory price increases are showing up directly in pricing and lead times — the 512GB M5 Ultra configuration has been pushed to a late October ship date due to memory costs, and the two-unit purchase limit reflects heavily concentrated demand against constrained supply.

Conclusion: The Local vs. Cloud AI Debate Enters a New Phase

From the M6 to the M5 Ultra, the logic behind this launch is clear: use 2nm process technology and massive unified memory as leverage to push local AI inference to the limits of what desktop hardware can do. When a single Mac Studio can run a 100-billion-parameter model offline, the cost and privacy calculus between local AI and cloud AI shifts meaningfully. At the same time, the pricing and availability pressure caused by rising memory costs is a reminder that the broader democratization of AI hardware still runs up against the hard realities of the underlying supply chain.