What a Mac Mini and a Mac Studio Actually Run

This is the first entry in a series we're calling Processing Gear: real specs, real prices, and what a given machine can actually do, checked against the marketing. Apple's own component costs moved this year, which changes both what a machine costs and, less obviously, what it's capable of. We're starting with the two Apple Silicon machines a small business is most likely to actually buy.

Why the price moved

Apple raised prices across the Mac and iPad lineup in June 2026. The company's own reasoning, on the record: CEO Tim Cook called the increases “unavoidable because of memory chip shortages,” driven by “the massive number of AI datacenters being built.” Memory makers have been redirecting production capacity toward the high-bandwidth memory that AI data centers buy, leaving less of the ordinary DRAM that goes into a laptop or a desktop. Cook told the Wall Street Journal, “we've been trying to shield our customers from the increases, but the situation has become unsustainable.” Apple's shares fell 6.1% the day the increases were announced, their worst day in more than a year. This isn't a one-time bump tied to a single product cycle. It's the same AI boom that makes this hardware worth buying in the first place, showing up on the cost side too.

What the two machines actually cost and hold, right now

Checked directly against Apple's own current storefront, not a reseller or a rumor site:

Two corrections worth stating plainly, because both circulate as received wisdom in hardware-advice writeups. First: 48GB is the real ceiling for a Mac Mini right now. There is no 64GB Mac Mini configuration on Apple's storefront. Second, and this one we got wrong ourselves in an earlier pass on this research: 96GB unified memory is not an M4 Max configuration at all. The M4 Max Mac Studio tops out at 64GB. Getting to 96GB on Apple's live configurator forces a chip swap to M3 Ultra. The interface literally presents it as “your selection requires a change to your configuration” and upgrades the whole machine, not just the memory. And 96GB turns out to be the ceiling for the entire Mac Studio line, M4 Max or M3 Ultra: Apple's own spec sheet lists no “configurable to” option above it for either chip, and neither does the live configurator, at either M3 Ultra core-count tier.

What that memory actually buys you, in plain arithmetic

“How many billion parameters can it run” is a real, answerable question, but only once you fix the quantization level, because that's what actually determines how many bytes each parameter takes up in memory. Cut the precision and the model gets smaller and a little less accurate; that trade is normal practice, not a corner cut. The arithmetic is simple enough to check yourself:

• 4-bit quantization (the common choice for running a model locally): roughly 0.5GB per billion parameters. A 70B model needs about 35GB just for the weights.
• 8-bit quantization (higher fidelity, double the footprint): roughly 1GB per billion parameters. A 70B model needs about 70GB.
• Full precision (FP16): roughly 2GB per billion parameters. A 70B model needs about 140GB, before accounting for the context window or anything else running on the machine.

That's weights only. Running the model also needs headroom for the context window (the conversation it's holding in memory) and the operating system, so treat these as floors, not comfortable operating points. On a 48GB Mac Mini at 4-bit quantization, an 8B model runs with plenty of room to spare; a 70B model's 35GB of weights leaves very little headroom for anything else, and a 120B-class model's 60GB of weights doesn't fit at all, full stop, with or without headroom. On a 96GB Mac Studio, that same 120B model at 4-bit needs about 60GB of the 96GB available, which is workable with real headroom left over.

What this means if you're buying a build

We build systems that include Carl, our conversational interface, tuned to one of three capability tiers depending on how demanding the reasoning load is: everyday operations, heavier multi-department reasoning, or complex and regulated work that wants the most capable local model. That third tier is the one the arithmetic above is actually about: it needs meaningfully more machine than the first tier does, and the gap isn't small.

This post used to say we weren't going to state a new hardware number here. Squaring published guidance against this arithmetic was a live internal question we were still working through at the time. It's settled now, the same way every other number in our pricing gets settled: explicitly, priced by formula, not implied by a blog post. Two role-named units are priced at real Apple configured cost plus a flat 25% for BOP to acquire and install them, scaling with which of Carl's three reasoning tiers you need: ~$1,000 for everyday-tier reasoning (a base Mac Mini M4), ~$3,000 for heavier multi-department reasoning (a maxed 48GB Mac Mini M4 Pro), and ~$6,625 for the top tier (a 96GB Mac Studio M3 Ultra, the machine that reasoning load actually needs, not a Mac Mini). Add a flat ~$1,000 Client Unit and the full hardware line runs from about $2,000 to about $7,625 depending on tier. Full detail: BOP-11 v1.5 §M.