Mac Studio M5 Ultra: Apple's 512GB Local AI Machine Just Made the Cloud Optional

Yesterday Apple announced a Mac, and the pitch wasn't photo editing. It wasn't even "Apple Intelligence," the feature that spent most of 2025 as a punchline. The press release for the new Mac Studio leads with a phrase I never expected to read from Cupertino: the ultimate desktop for on-device AI.

Read that again. Apple — a company that spent the last year fumbling its own AI features — just shipped a $5,499 box whose entire marketing thesis is running other people's frontier AI models. Open-weight models. On your desk. With a straight face, they wrote "without counting tokens or worrying about rising cloud costs" into the official announcement. Someone at Apple has been reading your API invoices.

The machine in question is the Mac Studio with M5 Ultra, announced August 25 alongside a new M5 Max Studio and an M6 Mac mini. And underneath the marketing gloss, it's the clearest signal yet of where Apple thinks AI is going: not everyone wants to rent it.

The specs that actually matter

Here's the short version for anyone who doesn't speak Apple silicon fluently:

• 36-core CPU, up to 80-core GPU — the biggest GPU Apple has ever built, and the first Ultra with Neural Accelerators baked into every GPU core (the dedicated matrix-math hardware previously reserved for the lesser chips)
• Up to 512GB of unified memory — RAM the GPU can use directly, no copying, no VRAM ceiling
• 1.2TB/s of memory bandwidth — 50% more than the M3 Ultra
• 4.3x the peak AI compute of the M3 Ultra, and 9.8x the M1 Ultra
• In LM Studio, Apple claims 4x faster LLM prompt processing than the M3 Ultra

Now the honest footnote: 512GB is not new. The M3 Ultra offered it starting last year. What's new is everything around it. Last generation you got a walk-in closet with a cat flap for a door. Huge capacity, but moving data in and out of it was slow enough that big models crawled.

The M5 Ultra finally fixes the door.

Why bandwidth is the whole ballgame

Here's the thing most people miss about running LLMs locally: inference is mostly a memory bandwidth problem wearing a GPU costume. Every single token a model generates requires sweeping essentially all of its weights out of memory. The model isn't "thinking" slowly — it's reading itself slowly. Megabytes of model per token, times your tokens per second, equals your bandwidth bill.

So 1.2TB/s matters more than any CPU or GPU core count on the spec sheet. Some context:

• An RTX 5090 — the fastest consumer Nvidia card — has 1.79TB/s of bandwidth, but only 32GB of VRAM. Great speed, nowhere to put a big model.
• An H100 has 3.35TB/s and 80GB. It also costs about $30,000, if you can find one.
• The M5 Ultra has 1.2TB/s and 512GB.

That's the trade Apple is making, and for LLMs it's the right one. A 5090 is faster at small models. Nothing at any consumer price point touches the M5 Ultra at big ones, because big models live and die by capacity first, bandwidth second. The arithmetic is brutal and unforgiving: a 4-bit quant of a 400-billion-parameter model is roughly 200GB before you even allocate the KV cache. That doesn't fit in 32GB. It doesn't fit in 80GB. It fits in 512GB with room to spare.

"Realistically, what can I run on this?"

That's the top question in every thread about this machine — someone asked it verbatim on Hacker News within an hour of the announcement. The honest answer, by memory tier:

• 128GB (M5 Max Studio, $2,499+): 70B-class models comfortably, 120B-class quants with squeezing. Everything a normal person needs.
• 256GB: 200B+ dense models. The models that beat GPT-4 from two years ago, at home.
• 512GB: 400B-class dense models, or the mid-tier MoE frontier models, unquantized. One Hacker News commenter with deep knowledge of the die architecture estimated a DeepSeek V4-class model running unquantized at roughly 50+ tokens per second generation and ~1,000 tokens per second prefill on the Ultra — genuinely competitive with cloud latency for single-user workloads.

And then there's the weird one. Apple is explicitly pitching clustering: connect four Mac Studios over Thunderbolt 5 with RDMA, and they behave like one machine with a shared ~2TB memory pool. Apple claims up to 3x the inference throughput of a single system — which, do the math, is about 75% scaling efficiency, and is honestly better than I expected over what is fundamentally a very fast cable. Four of these boxes will set you back somewhere north of a used car, but they'll run models that don't fit on any single GPU on earth, including Nvidia's. It's the most expensive LEGO set Apple has ever sold, and I kind of love that it exists.

The gap everyone's complaining about: there's no 1TB option. One commenter worked out the likely reason — the Ultra is two Max dies fused together, and 32GB base chips × 16 = 512GB. It's a packaging limit, not a strategy. Still, 512GB sits right at the edge of "enough" for the trillion-parameter class at 4-bit. Next generation, presumably, jumps the wall.

The pricing is where the fairy dust runs out

Base M5 Ultra Studio: $5,499. That gets you 96GB of memory. Yes, you read that right — the machine whose headline feature is 512GB of memory ships with less than a fifth of it. "$5,500 for 96GB of RAM" was, fairly, one of the top reactions yesterday.

From there the configure-to-order ladder gets steep. Early reports put the 256GB config around $10,000, and the 512GB version — which doesn't even ship until late October, its pricing apparently still being finalized — will climb well past that. This is happening against the backdrop of the worst RAM pricing in a decade; memory costs are up as much as 500% this year, and Apple passes that along with its customary markup. Nobody should pretend this is a value play.

Here's the counterargument, though, and it's the one the cloud crowd keeps missing: tokens are rent, memory is ownership. A frontier API subscription costs $200/month per seat, forever, with rate limits and a privacy policy you didn't write. If you're a lawyer, a doctor, a defense contractor, or just a company whose codebase is a trade secret, "just use the API" was never actually on the menu. For agent workloads — the ones that sit in a loop for hours, reading files, re-reading context, burning a million tokens on tasks you didn't even watch — the meter running is the product killer. An agent with no API bill is a different species of tool.

One Hacker News commenter nailed the skeptic's case: you can rent frontier cloud models for less than the monthly finance charge on this thing. True! If you are one person who wants to chat with a smart model, buy the cloud subscription and a nice dinner instead. This machine is for people shipping local AI products, bound by compliance, or allergic to the idea that their assistant's brain lives in someone else's warehouse.

The quiet part: Apple is building the software stack too

The hardware is only half the story, and it's the less interesting half.

Buried in the announcement: a brand-new macOS framework called Core AI — a first-party way to build, run, and deploy models on Apple silicon, with full-scale local LLMs as the explicit target. MLX, Apple's open-source ML framework, keeps maturing. LM Studio — the de facto "Steam for local models" — got name-checked with a new release, and Apple is now openly talking about on-device coding agents in Xcode.

For years the local AI scene ran on community tools and a prayer: llama.cpp, Ollama, a thousand quantization formats, and Apple pretending not to notice. Now Apple is shipping first-party inference infrastructure and putting "run enormous LLMs entirely on device" in a press release. When a company that ignores a market for a decade suddenly bundles it into the OS, the market has arrived.

Also worth savoring: the same day Apple announced this, OpenAI was on the front page of every tech site bragging that its new datacenter chip beats Nvidia's. Both stories are true at once, and that's the actual state of AI in 2026: the frontier needs warehouses full of custom silicon that cost more than a navy destroyer — and a meaningful chunk of the market would rather not depend on any of it. The future isn't cloud OR local. It's cloud for the hard problems, and a quiet, fast, private box under your desk for everything else. Apple just picked its side of that split, aggressively.

So should you care?

If you build with AI: yes, even if you never buy one. The M5 Ultra is the most mainstream validation yet that local inference is a real category, not a hobbyist curiosity. Open-weight model makers — DeepSeek, Qwen, GLM, Kimi, MiniMax — now have a flagship consumer machine explicitly designed to run their biggest models. That feedback loop (serious hardware ↔ serious open models) is what pulls the whole local ecosystem up.

If you want in cheaply: the M5 Max Studio at $2,499 with 128GB runs more model than most people will ever need. The new Mac mini starts at $899 — but note the base M6 caps at 32GB of memory and 170GB/s of bandwidth, which is fine for small agentic models and nothing more. Apple sells you the dream at $899 and the reality at $5,499-plus. That's not cynicism, that's a spec sheet.

And if you're the kind of person who read this far because you've already got opinions about KV cache sizes — you're exactly who we built CopperRiver for. It's a desktop AI assistant for your Mac that runs on those same open-source models and actually does things: browses, runs terminal commands, reads your files, automates the boring parts of your day. Local-first, no token anxiety. The M5 Ultra is overkill for it. The Mac you already own probably isn't.

The cloud is someone else's computer. Apple just spent a whole press release arguing yours is due for a promotion. For once, the marketing and the math agree.