Every Mac, priced

Converted at the ECB reference rate for 2026-08-25, then adjusted to strip consumption tax — most stores display tax-inclusive prices and the U.S., Canada and Brazil don't, so headline figures aren't comparable on their own.

The right-hand column is the honest comparison: what a gigabyte of memory actually costs once currency and consumption tax are removed. Canada, Singapore and Australia are all exactly 1.5× the U.S. number in their own currency, and the U.K. and Switzerland price at numeric par with the dollar.

Both ceilings are arithmetic, not benchmarks. Generation is bounded by memory bandwidth divided by bytes read per token, which for a mixture-of-experts model is its active parameters. Prefill is bounded by compute: roughly two floating-point operations per parameter per token, against an estimated —. Apple does not publish GPU throughput, so the compute figures scale from the measured ~70 TFLOPS of the 40-core M5 Max. Expect measured throughput around 30–50% of these numbers, and lower still for models with very small active-parameter counts, where routing and attention overhead dominate rather than weight reading.

Published reference points, for calibration against the ceilings above: on gpt-oss 20B an RTX 5090 measures around 205 tok/s of generation against roughly 50 for a DGX Spark, despite the Spark holding four times the memory. A Ryzen AI Max+ manages about 50 tok/s on a 27B. Bandwidth, not capacity, is what you feel per token.

The split is structural rather than technical. NVIDIA sells more bandwidth per card and far less memory per dollar, because generous VRAM on prosumer cards would undercut their datacenter business. Apple sells one large unified pool at moderate bandwidth. So Macs are good at fitting large models quietly, and NVIDIA is better at prompt processing and training, where compute rather than bandwidth is the binding constraint. Specifications and prices are approximate and current as of August 2026.

The M6 is the only chip whose bandwidth depends on how much memory you buy: 153GB/s at 16GB, 170GB/s at 24GB and above. It's the one upgrade in the line that buys two things at once.