Apple Just Priced the Local Thesis

Here’s a sentence from a press release Apple put out on August 25:

“Combined with up to 512GB of unified memory and 1.2TB/s of memory bandwidth, 50 percent higher than before, Mac Studio lets users run massive models entirely on device with complete privacy — without counting tokens or worrying about rising cloud costs.”

The AI labs rent you a meter. Apple sells you a box. And Apple just used its own press release to tell you the meter is a problem you can buy your way out of. Same day, it put a price on the way out: the new Mac mini starts at $899, and the new Mac Studio tops out at a starting price of $5,499. On September 22 the machines started showing up.

I didn’t see this coming — Apple going all in on local compute. I’ve been arguing two things in this section that Apple has now built a product line around. In June I wrote that “[t]he strongest confidentiality posture is client data that never leaves the building”, and that “[o]n Apple silicon there is a specific structural advantage: unified memory.” In July I wrote that the value in this business ends up with whoever owns compute, and that a firm should “[a]ssume local gets better every year.” Anybody can make a forecast. A bet with a price tag on it means something.

It’s also a sales pitch, so read it like one. And I’m not a neutral reader: this firm runs on Macs, and the local setup I’ve written about in this section runs on Apple silicon.

What $899 to $5,499 gets you

Start with the parts that aren’t marketing: prices, memory, bandwidth, and the port that decides whether you can chain machines together. None of this is a performance claim.

Apple shipped software too. The Studio release calls Core AI “a brand-new framework for building, running, and deploying AI models on Apple silicon.” MLX, which the release calls “Apple’s open-source machine learning framework optimized for Apple silicon,” is the runtime I talked about in Run It Yourself, Then Let the Client Choose.

Now hold that table up against the memory ladder from my July 27 tool essay. That ladder put 16GB at “[r]eal classification and extraction work. A genuine starting point, not a toy”, and 64GB at “the practical target for a firm that wants local drafting.” So the $899 mini is a box for sorting documents and pulling facts out of them. The M5 Pro mini is the first machine in the new lineup you can configure up to the drafting tier. Apple lists it starting at $1,699 and says it “supports up to 64GB”; Apple doesn’t say what the 64GB version costs.

The 512GB ceiling isn’t new. Apple’s own footnote says the machine it benchmarked against was a previous-generation Mac Studio with M3 Ultra and “512GB of unified memory”. Same memory as before. What got better at the top is bandwidth.

And you can’t buy the 512GB one yet. Apple said so on August 25 and again on September 22, in one line: “Mac Studio with 512GB of unified memory is coming in late October.”

Treat Apple’s numbers like the other side’s expert

Apple’s releases are full of multipliers: “up to 4.3x faster AI performance” for the Studio, “up to 4x faster AI performance” for the M6 mini, “[u]p to 9.8x faster LLM prompt processing in LM Studio when compared to Mac Studio with M1 Ultra”. Every one is Apple’s own number, measured by Apple, on systems Apple picked, and every one starts with “up to”. Treat them the way you’d treat the other side’s expert. He’s not lying. He’s also not working for you.

The two August 25 releases don’t even use the same number. The Studio release says the M5 Ultra hits “up to 4.3x the peak AI compute performance of M3 Ultra”. The chip release says the same chip offers “up to 4.5x the peak GPU compute for AI compared to M3 Ultra”. Same chip, same company, same day, two numbers. A multiplier tells you what the author decided to measure. That’s about all it tells you.

The more useful thing is that Apple’s own copy says which part of the machine buys which kind of speed. The chip release credits the M6’s new GPU Neural Accelerators with “significantly faster prompt processing when interacting with on-device LLMs.” It credits the memory with something else. Of the M5 Ultra’s 512GB of memory and 1.2TB/s of bandwidth, it says “[t]his lets users store huge datasets entirely in local memory, increase the tokens-per-second speed, and run huge LLMs with hundreds of billions of parameters entirely on device.”

Prompt processing is how fast the machine reads what you hand it. Tokens per second is how fast it writes back. Go by Apple’s own labels and the LLM multipliers are reading-speed numbers, and the headline multipliers are peak-compute numbers. The memory, which Apple credits for the writing speed, got nowhere near four times better. Capacity stayed at 512GB. Bandwidth moved less than the headlines: the M5 Ultra’s 1.2TB/s is “50 percent more than M3 Ultra”, and the M6’s 170GB/s is “a 10 percent increase over M5”.

For a law firm, that’s mostly good news. The work I’ve said belongs on your own hardware is reading work. Sorting an inbound document, pulling the dates and parties out of an order, a first pass for relevance: long input, short output. A machine Apple says reads up to four times faster, and writes somewhat faster, fits that work. The hard writing — drafting an argument, holding a long analysis together — still goes to the best model you can get, and nothing in these releases changes that.

Clustering gets the same treatment. Apple says “[a] cluster of four Mac Studio systems delivers up to a remarkable 3x faster AI inference than a single system.” Four machines, up to three times the speed. At the M5 Ultra’s starting price, that’s $21,996 of hardware. If you’re a lab that needs to hold a model too big for one machine, clustering is the only way to run it at all, and the math doesn’t matter. I don’t know of a law firm workload that needs it.

What a new chip doesn’t fix

A new chip changes the price and speed of running a model in your own office. It doesn’t touch four other things, and the press releases don’t mention any of them.

It won’t make the model smarter. A Mac Studio runs open-weight models. Apple’s releases talk about “frontier-class” models on your desk. That’s Apple’s phrase. It isn’t a measurement. Whatever the best open-weight model can do on the hardest multi-step legal analysis, it does no better on a $5,499 machine than on a rented server. Hardware makes a model faster. It doesn’t make it smarter.

It still makes things up. A local model fabricates a citation exactly as readily as a remote one. You still check every authority against the source before anybody relies on it.

It can still walk out the door. “[C]omplete privacy”, in Apple’s sentence, means privacy from the model vendor. It doesn’t mean privacy from a thief, from a sync client quietly uploading a working folder, from an unencrypted backup drive, or from whoever walks up to an unlocked screen. Rule 1.6(c) requires “reasonable efforts to prevent the inadvertent or unauthorized disclosure of, or unauthorized access to, information relating to the representation of a client”, and a stolen Mac Studio with an unencrypted disk is a disclosure no matter how local the model was.

Somebody else made the model. Apple sells the hardware and the frameworks. The model you’ll actually run gets downloaded from someone else. As I laid out in Executable Code You Are Calling a Model, a downloaded checkpoint is an untrusted executable until proven otherwise, and a new Mac makes it more likely you’ll download one this quarter.

The footnote Apple hopes you skip

Apple’s September 22 release carries a footnote about its own assistant. If you’re a lawyer trying to figure out what “local” means on a Mac, it’s the most important sentence in all four releases:

“Certain Apple Intelligence features that rely on server-side models are subject to daily usage limits, including but not limited to Siri AI, intelligent photo editing tools, Image Playground, and AFM 3 Cloud models in Shortcuts. Daily limits may vary by feature, request complexity, system demand, system policies, and other factors. Expanded access to such features will be available for a fee in the future.”

Start with Siri. Apple’s own flagship assistant isn’t purely local. The same releases say Siri AI can “draw on personal context to help users find what they need in the moment across their messages, emails, and photos”. On a lawyer’s Mac, those messages and emails are client communications. The footnote tells you at least some Siri AI features run on server-side models. What they send, and on what terms, is in Apple’s Apple Intelligence terms, not the press release, and Opinion 512 is plain about the baseline at page 7: “all lawyers should read and understand the Terms of Use, privacy policy, and related contractual terms and policies of any GAI tool they use”. Buying a machine to run models locally doesn’t make every AI feature on it local. Read the terms before you turn the assistant on for a machine that holds client files.

Then there’s the meter. “Expanded access to such features will be available for a fee in the future” is a company that sells boxes telling you it’s going to sell usage too. Apple is hedged. It sells compute to the people who want to own it, and it’ll rent server capacity to the people who want convenience. That line in the press release about not counting tokens is one side of the hedge.

Then the calendar. The 512GB version ships in late October. If you read “run massive models entirely on device” on September 22, you couldn’t yet buy the machine that sentence describes at its biggest.

And $5,499 is real money for a solo. That’s the starting price. Apple’s own lease puts the starting M5 Ultra Studio at $110.10 a month over 36 months, and the M5 Max Studio at $48.99. Set those against your monthly API bill. For a lot of small firms, the bill will be smaller.

What I’d do this quarter, in order

In Running Models on Your Own Hardware I said test first, buy second. Apple’s announcement doesn’t change the order. It changes the prices at step five.

1. Pick one narrow, read-heavy task you do all the time. Sorting inbound documents is still the best place to start.
2. Build your test set before anything else. Thirty labeled examples from your own work, measured against a hosted model as the baseline, the way I laid it out in How Do You Know It Works?. Without that number, a purchase is a hunch.
3. Run the smallest local model that might work, on a machine you already own. If a 4B or 8B model clears your bar on classification, the $899 mini is your whole capital budget.
4. Let the test pick the memory. If the job needs a 27B or 32B model with real context, the July ladder says 64GB. That means an M5 Pro mini configured up, or an M5 Max Studio. Buy what the measurement says, not what the keynote says.
5. Price it against your API bill, Apple’s lease figures included, over three years. Then add the setup and upkeep time, which I said in June people underestimate. They still do.
6. Lock it down before it holds anything. Full-disk encryption, no sync client watching the folders the model reads from or writes to, an encrypted backup, and a rule about what gets downloaded and from where.
7. Decide about Siri AI on that machine on purpose, after you’ve read Apple’s terms. Don’t just click through the setup screen.
8. Wait on the Ultra. The 512GB version doesn’t ship until late October, and I don’t know of a small-firm workload that needs it. Let somebody independent test it before you pay for Apple’s multipliers.

The one reason to move faster is a client who’s asked for it. In Run It Yourself, Then Let the Client Choose I described a tier election where a client can require that nothing about the matter leaves the firm’s own hardware. If you offer that tier, you need hardware that can honor it. Apple just priced that hardware starting at $899.

The July essay, sold at retail

In July I argued that the model itself is headed toward commodity, and that the lasting value lands with whoever owns compute, and under that, power. I wrote it about data centers, substations, and interconnection queues.

Apple’s move is the retail version. Don’t rent the customer compute. Sell it to them. Every firm that buys a Mac Studio to run models becomes a very small compute owner, and the value of that compute stays with the firm instead of padding a provider’s margin. I asked in July who’d own compute. Apple’s answer: you, one box at a time.

The July essay also named the most likely way it would turn out wrong: compute stops being scarce and the value moves again, “most likely to whoever owns distribution and customer demand.” Apple is set up to win either way. If compute stays scarce, it sells compute. If there’s a glut, it owns distribution: the devices, the operating system, and the assistant sitting on top of them with a fee coming “in the future.” That’s why I read August 25 as a well-hedged bet, not a reckless one. Take it seriously. Just don’t assume Apple needs it to pay off.

What would prove me wrong

Here’s how I’d know I blew this one.

• Apple moves the pitch to its servers. If Apple’s next hardware releases lead with metered, server-side Apple Intelligence features and quit leading with local, August 25 was a hedge dressed up as a bet, and I read it wrong.
• The top machine slips. If the 512GB Mac Studio doesn’t ship in late October like the release says, or ships and quietly disappears, the high end of the bet was marketing.
• Independent testing misses Apple’s numbers by a mile on the work lawyers actually do: long-document classification, extraction, summarization. Apple’s multipliers are Apple’s. If third-party tests on those tasks land way below them, the price-performance case in this essay goes down with them.
• The buyers send everything to the cloud anyway. If, a year from now, the small firms that bought these machines to run models locally are sending basically all their model work to hosted providers, then local was a feature people paid for and didn’t use, and the market passed on the confidentiality setup I’ve been pitching.

Sources

• Apple Newsroom, Apple introduces new Mac Studio with M5 Max and M5 Ultra (Aug. 25, 2026) — the “without counting tokens” sentence; memory ceilings (128GB M5 Max; 512GB M5 Ultra); 614GB/s and 1.2TB/s bandwidth; Thunderbolt 5 and RDMA clustering and Apple’s “up to 3x” four-system claim; Core AI and MLX; Apple’s “up to 4.3x” and prompt-processing multipliers; footnote identifying the 512GB M3 Ultra comparison system; Siri AI “personal context” description; Studio prices (“starts at” $2,499 and $5,499) and lease figures ($48.99 and $110.10 per month, 36 months); 512GB configuration “coming in late October”
• Apple Newsroom, Apple introduces M6 and M5 Ultra for a big leap in performance and AI compute (Aug. 25, 2026) — Apple’s “up to 4.5x the peak GPU compute for AI” figure; Neural Accelerators tied to prompt processing; bandwidth tied to tokens-per-second; bandwidth increases (50 percent over M3 Ultra; 10 percent over M5)
• Apple Newsroom, Apple unveils a more powerful Mac mini featuring the all-new M6 and M5 Pro (Aug. 25, 2026) — M6 memory (16GB standard, up to 32GB) and 170GB/s; M5 Pro up to 64GB and 307GB/s; Thunderbolt 4 on the M6 model and Thunderbolt 5 on the M5 Pro model; Apple’s clustering statement; “up to 4x” claim; mini prices (“starts at” $899 and $1,699)
• Apple Newsroom, The new Mac mini and Mac Studio are available today (Sept. 22, 2026) — availability date; prices and lease figures repeated from the August 25 releases; the Apple Intelligence server-side usage-limit and future-fee footnote quoted in the text; 512GB configuration “coming in late October” repeated
• ABA Standing Committee on Ethics and Professional Responsibility, Formal Opinion 512 (July 29, 2024) — p. 7 (read and understand the terms of any GAI tool used)
• Minn. R. Prof. Conduct 1.6 — paragraph (c) (reasonable efforts against inadvertent or unauthorized disclosure or access)
• Companion essays in this section: Running Models on Your Own Hardware (June 8, 2026), The Model Will Be Free. The Electricity Won’t. (July 27, 2026), Run It Yourself, Then Let the Client Choose (July 27, 2026), How Do You Know It Works?, and Executable Code You Are Calling a Model

Commentary on technology and the business of law. Not legal advice, not ethics advice, and not a hardware or vendor recommendation. Every performance multiplier above is Apple’s own claim from its own testing, attributed as such; prices, memory, bandwidth, and ship dates are as stated in Apple’s releases of August 25, 2026, and the availability release of September 22, 2026. The predictions and stated falsifiers are the author’s. No client information appears in this article.