CUPERTINO, 26 AUG 2026 — Apple has announced the M6, its first chip built on a 2-nanometre process, and put it in a Mac mini starting at US$899. Pre-orders opened on 25 August in 30 countries and the machine ships on 22 September.
The M6 has a 12-core CPU and a 12-core GPU, up from ten and ten in the M5, and the first dual Neural Engine Apple has shipped: two 16-core engines where previous parts had one.
The performance claims are measured against the M4
Apple quotes up to 40 per cent faster processor performance, up to four times faster on AI tasks, and roughly twice the graphics and storage speed. Those figures deserve their baseline stated, because Apple states it and most of the coverage does not.
The comparison is against the previous-generation Mac mini, which carried the 10-core M4, configured with 32GB of unified memory and 2TB of storage.
That is a fair comparison for someone replacing a Mac mini, which is who the claim is aimed at. It is not a chip-over-chip figure. The specification comparisons in the same announcement are drawn against the M5 — twelve cores up from ten — so the announcement compares cores to one generation and speed to the one before it.
Apple is not being inaccurate, and the 40 per cent is still easy to misread as a single-generation M5 to M6 improvement. It is two generations of gain, and the announcement does not go out of its way to say so.
What the node actually buys
Moving from 3nm to 2nm does not make an individual transistor switch dramatically faster. It lets more of them fit in the same area and use less power for the same work. The real gains now come from the power saving.
For a small desktop that matters more than it would elsewhere. The Mac mini's enclosure sets a hard thermal ceiling, so performance in that box is bounded by what can be cooled rather than by what the silicon could theoretically sustain. In a small chassis, lower power per operation translates directly into higher sustained performance. That is not true of a machine with room to bolt on a larger heatsink.
The other half of the node story is who else is on it. Leading-edge capacity is scarce and expensive; Samsung has been raising prices on its advanced nodes with 4nm lines running full. Getting first-of-node volume requires committing to it years ahead, and very few buyers can.
The benchmark machine is not the US$899 machine
Read the comparison configuration again. The reference system carried 32GB of unified memory and 2TB of storage, and a like-for-like comparison means the new machine was similarly equipped.
The entry Mac mini is 16GB and 256GB. So the headline figures describe a configuration well above the advertised price, which is normal practice and still worth knowing. The storage claim in particular is a statement about a 2TB drive rather than about the one in the base machine.
This matters most for the buyers the AI framing is aimed at. Memory is the binding constraint on running a model locally — the weights have to fit — and 16GB shared between the operating system, applications and a model is tight for anything substantial. A 256GB drive holds surprisingly few model files once a system and a working set are on it.
The dual Neural Engine is in every unit. The memory and storage to make full use of it are the options, and configuring toward them moves the price a long way from US$899. This is less a criticism of the design than a caution. The entry price and the peak AI performance do not describe the same machine.
Two Neural Engines rather than one larger one
The interesting detail is the split. Apple shipped two 16-core engines rather than a single 32-core one, which is an architectural decision rather than a marketing one.
A single large accelerator is more efficient for one big model. Two separate engines are better when several things need to run at once without interfering — a system-level model handling dictation or image analysis while an application runs its own, each with predictable latency rather than queueing behind the other.
This points to where on-device AI is heading. A single large engine assumes one model runs at a time, on user command. Two engines assume something is always running in the background, and that a foreground task cannot be made to wait for it. Partitioning also makes isolation easier to reason about, which matters once models handle a user's own data.
It is the same reasoning visible in server silicon, where Nvidia's Vera statically partitions its cores' resources between threads rather than sharing them dynamically. Both are choosing predictable latency over peak throughput.
A first-of-node part in an US$899 box
Leading-edge silicon usually debuts where the margins are highest and the die is smallest, because early yields are poor and every wafer is precious. A desktop computer at US$899 is not the obvious place to introduce a process.
There are two ways to read this and they are not mutually exclusive. The first is confidence, since you do not ship first-of-node volume into a mainstream product unless yields support the price. The second is scheduling, since the Mac mini refresh was due and the part was ready.
Either way the market effect is the same. The first 2nm processor most people can buy is a desktop at under a thousand dollars, which sets a public reference point for what the node costs long before anyone publishes a wafer price.
What this does to everyone else's 2nm timetable
The date is the part with consequences beyond Apple.
Japan's Rapidus is targeting 2nm mass production in 2027, a target the government has just sought a further ¥150bn to support. That schedule already described reaching a node the leaders were shipping. With Apple putting 2nm in a consumer desktop in September 2026, the gap is no longer theoretical: a year, at least, before a challenger's first volume, by which point the leaders will be qualifying what comes next.
None of that makes a second or third source worthless. It does mean the pitch has to be supply security and price rather than parity, and that is a different business.
What it means from here
For buyers in the region the practical question is unglamorous. If the workload is ordinary desktop computing, the 40 per cent is against a two-year-old machine and the upgrade case rests on age rather than on the node.
If the workload involves running models locally — and for a growing number of small businesses handling data they would rather not send anywhere, it does — then the dual Neural Engine and the four-times AI claim are the relevant figures, and they are the ones aimed at that use.
The wider signal is about where inference is expected to happen. A company putting its first 2nm part into a US$899 desktop with two neural accelerators is not building for occasional AI features. It is building for a machine that is expected to be running something most of the time.
View author profile → · Editorial policy
