M5 Ultra 512GB Arrives: 768GB Vaporware, 384GB GPU Limit, Which Memory for Your LLM?

Three Key Takeaways

1. "Availability" requires quotes — The 512GB version went on order in mainland China on October 29 with a 10–12 week delivery window (late November to early December). Apple has confirmed Mac and iPad supply constraints will persist through the end of 2026 due to memory shortages.

2. 768GB is essentially vaporware for this generation — Apple's official site still lists 512GB as the maximum. The 768GB figure comes only from a June Mark Gurman report stating Apple internally tested it but supply constraints might prevent launch. The 1.5TB target for M7 Ultra in 2028 is similarly a design goal, not a shipping guarantee.

3. 512GB buys capacity, not speed — It uses the same M5 Ultra die and 1.2TB/s bandwidth as the 256GB model, so token generation speed is identical. The extra ¥116,250 only lets you load larger models. Before paying, you must identify a specific model that exceeds 256GB; if you cannot name one, do not upgrade.

Clarifying "Availability"

Apple's August 26 announcement priced the M5 Ultra Mac Studio starting at ¥46,999 (96GB/1TB) and stated the 512GB configuration would be available "late October." That window was met: on October 29 the 512GB option went live at $22,299 (US) / ¥172,999 (China, 512GB+1TB). However, "on sale" does not mean "in stock." The quoted delivery lead times are 10–12 weeks for 512GB, 13–15 weeks for 256GB, 5–6 weeks for 96GB, while M5 Max models ship immediately. Apple's November 3 earnings call confirmed global DRAM shortages will keep Mac and iPad supply tight through the end of 2026; the US store also says "late November to early December."

Price comparison stings: the 2025 M3 Ultra 96GB→512GB upgrade cost only $4,000 (full system $9,499). Today the same upgrade costs $15,500 (full system $22,299), a per-GB increase from $9.62 to $37.26 — roughly 3.9×. Tim Cook warned in April that memory costs had risen significantly; this is a market-wide DRAM price pass-through, not an Apple-only markup.

Note that 512GB is only offered with the top-tier chip (36-core CPU + 80-core GPU). The base M5 Ultra (30C/64C) maxes out at 256GB, and M5 Max at 128GB. The ¥172,999 figure breaks down as: ¥46,999 base + ¥9,750 for 36C/80C + ¥116,250 for 512GB. Maxing storage to 16TB brings the total to ¥229,249.

Counter-intuitive math: two 256GB machines (each 1TB SSD) cost $21,598 combined — about $700 less than a single 512GB unit at $22,299 — and you get two chips and double the SSD. The trade-off is non-shared memory; clustering requires Thunderbolt 5 + RDMA support in your software stack.

768GB: Not Coming This Generation

June 28, 2026 — Bloomberg's Mark Gurman reports Apple tested up to 768GB unified memory on M5 Ultra Mac Studio, but notes supply constraints could delay launch and push top-tier price above $10,000.

August 26, 2026 — Apple's official press release states "up to 512GB unified memory" with no mention of 768GB .

October 28–29, 2026 — 512GB orders open; the online configurator still shows only 96/256/512GB tiers, no 768GB .

2028 — Gurman says M7 Ultra's design target is 1.5TB, but explicitly "depends on supply chain realities"; at current enterprise DRAM pricing, memory alone could exceed $35,000.

Background: Apple cancelled M6 Pro/Max/Ultra; this year only base M6 ships (in Mac mini). M7 base arrives H1 2027, M7 Pro/Max late 2027, M7 Ultra in 2028. Thus Mac Studio jumps from M5 Ultra straight to M7 Ultra — M5 Ultra is the only high-end Mac Studio refresh for nearly two years. Waiting for "more memory + next generation" means waiting until 2028, and even then 1.5TB is not guaranteed.

Scope reminder: 768GB (M5 Ultra internal test) and 1.5TB (M7 Ultra design target) are both from Bloomberg's Mark Gurman; Apple has never officially announced either, and both carry "subject to supply constraints" caveats. The only officially announced, currently orderable maximum is 512GB.

Your Local LLM Ceiling Is Actually 384GB

The most critical section: 512GB unified memory ≠ 512GB available to the model .

macOS defaults to wiring roughly 75% of unified memory to the GPU. On a 512GB machine that yields about 384GB out of the box. To access more, you must manually raise the limit, e.g.: sudo sysctl iogpu.wired_limit_mb=483328 (≈472GB). This setting does not persist across reboots , and you must never set it to the physical maximum — the kernel, WindowServer, and foreground apps need headroom. Community guidance suggests reserving 32–64GB; a practical stable range is 448–464GB . Check the current value with sysctl iogpu.wired_limit_mb; a return of 0 means the default is in effect.

Beyond that, you must reserve space for KV cache and framework overhead. Community testing shows GLM-5.3-Flash (177.6GB) runs fine at 64K context but OOMs at 128K; a 6-bit Flash-Next at 176.6GB memory pressure was killed by macOS. Rule of thumb: add 20% headroom on top of model weight size ; do not run the machine at its absolute limit.

Key insight from the table: the vast majority of everyday models fit under 200GB , so 96GB or 256GB suffices. 512GB provides zero speed advantage (same chip, same 1.2TB/s bandwidth; token write speed is bandwidth-bound). It only unlocks the 265–420GB tier — models like MiniMax M3, Mistral Large 4 2-bit, GLM-5.3 4-bit.

Above that, single-machine is a hard wall: Kimi K3 native 1.56TB, even 1-bit quantization needs ~610GB; DeepSeek V4 Pro 4-bit ~837GB — no single Mac can hold them . Community experiments clustered four 512GB Mac Studios via Thunderbolt 5 to run an 817GB quantized model, achieving ~14.7 tok/s (single-run community measurement, not a standardized benchmark). Apple officially claims four-node clusters deliver "up to 3× inference performance," not linear 4× memory scaling.

Common pitfall: do not estimate model size from its name . Qwen3.8-Flash-Next is labeled "125B" but includes ~55B extra lookup parameters; its actual 4-bit download size is 106–111GB, not the ~69GB implied by the parameter count. Check the Hugging Face repository download size before purchasing.

Which Tier Should You Buy?

Final verdict: 768GB is this generation's vaporware; 512GB is real but ships in 10–12 weeks; the true determinant of "can I run it" is not the 512GB sticker but macOS's default 384GB GPU wired limit, and whether you can name a specific model that absolutely requires 512GB.

Data Sources

Mainland China base price ¥46,999 (M5 Ultra 96GB/1TB), M5 Max ¥19,999, pre-order 8/27 / ship 9/22, 512GB "late October", 36C/80C, 1.2TB/s bandwidth, Thunderbolt 5 + RDMA 4-node cluster up to 3× inference: apple.com.cn official press release and purchase page (2026-08-26, re-verified 2026-11-12).

512GB China markup ¥116,250, 512GB+1TB system ¥172,999, 16TB max ¥229,249, only 36C/80C eligible, per-tier lead times: ithome.com (2026-10-29), kuaikeji.cn (2026-10-29), apple.com.cn configurator pricing.

US $22,299 base ($6,799 + $15,500 memory), 16TB $29,799, late Nov–early Dec ship, M3 Ultra 2025 96→512GB only $4,000 / system $9,499, per-GB $9.62→$37.26 (~3.9×), two 256GB systems $21,598 combined: 9to5Mac / Yahoo Tech / Channel Insider (2026-10-28 through November).

Memory crisis, earnings call confirms Mac/iPad supply tight through end of 2026, 512GB 10–12 week lead: AppleInsider (2026-11-03), ithome.com / sina.com.cn rebroadcast.

768GB internal test, supply constraints may block launch, M6 high-end cancelled, M7 Ultra 2028 with 1.5TB design target: Bloomberg Mark Gurman Power On (2026-06-28 / 07-12 / 08-25), ithome.com, 163.com, MacRumors rebroadcast; all are leaks, not official Apple announcements .

GPU default ~75% wired (512GB≈384GB), iogpu.wired_limit_mb tunable to ~472GB, non-persistent, reserve 32–64GB + 20% headroom, GLM-5.3-Flash 177.6GB OOM at 128K, MacStories & PriceZen / Pinggy community test aggregates (2026-10 to 11).

Model weight sizes (Qwen3.8-27B 16.1GB, gpt-oss-120b ~65GB, Flash-Next 4-bit 106–111GB, DeepSeek-V4-Flash 151GB, GLM-5.3-Flash 177.6GB, MiniMax M3 265GB, Mistral Large 4 2-bit ~380GB, GLM-5.3 4-bit 418GB, Kimi K3 1-bit ~610GB / native 1.56TB, DeepSeek V4 Pro 4-bit 837GB): Hugging Face repos and Fello AI / PriceZen / Pinggy / LLM Configurator aggregates (2026-10).

⚠️ Scope reminder : Model sizes above are aggregated from repository download sizes and community tests, not a unified benchmark ; quantization levels cause large variance. "Fits" means weights can reside; KV cache and context overhead are not included — long contexts need extra headroom. 2-bit/1-bit are aggressive, lossy compression. Four-node 14.7 tok/s is a single community run, not a comparable benchmark. Lead times and prices per apple.com.cn order page at time of purchase; 768GB and 1.5TB are leaks/design targets, never officially announced by Apple — do not treat them as shipping configurations.