Mac Studio M3 Ultra · 96 GB
819 GB/s of unified-memory bandwidth in a silent 3.6 kg box — the fastest desk-tier memory in this catalogue, buyable today.
Price
£5,299 inc VAT
Apple RRP since 25 Jun 2026 price rise (was £4,199) · 13–14 week lead times reported
verified 2026-07 · supply & lease options in every proposal
Apple- 96 GB
- Unified memory
- 819 GB/s
- Memory bandwidth
- 32-core
- Neural Engine
- ~300 W
- Typical inference
512 GB config discontinued Mar 2026
3× a DGX Spark
M3 Ultra, 28-core CPU / 60-core GPU base
silent — 480 W max system
The M3 Ultra's party trick is bandwidth: 819 GB/s to the whole unified pool — three times a DGX Spark — which is why 70B-class dense models that crawl on other unified-memory boxes run at ~20 tok/s here, silently, from a machine that idles in single-digit watts.
The honest constraint is the ceiling: after Apple withdrew the 256 GB and 512 GB options during the 2026 memory shortage, new orders cap at 96 GB — enough for gpt-oss-120b, Qwen3.5 122B and Devstral 2 at 4-bit, not for the 400 GB-class frontier weights. The M5 Ultra refresh expected late 2026 (tested to 768 GB) reopens the big-model story; until then, Thunderbolt 5 RDMA clustering is the scale-out path.
Prompt processing remains the platform weakness — MLX narrows the gap but long-context ingestion trails NVIDIA silicon. Right machine for chat, drafting and agent loops; wrong machine for bulk 100k-token document runs.
vs its siblings: The bandwidth outlier of the desk tier — and the only silent machine in the catalogue.
Memory, to scale
96 GB model-visible · bandwidth is the speed limit
Unified memory
96 GB · 819 GB/s
LPDDR5x
For scale
DGX Spark — 128 GB @ 273 GB/s
RTX PRO 6000 — 96 GB @ 1.79 TB/s
DGX Station GB300 — 748 GB coherent
macOS 26.2 RDMA over Thunderbolt 5 + MLX distributed/exo — clusters of Studios pool memory for trillion-parameter models.
What it actually runs
Declared from research and benchmarks, not computed marketing — tokens-per-second figures are cited where a real measurement exists.
- gpt-oss-120bMXFP4with headroom~60 tok/s via MLX (M3 Ultra, verified)
- Qwen3.5 122BQ4with headroom10B active MoE — strong fit for the bandwidth
- Devstral 2Q4with headroomdense 123B coding model — runs properly at 819 GB/s
- Llama-70B-class denseQ4with headroom~19–20 tok/s (bandwidth-derived estimate)
- GLM-5.2 / frontier class4-bit2+ units linkedneeds the discontinued 512 GB config or a TB5 cluster
The full sheet
Compute
- Chip
- Apple M3 Ultra — 28-core CPU (20P+8E), 60-core GPU base; 32c/80c upgrade
- Neural Engine
- 32-core
Memory
- Unified memory
- 96 GB @ 819 GB/s (256/512 GB options withdrawn H1 2026)
- Model capacity
- ~70 GB usable for weights — 120B-class MoE at 4-bit
Storage & I/O
- SSD
- 1 TB base, to 16 TB
- Networking
- 10 GbE standard · Wi-Fi 6E
- Ports
- 6× Thunderbolt 5 (120 Gb/s) — the clustering fabric
Software
- Serving
- MLX (performance path) · llama.cpp · LM Studio · Ollama
- Note
- No CUDA — vLLM/TensorRT-LLM ecosystems unavailable
Where it shines
- 819 GB/s — dense 70B models actually run at usable speed
- Silent, tiny, sips power; the easiest machine here to live with
- Thunderbolt 5 RDMA clustering is real (4 nodes ≈ 2 TB pooled)
- Standard business procurement — it's a Mac
The trade-offs
- 96 GB ceiling on new orders — the 512 GB config is gone until M5 Ultra
- Prompt processing far slower than NVIDIA — long-context work suffers
- No CUDA: MLX/llama.cpp only, no vLLM-class production serving
- June 2026 price rise took the base from £4,199 to £5,299
Buy this box for
Understanding Apple Mac Studio
Apple M3 Ultra
Apple silicon's unified-memory architecture gives the GPU the entire RAM pool at 819 GB/s — 3× a DGX Spark's bandwidth. The M3 Ultra's 512 GB configuration became the folk hero of local AI, running DeepSeek R1 671B at 4-bit (~18 tok/s) and even Kimi K2-class trillion-parameter models at ~3.5-bit, silently, under 300 W.
The honest caveat for buyers today: the memory shortage killed the big configurations. Apple withdrew the 512 GB option in March 2026 and the 256 GB option by May — new orders cap at 96 GB (£5,299 since the June 2026 price rise). The 512 GB machines circulate refurbished and second-hand, and the M5 Ultra refresh expected late 2026 has been tested to 768 GB. Clustering is the workaround that works now: macOS 26.2 added RDMA over Thunderbolt 5, and four M3 Ultras pool ~2 TB to serve trillion-parameter models at ~25 tok/s.
The other trade-off is prompt processing: no CUDA, so you serve through MLX or llama.cpp, and long-context ingestion is far slower than NVIDIA silicon even though generation speed is strong. Right machine for chat and agent workloads; wrong machine for 100k-token document dumps.
Sources & verification
Specifications and prices verified 2026-07 against the sources below. The memory shortage is repricing this market monthly — we re-verify at quote.
Compare against the rest of the catalogue, or have us spec it against your workloads.
Back to catalogueSpec the Mac Studio M3 Ultra · 96 GB for your workload
We'll size it against your real token volumes, quote supply and lease options, and stand up the models with evals that prove they do the job — or tell you honestly that a different box (or the API) wins.