Mac Studio — M5 Max
$2,49936 GB unified · 512 GB SSD · 18-core CPU · 32-core GPU
Hardware · Kigali
One Mac Studio. One desk.
A workstation in Kigali that runs the models locally — a Mac Studio M5 Max in its default configuration, a Studio Display, fibre with a failover line, and a UPS that keeps it all up.
- One-time
- $5,276
- Monthly
- $93
- On battery
- 73 min
What's on the desk
Five pieces, all shipping today. Apple US list prices — add freight and duty.
Studio Display (2026) — 27″ 5K
$1,599600 nits · P3 · 12 MP Center Stage · Thunderbolt 5 in, daisy-chain out
Magic Keyboard with Touch ID
$149The Mac Studio ships without one. Touch ID unlocks and signs in
Magic Trackpad
$129Force Touch, USB-C charging
APC Back-UPS Pro 1500
$300BR1500MI, the 230 V model · 1500 VA / 865 W, line-interactive AVR. Rides through brownouts and covers an outage long enough to sleep cleanly
The one box
The default configuration. Nothing upgraded, nothing missing.
- Chip
- Apple M5 Max
- CPU
- 18-core (6 super + 12 performance)
- GPU
- 32-core, Neural Accelerators
- Neural Engine
- 16-core
- Memory
- 36 GB unified · 460 GB/s
- Storage
- 512 GB SSD
- Rear
- 4× Thunderbolt 5 · 10 Gb Ethernet · 2× USB-A · HDMI
- Front
- 2× USB-C · SDXC
- Wireless
- Wi-Fi 7 · Bluetooth 6 (N1)
- Software
- macOS 27
The site
Kigali has what the Erkowit site lacked: a grid, hands nearby, and a licensed link.
- City
- Kigali, Rwanda
- Elevation
- ~1,567 m
- Ambient
- ~20–27 °C — no cooling load
- Grid
- Stable · short outages
- Link
- Fibre + LTE failover
- Staffing
- On the desk · reachable remotely
The live cost model
Switch line items off, or add the memory upgrade, and watch the one-time cost, the monthly bill and the battery runtime move. The 128 GB upgrade adds $2,600 — and can only be bought at checkout.
The box
$2,499Mac Studio — M5 Max
36 GB unified · 460 GB/s · ~90 W load · 9 W idle · 0.86 kWh/day
The desk
$2,277The link
$200Power
$300What to run on it
Frontier weights are closed — no hardware buys them. What runs is the open-weight tier. The rule: MoE with low active parameters, never dense — and in 36 GB, the 30B class is the comfortable ceiling.
Llama 3.1 8B 4-bit8B dense · ~5 GB · fastTrivial for the box. Good for classification and enrichment
gpt-oss-20b~4B active (MoE) · ~13 GBGeneral reasoning that leaves most of the 36 GB free
Qwen3-Coder-30B-A3B 4-bit3B active (MoE) · ~17 GB weightsThe target. Fits with room for context — low active params is what 460 GB/s wants
gpt-oss-120b / 80B-class MoE~45–65 GB weights · needs the 128 GB boxDoes not fit in 36 GB. Only with the upgrade, decided at checkout
Dense Llama 3.1 70B70B dense · ~40 GB 4-bitNever. Too big here, and bandwidth-bound on any Mac
macOS and MLX, honestly
The box serves models through Apple's MLX (mlx-lm, Ollama's MLX engine) — unified memory means the whole model sits within the GPU's reach with no copies. The price: no CUDA, no vLLM, and mlx_lm.server keeps a per-request KV cache, so the box is excellent for one person or a serial queue, not dozens of concurrent requests. That matches our size today.
The roads not taken
Three reasonable alternatives, each rejected or deferred for a specific reason rather than a preference. Figures checked October 2026.
NVIDIA DGX Spark
rejected128 GB
for ~$4,000, against 36 GB here
273 GB/s vs 460 GB/s
CUDA · vLLM · DGX OS (Ubuntu)
Headless — no desk, no display workflow
Holds bigger models, decodes slower, and is a second headless machine beside the desk. If 80B-class models become the need, the +$2,600 upgrade keeps it in one box.
M5 Ultra now
watch+$3,000
for 2.6× bandwidth and 96 GB
96 GB · 1.2 TB/s · from $5,499
Up to 512 GB (slipping to 2027)
480 W max continuous rating
Real speed, bought before there is a queue to speed up. It is Tier 4 — revisit when measured saturation says the Max is the bottleneck.
Studio Display XDR
rejected$3,299
against $1,599 for the Studio Display
Mini-LED · HDR · 120 Hz
Same 27″ 5K canvas
Built for colour grading
Code, docs and dashboards gain nothing from HDR. The $1,700 difference is a fibre line for two years.
Buy the class, never the SKU
Apple refreshes its chips roughly yearly, and each generation raises memory bandwidth — precisely what caps decode. Our Tier 2 trigger may not fire before the M6 generation — so the spec reads “Mac Studio, Max-class chip, default configuration of the generation current at trigger time,” never a part number. Written that way, buying the box last is not only cash discipline; the plan upgrades itself while it waits.
When the power or the link drops
Kigali's grid is steady, not perfect, and the box gets used remotely whenever nobody is at the desk. So recovery is built out of simple, reliable parts.
UPS with AVR
Brownouts never reach the box; a real outage gets minutes to finish the job and sleep cleanly
Start up after a power failure
One macOS Energy setting, so every long outage self-heals without a human
LTE failover on a second carrier
The router cuts over in seconds, and remote login (Tailscale, Screen Sharing) keeps working through it
Backups and an idempotent queue
Time Machine hourly, an off-site copy to S3, and every job safe to re-run after a cut
Tiers and triggers
Note the shape: the link comes first because it is cheap and protects delivery today, and the box waits on revenue. Running models locally then costs nothing extra — it is software on a machine we already bought.
Today
$0Laptops the team already owns, plus cloud.
Trigger: Current state
The link
now~$500 + ~$80/moBusiness fibre, an LTE failover on a second carrier, and a UPS for the router. The one piece every later tier stands on.
Trigger: None — affordable now, and it protects delivery today
The workstation
~$4,800Mac Studio M5 Max in its default configuration, Studio Display, Magic Keyboard and Trackpad, Time Machine SSD. Hermes, the CRM mirror, git and staging move onto the box.
Trigger: First paying school, or non-dilutive funding
Local models live
$0 — softwareMLX serving on the box: enrichment, classification and a 30B-class coder run locally in 36 GB; the cloud keeps the frontier work.
Trigger: 3 paying schools / ~$3K MRR sustained 3 months
Scale the box
~$5,100–9,500Memory cannot be added to a Mac later, so scaling means a second box: an M5 Ultra (96–512 GB) or a 128 GB Studio clustered over Thunderbolt 5. The first Studio moves to a teammate's desk.
Trigger: Measured saturation — the queue is waiting on the machine, not on the link
Customer plane migration
IncrementalSelected tenants on owned iron.
Trigger: 12 months of measured uptime, and a customer who wants on-prem and pays for it