The 512GB Mac Studio Missed Its Window: Late October Is Here and You Still Can't Order One (2026)

By RunAIHome Team · 12 min read
mac-studioapple-siliconunified-memorylocal-llmhardwarebuying-guide

TL;DR: Apple said the 512GB M5 Ultra Mac Studio would ship in late October. It’s October 27, there is still no order button and no price, and Macworld now reports delivery estimates for high-end Studios slipping toward January. If your models fit in 256GB, buy that config now; if they don’t, don’t panic-pay $16,000+ for a used M3 Ultra.

Wait for M5 Ultra 512GBM5 Ultra 256GB (buy now)Used M3 Ultra 512GB
Best for265GB+ MoE quants in one quiet boxEverything up to ~190GB wired (gpt-oss-120b, 671B at 2-bit)Same 265GB+ class, needed this month
Price / CostUnannounced (~$15K speculated); reports point to January$9,499, shipping since Sep 22$16,000–$19,750 recent sold prices
The catchNo date, no price, DRAM crunch could slip it againMiniMax M3 Q4 (265GB) will never fit819 GB/s vs 1.2 TB/s, no warranty, ~2× the 256GB price

Honest take: The only reason to pay a used-market premium for yesterday’s 512GB machine is a hard deadline. For everyone else, the 256GB M5 Ultra at $9,499 is the config you can actually buy — and the 500GB-class model wave landing right now mostly doesn’t fit in 512GB anyway.

Back in August we told readers eyeing the 512GB tier of the new Mac Studio M5 Ultra the same thing Apple did: orders open in late October. Our launch-day breakdown flagged that the config had no price and couldn’t be pre-ordered. Two months later, that’s still the complete list of facts Apple has published about it.

It’s now October 27. “Late October” has four days left, no order page has appeared, no price has been announced, and the reporting around Apple’s memory supply has gotten worse, not better. If you’ve been sitting on $10,000-plus waiting for this machine — and our own earlier coverage told some of you to do exactly that — here’s where things actually stand and what to do about it.

The timeline: what Apple promised vs. what exists

The checkable facts, in order:

DateWhat happened
Apr 2026Bloomberg’s Mark Gurman reports the M5 Mac Studio slipping from summer to ~October, blaming DRAM and NAND shortages
Aug 25, 2026Apple announces M5 Max / M5 Ultra Mac Studio; every config orderable except 512GB, listed as “late October” with no price
Sep 22, 2026M5 Max and M5 Ultra 96GB/256GB configs ship
Oct 27, 2026Still no 512GB order page, no price; Macworld reports delivery estimates on high-end configs slipping to January

We’ve been re-checking for an order page on nearly every publishing run since October 18. Nothing. Apple hasn’t formally retracted “late October” — but a config that goes from announcement to order page to delivery inside four days would be a first for this product line, and the coverage is already written in the past tense: Macworld’s benchmark write-up of the M5 Ultra carries the headline “worth waiting 4 months to get one,” and its buying guide warns that build-to-order configs are quoting January.

The cause isn’t mysterious. The same DRAM shortage that has DDR5 kits at 4× their 2025 prices already forced Apple to pull the M3 Ultra’s 512GB config in March 2026 and its 256GB tier by May — the old flagship spent its final months on sale as a 96GB-only machine. A 512GB unified-memory SKU is exactly the product a memory supercycle strangles first, and no analyst we’ve seen forecasts real relief before new fabs come online in 2027–2028.

Why the 512GB tier matters more now than it did in August

When Apple announced this config, the honest question was “what would you even run in 512GB?” October answered it. Three releases in the last two weeks alone target exactly this capacity class — and the answers cut both ways.

What 512GB uniquely runs. MiniMax M3 is the clearest case: Unsloth’s 4-bit UD-Q4_K_XL GGUF is ~265GB. That’s over the 256GB Studio’s ceiling before you load a single token of context — the 512GB machine is the only single consumer box that holds the 4-bit quant with room for KV cache. The same math covers DeepSeek’s 671B line at 4-bit (~404GB), which the old M3 Ultra 512GB ran at a measured 17–18 tok/s under 200W.

What 512GB still can’t run. The frontier kept moving while Apple’s order page didn’t. Kimi K3.1 Pro, whose weights land November 4, is 2.8T parameters — 1.56TB in its native MXFP4, triple this machine’s memory. Mistral Large 4 (1T total, ~52B active, weights due October 31 per the Hugging Face placeholder) pencils out to roughly 620GB at Q4_K_M-class quantization using the ~0.62 GB-per-billion-params rule that has held across recent GGUFs — over the line; it would take a ~3-bit quant (~470GB, our estimate) to squeeze in. Treat both of those sizes as derivations until real GGUFs post, but the direction is clear: “512GB runs everything” was never true, and it’s less true every month.

One operational detail if you’re sizing models against these numbers: macOS wires only about 75% of unified memory to the GPU by default, so even a 512GB machine gives the GPU roughly 384GB out of the box — not enough for the DeepSeek 671B Q4 case above. The fix is one command:

$ sudo sysctl iogpu.wired_limit_mb=458752
iogpu.wired_limit_mb: 0 -> 458752

That raises the wired ceiling to 448GB — the setting the measured M3 Ultra DeepSeek runs used — while leaving the OS ~64GB. It resets on reboot, so put it in a LaunchDaemon if the box is a dedicated inference server. Check your own target model and context against any candidate config with our VRAM calculator before you spend anything.

The used M3 Ultra 512GB market: softer, still a bad deal

Every week the new 512GB machine doesn’t ship is a week the only alternative — a used Mac Studio M3 Ultra 512GB — holds its scarcity premium. The current numbers, from eBay listings and macpro-la’s October price guide:

September 2026Late October 2026
Asking prices$18,250–$27,500$15,499 (open-box)–$24,999
Completed sales—$16,000–$19,750 (Sep 10–Oct 6 tracked sales)
Original 2025 price$9,499—

The floor has come down about $2,750 since we covered this market in September — but a machine that sold new for $9,499 eighteen months ago still transacts at $16,000+. You’d be paying roughly 1.7× the price of a new 256GB M5 Ultra for 819 GB/s of bandwidth against the new machine’s 1.2 TB/s (a 46% decode deficit on anything bandwidth-bound), zero warranty, and a chip two generations back.

Our September verdict was “don’t pay it, wait for the new 512GB in late October.” The second half of that advice just expired, so let’s be precise about the first half: the used premium only makes sense if a specific 265GB-class workload has to be running locally before roughly February. If that’s genuinely you — a contract, a compliance deadline, a product demo — the $16K machine is a tool and tools cost money. If “need” actually means “want,” January is not a long time to save $6,000-plus.

What the speeds actually look like

For the decision below, here’s the measured-vs-estimated picture across the three machines. Measured numbers are from community llama.cpp/MLX runs; estimates are bandwidth-scaled and marked.

ModelM3 Ultra 512GB (819 GB/s)M5 Ultra 256/512GB (1.2 TB/s)
Llama 3.3 70B Q4_K_M16–19 tok/s measured23–27 tok/s
gpt-oss-120b (MXFP4)79.7 tok/s measured~120 tok/s, early community run
DeepSeek 671B-class Q417–18 tok/s measured, <200W~25 tok/s (estimate, 512GB only)
MiniMax M3 UD-Q4_K_XL (265GB)fits, no public benchmark(512GB only; est. 70–95 tok/s from ~11.5GB/token active)

Early community numbers also show engine choice mattering on the new chip — DeepSeek V4 Flash reportedly moves from ~40 tok/s under llama.cpp to ~53 tok/s under MLX on an M5 Ultra, a 34% gain for free. Whatever config you land on, test both runtimes before concluding the hardware is the bottleneck.

Note what the table doesn’t show: nothing in the 500GB class decodes fast because of 512GB. Capacity buys you entry; bandwidth sets the speed, and 1.2 TB/s divided by a big dense model’s per-token reads is still a two-digit number. If your real workload is 70B–120B-class models, the 256GB — or even the M5 Max at $2,499 — delivers the same tokens per second as the 512GB machine would.

What to actually buy

Prices as of October 2026, all verified above:

Your situationThe machinePriceWhere
Your biggest model fits ~190GB wired (gpt-oss-120b, 70B–120B class, 671B at 2-bit)Mac Studio M5 Ultra 256GB$9,499Check price
You need 265GB+ locally before February — hard deadline, not preferenceUsed Mac Studio M3 Ultra 512GB$16,000–$19,750 (sold range)Check price
You want CUDA, a warranty, and max single-box decode speed instead of max capacityRTX PRO 6000 Blackwell 96GB$13,998–$19,999Check price
You can wait and want the 512GBKeep the cash; re-check Apple’s configurator monthly$0 today—
Undecided — want to feel 500GB-class MoE speeds before committing five figuresRented multi-GPU pod (4× A100 80GB and up)pay per hourRunPod

The row most people should pick is the first one. The gap between “what I’d like to be able to load someday” and “what I run every day” is where five-figure mistakes live — and if the 512GB tier does appear at the speculated ~$15,000, the 256GB config plus $5,500 in your pocket will still look like the right call for anyone not running MiniMax M3 or a 671B daily. We made the same argument about the RTX PRO 6000 vs. this machine: buy for the workload you have.

If the models you’re waiting to run are coding agents, remember the API escape hatch exists while you wait — the economics of BYOK backends for Cursor/Cline-style harnesses are covered on our sister site aicoderscope.com, and the self-hosted serving stack side lives at aifoss.dev.

FAQ

Has Apple officially delayed the 512GB Mac Studio? No formal statement. The facts: Apple said “late October” on August 25, no order page or price exists as of October 27, and Macworld reports high-end Studio delivery estimates slipping to January. Apple missing its own stated window without comment is how the M3 Ultra’s memory tiers disappeared in 2026, too.

Why is a $9,499+ computer delayed by a RAM shortage? Because 512GB of LPDDR5X per unit is exactly what’s scarce. DRAM makers have shifted capacity to HBM for AI accelerators; consumer DDR5 has roughly quadrupled since summer 2025, and Apple already cut the old M3 Ultra’s 512GB and 256GB configs for the same reason. High-capacity unified memory competes directly with the most supply-constrained silicon in the industry.

Will the 512GB M5 Ultra run Kimi K3.1 Pro when the weights land November 4? No. K3.1 Pro is 2.8T total parameters — about 1.56TB in its native MXFP4. No single consumer machine runs it; see our K3.1 Pro hardware breakdown for what the realistic access paths are.

Is the 256GB config also delayed? No — it’s been shipping since September 22. Macworld’s warning is about build-to-order estimates stretching on high-end configs, so if you order a 256GB machine, screenshot the delivery estimate at checkout and treat multi-week quotes as normal right now.

What will the 512GB config cost when it ships? Apple hasn’t said — that’s half the story here. Coverage speculates around $15,000 based on the $4,000 Apple charges for the 96→256GB step and its June 2026 memory-price increases, but there is no price list. Budgeting more than the $13,499 a naive doubling suggests is prudent.

  • Mac Studio M5 Ultra — $5,499 (96GB) / $9,499 (256GB); the 256GB config is the one you can actually buy today and covers everything up to ~190GB of wired model weight.
  • Mac Studio M3 Ultra 512GB (used) — $16,000–$19,750 recent sold prices; deadline-driven buyers only.
  • RTX PRO 6000 Blackwell 96GB — $13,998–$19,999 street; the CUDA alternative when decode speed matters more than raw capacity.

Sources

Last updated October 27, 2026. Prices, ship dates, and model availability change quickly — verify current numbers before purchasing.

Was this article helpful?

Get the numbers before you buy

New GPU and mini-PC benchmarks, VRAM thresholds, and price checks — sent when there's something worth acting on, not on a schedule. No spam, unsubscribe anytime.