Apple’s Desk-Scale AI Play: First 2nm M6 and a Quad-Die M5 Ultra Arrive Together

Apple broke its own silicon rules twice on August 25. The M6 became the company’s first chip built on a 2-nanometer process, debuting inside a refreshed Mac mini. Hours later, the M5 Ultra arrived in the Mac Studio as the first M-series design to bolt four dies into one package. One targets everyone; the other targets anyone who wants a small data center under their desk.

Both machines can be pre-ordered today and ship September 22, per Apple’s announcement and coverage by Unite.AI and Gigazine.

M6: the mainstream chip, rebuilt at 2nm

The headline claim from Cupertino: the M6’s CPU core is, in Apple’s own testing, the quickest single-thread performer on the market. The twelve-core layout behind that claim reads like a pyramid — two “super” cores up top for brute-force single threads, four mid-tier performers beneath them, six efficiency specialists handling background work. Apple credits that hierarchy with a 1.2x multithreaded gain over the M5.

Graphics got the same treatment. Each of the dozen GPU cores carries its own Neural Accelerator, which Apple credits for roughly a 30% jump in peak AI compute versus last year’s M5. A new twin-engine Neural Design doubles peak machine-learning throughput again, and system frameworks can drive both engines at once. Memory tops out at 32 GB, feeding off 170 GB/s of bandwidth — a modest ten-percent step up, but enough, Apple argues, to run language models locally for private agentic tasks.

M5 Ultra: four dies pretending to be one

The Studio’s new heart is stranger engineering. Two dual-die M5 Max packages are fused through an upgraded UltraFabric link pushing more than 4.4 TB/s between dies, with connection density up six-fold. The result reads to software as a single processor: as many as 36 CPU cores, 80 GPU cores — every one with a Neural Accelerator — plus a 32-core Neural Engine and dedicated media hardware including four ProRes engines.

Two numbers do the heavy lifting. Peak AI-oriented GPU compute is claimed at 4.5 times the M3 Ultra. And the memory ceiling reaches 512 GB at 1.2 TB/s — half again faster than its predecessor — which is what lets Apple state that models spanning several hundred billion parameters can live and run entirely on the desktop, no cloud round-trip required.

In the release, Sri Santhanam, Apple’s vice president of Silicon Engineering, framed the pair as “the next giant leap in performance and AI compute” and called the M5 Ultra “the most powerful M-series chip yet.”

Pricing starts at $899 for the mini and $2,499 for the Studio configuration; a leasing option called Apple Upgrade begins at $48.99 monthly over 36 months, according to Unite.AI. Buyers wanting a middle path can spec the mini with an M5 Pro and 64 GB of memory.

Read the fine print on those numbers

Every performance figure above comes from Apple measuring its own preproduction hardware in August 2026 against prior generations, using what the company describes as industry-standard benchmarks. Independent numbers will take longer. It is also worth noting Apple never names the foundry producing the 2nm node — the company has relied on TSMC for every prior generation, as Unite.AI observes.

There is a quieter story in the launch sequence, too. TNW reported back in June — correctly, it turns out — that Apple would skip conventional M6 Pro and Max variants this cycle and slot an M5-based Ultra into the Studio as a stopgap. That is precisely what shipped, suggesting the 2nm transition is being staged deliberately rather than rolled out all at once.

Analysis: a memory ceiling aimed at a shortage-shaped market

The specification lands at an awkward moment. TNW’s Ana Maria Constantin observes that an industry-wide DRAM shortage has already stripped shelves of high-memory Macs and pushed prices upward across the board — yet Apple is leaning its premium pitch on half a terabyte of unified memory per box.

The bet makes sense once you look at who buys it. For hospitals, law firms and government ministries, the interesting feature is jurisdictional rather than computational: prompts, files and customer records processed locally simply never leave the premises. Constantin’s caveat is fair — an American design fabricated in Taiwan settles where data sits, not who supplies the compute, so this is privacy by architecture rather than true sovereignty. Even so, a desk-sized machine that satisfies strict data-handling rules is arguably the most practical enterprise product Apple has shipped in years.

The strategic subtext: while cloud vendors fight for trillion-dollar data-center sites, Apple is selling the escape hatch from that queue. Watch whether the 512 GB tier sells through despite shortage pricing — it will tell you how big the local-inference market really is.

Sources

Benchmark figures are Apple’s own August 2026 measurements on preproduction units and remain unverified independently.