Apple announced the M6 on Monday and on paper it's a genuinely big deal: the first chip shipping anywhere on TSMC's 2nm process. What's odd is where it landed. Apple gave the 2nm debut to the Mac, not the iPhone, which is not how this usually goes — the iPhone eats first, always, because that's where the volume is. Reading between the lines, 2nm yields probably aren't ready for iPhone-scale production yet, and the Mac is a nice low-volume place to shake the node out.

The chip itself: 12 CPU cores split into two "super" cores, four performance and six efficiency. A 12-core GPU with a Neural Accelerator in every core. And the first dual Neural Engine Apple has shipped — two 16-core engines that the system frameworks will apparently drive at the same time. Bandwidth goes to 170GB/s, up from 153GB/s on the M5.

For the new Mac mini, Apple quotes up to 40% faster CPU, 2x graphics and 4x AI versus the M4 mini. Standard asterisk applies: Apple's benchmarks, Apple's choice of benchmarks, preproduction hardware. But even with the marketing discount applied, neural accelerators in every GPU core plus two NEs running in parallel is the kind of change that shows up in actual local inference and not just on a slide.

Then you scroll down to the memory row. 16GB standard, 32GB maximum.

The chip got better while the memory got scarce

Here's the part I find funny in a way I don't especially enjoy. The M6 mini starts at $899. The M4 mini launched at $599 in October 2024, got quietly bumped to $799 in June, and now sits at $899 for the same 16GB/256GB entry config. That's fifty percent more money in under two years for the entry-level box. Apple's stated reason for hiking prices across the Mac line is component costs — specifically RAM.

And they're not making that up. DRAM spot prices are up somewhere around 700% over the past year. Contract prices jumped more than 50% quarter-over-quarter coming into 2026. Data centres are now absorbing an estimated 70% of global memory production, because the memory makers moved wafer capacity off commodity DRAM and onto HBM, which sells for AI money. Dell, HP and Lenovo have all pushed PC prices up 15-20%. Phone vendors are quietly shrinking mid-range RAM configs. SK Hynix has suggested this doesn't fully unwind until sometime past 2030.

So: the box on your desk costs $300 more and stops at 32GB because of the AI buildout. And the main reason a lot of us want that box — running models locally, keeping data on our own hardware, not renting inference by the token — is a reaction to that same buildout. It's circular in a way that's hard to be cheerful about.

For local inference, memory is the whole game

Anyone who's actually run models locally knows you hit the memory wall long before you run out of compute. Weights have to fit, full stop. 32GB of unified memory is a genuinely useful machine — you can sit comfortably around the 20-30B range at 4-bit quantisation with room left for the OS and your editor — but it's a hard ceiling, and it's roughly the same ceiling this tier has had for a while now while the models kept getting bigger.

Which makes the M5 Pro mini at $1,699 a slightly strange pitch. It gets you more cores (15, configurable to 18) and 24GB base. If local inference is why you're buying, cores are not your bottleneck, and you're paying Pro money to move the wrong number. The genuinely large memory pools live up in Studio and Ultra territory, at prices that stopped being "small box on the desk" a long time ago.

None of which is to say the M6 is bad. It's a good chip, aimed with real precision at the workload people actually have in 2026, and shipping it on 2nm ahead of the iPhone is a surprising and kind of admirable call.

But "great silicon, constrained memory, higher price" is a strange thing to ship into the exact moment when the whole argument for on-device AI is getting off the meter. Meanwhile the hyperscalers are penciling in somewhere between $660B and $860B of capex for 2026, depending on whose estimate you trust — Amazon around $200B, Alphabet $175-185B, Microsoft $120B+, Meta $115-135B. That spending is the direct upstream cause of what you're now paying for RAM. Which means you're covering their data centres twice: once in the memory tax on your hardware, and again if you give up and just call the API.

Pre-orders are open now and it ships on September 22. And I'll probably still end up wanting one, which is the most irritating part of all this.