New Mac Studio with M5 Max and M5 Ultra

Original Article ↗ Hacker News Discussion ↗

Apple’s new Mac Studio with M5 Max and M5 Ultra chips is being received as a powerful local AI workstation, thanks to up to 512GB of unified memory and very high memory bandwidth that make it attractive for running large language models on-device. At the same time, many are alarmed by the pricing, especially for high-RAM configurations in the midst of a global “RAMpocalypse,” and debate whether buying such hardware beats renting GPUs or using cloud AI services. The launch is also reigniting questions about whether to favor desktops over laptops, how much future-proofing unified but non-upgradable memory really offers, and where Apple fits relative to NVIDIA-based systems for serious AI workloads.

Pricing & Regional Differences

  • Many commenters focus on high prices, especially in Europe where a well-specced M5 Ultra configuration exceeds €6,600.
  • US vs EU comparisons highlight VAT inclusion in EU pricing and varying US sales tax; some consider buying in low‑tax or tariff‑free regions (with reminders that technically import taxes still apply).
  • Several people compare current prices to historically expensive PCs, noting that the “computer you really want” still costs around a few thousand dollars, but high‑end AI configs now far exceed that.

RAM Capacity, Shortages, and Options

  • RAM pricing is called “insane”: +$4,000 for 256 GB unified memory, with 512 GB expected to be near or above $20k.
  • Multiple posts reference a broader “RAMpocalypse”: DRAM capacity is reportedly sold out through 2027 and big buyers (including Apple) may have under‑reserved.
  • Many users wish for a 1 TB unified memory option; some say they’d pay $20–30k because that changes which models can run locally (e.g., ~1T+ parameter 4‑bit models).

Performance, Unified Memory & LLM Workloads

  • Unified memory is praised for letting CPU and GPU share a large, fast pool, ideal for local LLMs and large-context workloads.
  • M5 Ultra’s 1.2 TB/s bandwidth is repeatedly highlighted; some compare it favorably to consumer GPUs on bandwidth, but others argue Nvidia workstation GPUs will still dominate raw compute.
  • There’s debate over energy efficiency and tokens‑per‑watt: some claim Macs win; others say optimized Nvidia rigs can match or beat them.
  • Bandwidth vs latency tradeoffs are discussed; high bandwidth helps LLM prefill, but pointer‑heavy workloads rely more on latency and caches.

Desktop vs Laptop / Thin‑Client Setups

  • Many are reconsidering a Mac Studio (or Mini) plus a lightweight laptop or iPad, remoting in via Tailscale, SSH, or Apple Screen Sharing.
  • Several describe using Studios/Mac Minis as always‑on home servers for LLMs and dev, with laptops (or Neo/iPad) as thin clients.
  • Thermal throttling and fan noise on MacBook Pros under sustained LLM workloads push some toward desktops.

Cloud vs Local AI Economics

  • Some argue renting GPUs or using API providers is still cheaper for experimentation and bursty workloads; local hardware sits idle most of the time.
  • Others with heavy, continuous or privacy‑sensitive workloads report cloud AI bills in the five figures per month and justify expensive local hardware as “prepaying” for years of compute.

Expandability, Form Factor & Future‑Proofing

  • Lack of user‑replaceable RAM/SSD and absence of a modular tower Mac draw criticism; some want a reusable chassis with swappable mainboards.
  • Others counter that SoC + on‑package memory is what enables these bandwidth numbers, making traditional upgradability technically and economically difficult.
  • Thunderbolt 5 clustering of multiple Studios is discussed; Apple claims speedups for up to four nodes, but bandwidth and RDMA limitations are noted.

Other Technical Notes & Reactions

  • PCIe Gen 6 SSD architecture with up to 2× storage speed intrigues some; questions remain about thermals and real‑world benefit.
  • M6’s quiet launch (Mini only, 32 GB cap, no Pro/Max/Ultra) suggests M7 will be the bigger AI‑focused step.
  • Overall sentiment mixes excitement about serious local‑AI capabilities with frustration over pricing, RAM caps, and long‑term OS/hardware lock‑in.