Mac Studio Memory Roadmap

1. Why this matters now

For local LLM work, Apple’s pitch has never been “beats an H100 at FLOPs.” It has been: one coherent memory pool big enough to hold the weights, quiet enough to live under a desk, efficient enough that a full rack of GPUs is overkill for a single researcher’s agent loop.

The March 2025 M3 Ultra Mac Studio made that concrete with a 512GB unified-memory ceiling — Apple’s “most memory ever in a personal computer” line. Then the global DRAM shortage ate the high SKUs. By early 2026 the 512GB option disappeared from the configurator; later storefront snapshots showed M3 Ultra configs pinned near the 96GB floor while delivery estimates stretched months out.

Capacity is the product. Bandwidth is the tax. Supply decides which SKUs exist.

So when Bloomberg-sourced reporting says Apple has tested 768GB on an M5 Ultra Studio and is looking at 1.5TB on an M7 Ultra — possibly never sold to the public — that is not trivia. It is the entire local-AI value proposition for this form factor, under a memory market that is actively hostile to huge consumer configs.

2. What we already know: M3 Ultra 512GB

Anchor numbers from Apple’s own Mac Studio tech specs (current-generation Ultra SKU as shipped in 2025):

Side-by-side on the same product page: M4 Max Studio configs advertise 410 GB/s (32-core GPU class) or 546 GB/s (40-core GPU class). Ultra is still the bandwidth and capacity king of the consumer Mac line — just not an HBM monster.

3. Hints teased out of the rumor mill

None of the following is Apple-confirmed silicon. It is a chain of secondary reporting (MacRumors, Macworld, 9to5Mac, Tom’s Hardware) paraphrasing Mark Gurman’s Bloomberg Power On notes and related supply-chain pieces. Treat every number as tested / planned / supply-gated, not as a store SKU.

Hint A — M5 Ultra has been tested to 768GB

June 2026 coverage: Apple plans an M5 Ultra Mac Studio with roughly 36 CPU cores and 80 GPU cores, and has tested support for up to 768GB of unified memory — a step above the prior 512GB ceiling. Same articles stress that supply constraints may block shipping that SKU at launch.

That is the clean answer to “will there be another 500 / 512 class high end?” The technical ceiling in the rumor set is 768, not a rehash of 512. Whether you can buy 768 is a separate, uglier question.

Hint B — Internal code J246, earlier 2026 slipped to later 2026

Macworld notes an internal code J246 for the M5 Ultra Studio, replacing the March 2025 M3 Ultra machine. Earlier expectations of a first-half 2026 ship slid toward an October 2026 window after memory-chip supply and price shocks. By late July 2026, broader Mac roadmap pieces still put M5 Ultra “this year” but with timing described as memory-supply dependent and currently unclear.

Hint C — Pricing hints are illustrative, not SKUs

After the June 2026 Mac price increases (Ultra base climbing to the mid-$5k range), secondary pieces floated that a fully loaded high-memory M5 Studio could clear $10,000. Forum math goes higher. None of that is a leaked Apple price sheet — it is extrapolation from DRAM cost and prior $4k 512GB uplifts.

Hint D — 1.5TB is an M7 Ultra story, and maybe not a product

Gurman-sourced Macworld reporting: while preparing M5 Ultra, Apple is also working on an M7 Ultra that supports up to 1.5TB of memory — roughly double the planned M5 Ultra high mark. The same coverage warns the public may never be able to buy that configuration; internal AI-server use and RAM-market conditions gate what becomes a BTO option. Arrival talk centers on ~2028–2029, not the next refresh.

Hint E — M5 Ultra for internal AI servers too

Separate from the Studio: Apple is described as building internal AI servers using M5 Ultra. Apple does not sell servers, so this is Apple-for-Apple capacity. It still matters: silicon that must host large on-device / on-prem inference for Apple’s own stack is more likely to keep extreme memory configs in the design envelope even if the public never sees the top bin.

4. Chip roadmap: M5 → skipped M6 high-end → M7

This is the part people get wrong when they say “wait for M6 Ultra” or “M7 next year.” The rumor consensus as of mid–late July 2026:

Practical planning rule: if you are buying a desk-side Apple box for models in the next 12–18 months, the decision is M3 Ultra (if you can find memory) vs wait for M5 Ultra. “Wait for M6 Ultra” is a dead branch in the current rumor tree. “Wait for M7 Ultra / 1.5TB” is a 2028-class bet with non-trivial chance the top bin never becomes a consumer SKU.

5. Memory bandwidth by generation

Capacity decides which models fit. Bandwidth decides how fast tokens leave the machine once weights are resident — especially decode, long-context attention traffic, and multi-agent concurrency.

Editorial pieces have floated M5 Studio configs “exceeding 600 GB/s.” Treat that as a floor-ish narrative for high Max/Ultra class, not a measured Ultra number. Until Apple prints a GB/s figure on a press slide, do not bake 1.2 TB/s into a purchase model.

6. What models you can run (capacity math, not magic)

Rough GGUF / MLX planning weights (order-of-magnitude; KV cache and runtime overhead still bite):

What 512GB actually unlocked

The M3 Ultra 512GB machine’s superpower was never “fastest tokens in the lab.” It was:

• Resident 400B-class Q4 without sharding across cards.
• Long context without immediately swapping weights off GPU memory — everything is the same pool.
• Multi-process local agents — several mid-size models or one giant + embedding + rerankers without PCIe stage-in choreography.
• Quiet always-on inference for private code, medical-adjacent notes, or air-gapped work where cloud is a non-starter.

What 768GB adds over 512GB

7. Mac Studio vs RTX PRO 6000 vs DGX Station

Three different answers to “local AI on a desk.” Comparing them as if they were the same product is how people waste money.

How to choose without self-gaslighting

• Pick Mac Studio if your bottleneck is model size + privacy + noise + single-box simplicity, and you live in Apple’s toolchain (or accept MLX/llama.cpp as primary).
• Pick RTX PRO 6000 if your models fit in 96GB (or you multi-GPU), you need CUDA, and tokens/s + pro graphics matter more than holding a 400B dense Q4 alone.
• Pick DGX Station if you are essentially buying a personal mini-cluster — coherent ~0.75 TB, HBM-class bandwidth, trillion-parameter marketing claims — and budget/power/support match that ambition.

8. Is 768GB worth upgrading from M3 Ultra 512GB?

Upgrade for capacity and architecture — not for a free 50% tok/s.

Relative to a fully optioned M3 Ultra 512GB that you already own and use: +256GB is meaningful only if you have a workload that hits the wall. Relative to a stripped current storefront (96GB Ultra while 512 is gone): any high-memory M5 Ultra that actually ships is a different machine class again.

Price discipline: if high-memory configs land in the $10k–$15k+ rumor band while used or leftover 512GB Ultras exist, do the dollars-per-GB math including resale. DRAM scarcity has already shown Apple will delete SKUs rather than eat infinite margin compression.

9. Buy / wait checklist

• Need huge models this quarter? Hunt remaining high-memory M3 Ultra inventory / secondary market; do not assume 512 returns.
• Can wait into late 2026? Watch for M5 Ultra announcement with an explicit max unified-memory line item — and whether 768 is build-to-order or “coming later.”
• Planning 2027+? Assume no M6 Ultra Studio. Revisit M7 Pro/Max (late 2027) and M7 Ultra (2028) only if 1.5TB remains in public copy.
• CUDA-native or tok/s-max? Price RTX PRO 6000 workstations or multi-GPU boxes in parallel; do not force Apple silicon if the stack is wrong.
• Enterprise deskside? Put DGX Station on the spreadsheet with power, support contracts, and software — not just the 748GB headline.

10. Sources & confidence

Official / primary-ish

• Apple Newsroom — Mac Studio / M3 Ultra (March 2025) — 512GB “most ever” framing.
• Apple Mac Studio technical specifications — M3 Ultra 819 GB/s; M4 Max 410 / 546 GB/s.
• NVIDIA RTX PRO 6000 Blackwell — 96 GB GDDR7, 1,792 GB/s, 600 W class.
• NVIDIA DGX Station — ~748 GB coherent memory, GPU HBM bandwidth in the multi-TB/s class, Grace + Blackwell Ultra deskside.

Rumor / secondary (memory ceiling & roadmap)

• MacRumors — M5 Ultra Mac Studio, up to 768GB (June 25, 2026)
• MacRumors — 512GB option removed (March 2026)
• Macworld — 1.5TB Ultra / M7; may never ship publicly
• MacRumors — M7 Ultra Studio ~2028, cooling
• 9to5Mac — skip M6 Pro/Max
• Tom’s Hardware — M7 Ultra 1.5TB / Blackwell-class AI language
• MacRumors — Mac overhaul; supply-dependent timing (July 22, 2026)
• Bloomberg Power On / Mark Gurman newsletters (paywalled primaries behind much of the above)

Confidence tags