Apple says the M6 is up to four times faster at AI. Every number in that promise was chosen by Apple.

Apple unveiled its first 2-nanometer chip on Tuesday, and within an hour the number everyone had was this one: up to four times faster AI. It is a good number. It is round, it is large, and it fits in a headline. Here are the three questions almost nobody asked before reposting it: four times faster at what, measured how, and compared to which chip? Answer those and the four shrinks — not to nothing, but to something a good deal more honest than a slide.

Let me say the fair part first, because I am not here to tell you the M6 is vapor. It plainly isn't. It is Apple's first chip on a 2-nanometer process, which is a genuine manufacturing step, not a marketing one. It adds two CPU cores and two GPU cores over the M5. It introduces the first dual Neural Engine — two 16-core engines instead of one. It lifts memory bandwidth to 170GB/s. Those are real changes to real silicon, and a generational bump is a reasonable thing to expect from them. My argument is not with the chip. It is with the numbers stapled to it, every one of which was produced by the company that would like to sell you the chip.

Compared to what?

Read Apple's own materials side by side and the baselines move under your feet. The headline "up to 4x faster AI," along with "up to 40 percent faster CPU" and "twice the graphics," is measured against the M4 — the chip in the outgoing Mac mini, two generations back. Turn to the page describing the M6 itself and the comparisons change partners: a "nearly 30 percent" increase in peak GPU compute for AI, this time against the M5; a Neural Engine with "up to 2x the peak compute over previous generations"; a CPU that is "up to 1.2x faster" multithreaded than the M5. Bandwidth is a "10 percent increase over M5" — unless you prefer the "2.5x over M1" framing, which is also on offer.

None of that is fraud. It is gravity. When you have several predecessors to choose from, each metric quietly picks the ancestor it beats by the widest margin. Against the M5, the honest generational story is roughly 20 to 30 percent on the parts that matter for AI — respectable, unremarkable, the sound of Moore's law still breathing. Against the four-year-old M1, everything is a multiple, because four years is four years. Against the M4, you get the four-times figure, because the Mac mini skipped a generation and the marketing gets to bank two years of gains as if they were one. The trick is not any single comparison. It is presenting them together, in the same breath, so the reader adds up numbers that share no baseline.

The missing number

Now the thing that isn't there. "Up to 4x faster AI" is a multiplier with no anchor. Four times faster than what, in absolute terms? Apple did not publish a figure — not a trillion operations per second, not a tokens-per-second on a named model, nothing you could hold. A multiplier without an absolute is a claim you cannot check, because you cannot even locate its starting point. It is the oldest move in the benchmark book: quote the ratio, withhold the base, and let the reader assume the base was high.

"Up to" is doing the rest of the lifting. "Up to" means best case — the single workload, at the single precision, under the single condition, where the new hardware pulls furthest ahead. Neural-engine throughput is almost always quoted at low numerical precision, because dropping from 16-bit to 8-bit or lower roughly doubles the operations per second on paper without the chip working any harder. Apple did not say which precision produced the four, or which model, or which task. It said "AI," a word that now covers everything from running a chatbot to sharpening a photo, and let you fill in the flattering version. A number that broad is not a measurement. It is a mood.

A multiplier is not a measurement. “Up to four times faster AI” tells you Apple ran two tests and divided one by the other. It does not tell you which two tests, at what precision, against which chip. — Phuong Nguyen

The ceiling is the memory, not the Neural Engine

Here is the part the AI framing quietly steps around. Whether you can run a real language model on your desk is decided less by how fast the Neural Engine is than by how much memory the machine has and how quickly it can move it. A model has to fit. A capable open model in 4-bit compression wants somewhere around sixteen to twenty gigabytes before you have typed a word, and every token of context eats more. The base Mac mini with the M6 ships with 16GB of memory and a 256GB drive, at a starting price that crept up to $899 — and it tops out at 32GB. You can make the Neural Engine as quick as you like; it cannot run a model that will not fit in the memory beside it.

Watch which machine Apple reaches for when it wants to say the impressive thing about local models. It is not the mini. It is the Mac Studio with the M5 Ultra, which scales to 512GB of unified memory and is the one Apple says can “run huge LLMs with hundreds of billions of parameters entirely on device.” That claim is true — of a machine with half a terabyte of memory and a four-figure price, not of the $899 box the four-times-faster number is attached to. The mini can run smaller models nicely, and for a lot of on-device tasks that is genuinely enough. But “run LLMs on device” is a sentence with an asterisk, and the asterisk is spelled in gigabytes.

What is actually new

So that you don't mistake skepticism for a verdict: several things here are real and worth having. Moving to a 2-nanometer process is the hard, expensive kind of progress, and Apple is first among its peers to ship it in a consumer machine. The dual Neural Engine is a genuine architectural change rather than a clock bump, and splitting the work across two engines is a sensible way to buy throughput. More cores and 10 percent more bandwidth are not nothing. If you are replacing a machine from the M1 or M2 era, the jump you feel will be large and the multipliers will, for once, roughly match your experience — not because the marketing was honest but because four years of hardware really did pass.

The point is narrower and it matters: a real chip does not need invented numbers, and the moment a company reaches for a baseline two generations back and a multiplier with no absolute, it is telling you the honest figure was less exciting than it wanted. The honest figure — around a quarter faster than last year on the AI-relevant parts — is fine. It sells a chip. It just doesn't trend.

The test that would settle it

None of this is checkable yet, which is its own small tell: the Mac mini with the M6 does not ship until September 22. Everything circulating now is a first-party claim about hardware no independent reviewer has run. The test that would settle it is boring and will arrive in about a month — the same open model, at the same precision, run on an M4, an M5 and an M6, tokens per second reported with the workload named, on machines someone other than Apple paid for. When those charts land, we will know what the four was really worth. My guess, on the record, is that it lands nearer to “about thirty percent faster than an M5 at the AI tasks most people actually run,” and that this will be a perfectly good chip described, for one week in August, by a number that was never a measurement. If the chart only goes up, look harder. And when every bar on it has a different baseline, look harder still.