Apple's M6 chip arrives on a two-nanometre process

On 26 August, Apple revealed its next generation of desktop and workstation silicon, and the two announcements point in different directions. The M6 is Apple’s first chip on a two-nanometre process, built for the mainstream of its lineup, with the fastest single-threaded performance the consumer industry has measured, according to the company. The M5 Ultra, announced alongside it, is the opposite kind of engineering: a quad-die design on the previous node, built for the heaviest work Apple sells, with up to a 36-core CPU, an 80-core GPU and inter-die memory bandwidth above four terabytes per second.

The two-nanometre transition is the part with industry-wide implications. Moving to a smaller process than the three-nanometre generation that has powered the last several years of chips is a manufacturing event, not just a product one, because it requires the foundry, TSMC, to have production ready at scale, and because the first customer on a new node sets the tone for everyone who follows. Apple’s M6 being built on 2nm means the supply chain has delivered, and it gives Apple a period of exclusivity on the node before the rest of the industry can follow, the same pattern that played out with earlier transitions.

The M5 Ultra is the more revealing product for where Apple’s silicon is going. A quad-die design, four chips on one package, is a construction Apple has not used before, and the target is explicit: running large language models locally and rendering complex 3D scenes, workloads that previously belonged to workstations with discrete graphics and cloud servers. The chip is a statement that Apple intends to serve the AI workstation market with its own silicon, at unified-memory sizes and bandwidth that make running substantial models on a desktop practical. That is a direct challenge to the assumption, widespread a year ago, that serious AI work had to happen in the cloud.

The timing of the announcement is as interesting as the silicon. Apple has been staggering its product releases, skipping the spring event and rolling out hardware through the year, and the M6 announcement arrived days after the new Mac mini and Mac Studio were refreshed on the M4 family. The pattern suggests a deliberate separation: the volume desktop line gets the proven node, the premium and workstation lines get the new one, and the company keeps both audiences buying in different quarters. It is a supply-chain management strategy as much as a product strategy.

The competitive read matters because the stakes have changed. For years the comparison for Apple silicon was with Intel and AMD on laptops; now it is also with the AI accelerators and workstations that professionals buy for machine learning. The M5 Ultra’s numbers, 80 GPU cores and multi-terabyte memory bandwidth, are aimed at that market, and the M6’s single-thread crown keeps the mainstream line unbeatable on everyday performance. Apple’s silicon programme, which began as a way to differentiate laptops, is now executing two strategies at once, one for the volume market and one for the most demanding professional work, and the 26 August event was the clearest view yet of both.

The honest caveat is that announcements are not benchmarks, and Apple’s performance claims have a habit of being measured against its own previous chips rather than the competition’s current ones. The M6 and M5 Ultra will be judged by the machines they ship in, by independent tests and by whether the AI workstation promise holds under real workloads. What is already clear is that the company that built its own chips to free itself from Intel’s roadmap is now driving the foundry roadmap itself, and that a two-nanometre, quad-die, multi-terabyte generation is what a self-designed silicon strategy looks like when it reaches full maturity.