Hardware

Every other document here configures software. This one configures atoms: the silicon, the screen, the link and the watts that have to exist before any of the engine runs locally.

Supersedes the Erkowit solar node. The previous version of this document sized an unmanned, off-grid DGX Spark node in eastern Sudan: a 6 kWp array, a 15 kWh battery, VSAT and a caretaker (in git history). Moving to Kigali removes the whole power plant. The grid is steady enough that a UPS replaces it, so the number that matters stops being kWp and becomes dollars.

Status: plan, not a purchase order. Tier 1 is affordable today; everything else is costed so the trigger is a decision rather than a scramble. See Tiers and triggers. The live cost model is at /hardware.

The site

The Erkowit plan was dominated by what the site lacked: grid, staff and a lawful satellite link. Kigali has all three, so the plan shrinks to the machine and the two things it actually needs from the room: clean power for a few minutes, and a link that fails over.

What "run models locally" can mean

Frontier weights are closed, and no hardware buys them. What runs locally is the open-weight tier, which tops out around 71–72 % on SWE-bench against 80–95 % for closed models. The cloud keeps the frontier work; the box takes enrichment, classification, the coder model and anything that should not leave the building.

Run MoE models with low active parameters — never dense 70B

Decode speed is set by memory bandwidth ÷ bytes read per token, and a mixture-of- experts model only reads its active parameters. In 36 GB at 460 GB/s:

In 36 GB the 30B MoE class is the comfortable ceiling. That covers enrichment, classification and a capable coder; the cloud keeps everything bigger.

macOS and MLX, honestly

The box serves through Apple's MLX (mlx-lm, Ollama's MLX engine). Unified memory puts the whole model within the GPU's reach with no host-to-device copies, so 36 GB on a Mac behaves like 36 GB of VRAM, minus what macOS itself holds.

The price: no CUDA and no vLLM. vLLM's PagedAttention is Linux + CUDA only, and mlx_lm.server keeps a per-request KV cache. So the box is excellent for one person or a serial job queue, and the wrong shape for dozens of concurrent requests. That matches our size today; if it stops matching, it shows up as measured saturation and fires Tier 4.

The one box

Mac Studio, M5 Max, Apple's default configuration: $2,499 (US list, October 2026; $2,299 education). The M5 Mac Studio shipped 22 September 2026.

Ports: 4× Thunderbolt 5, 10 Gb Ethernet, 2× USB-A, HDMI and a headphone jack on the back; 2× USB-C and an SDXC slot on the front. Wi-Fi 7 and Bluetooth 6 on Apple's N1 chip. Ships with macOS 27.

Memory is decided at checkout or never. Apple memory is part of the chip package; nothing can be added later. The default box is the cash-first choice and holds the 30B class. If 80B-class models are a known need on the day of purchase, the +$2,600 upgrade is the time to buy them, not a later Tier.

Buy the class, never the SKU

Apple refreshes its chips roughly yearly, and each generation raises memory bandwidth, which is exactly 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.

The display

Studio Display (2026): $1,599. 27″ 5K, 600 nits, P3, 12 MP Center Stage camera with Desk View, six speakers, two Thunderbolt 5 ports + two USB-C, daisy-chainable. The panel is the 2022 one; the new parts are the ports and the camera.

The Mac Studio ships with no keyboard or pointer: Magic Keyboard with Touch ID ($149) + Magic Trackpad ($129).

Connectivity

The fibre monthly is an estimate until a written quote replaces it. Pick the carrier for the LTE line so it does not share the fibre provider's upstream; two links that fail together are one link.

Ingress and bandwidth

The box does not take inbound traffic from the internet. Remote access goes over Tailscale (and macOS Screen Sharing over it); no ports are opened on the router. Model weights are the only large downloads: a one-off 30–65 GB per model, minutes on fibre and something to avoid on the LTE lane.

Power

Kigali's grid replaces the solar plant. What remains is an APC Back-UPS Pro 1500 (BR1500MI, the 230 V model; 1500 VA / 865 W, line-interactive, ~$300): the AVR absorbs brownouts, and the battery gives an outage enough minutes to finish the job and sleep cleanly. The 120 V BR1500MS2 is the US sibling; do not import that one.

UPS runtime

runtime (min) = (battery Wh × efficiency) / peak W × 60
              = (216 × 0.8) / 143 × 60  ≈ 73 min

Peak is the box under load (~90 W) + display (30 W) + fibre ONT/router (12 W) + LTE router (6 W) + SSD (5 W) = 143 W. With the 128 GB upgrade it rises to ~158 W (~66 minutes). Both are far under the UPS's 865 W limit.

Energy and the bill

At $0.25/kWh that is **$13/month**. Power is the smallest monthly line; the link is the largest.

When the power or the link drops

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 and sleep cleanly.
• Start up automatically after a power failure. One macOS Energy setting, so every long outage heals itself without a human.
• LTE failover on a second carrier. The router cuts over in seconds, and Tailscale keeps the remote session alive through it.
• Backups. Time Machine hourly to the Thunderbolt 5 SSD, plus an off-site copy to S3.
• Idempotent queue. A power cut mid-job is normal, not an incident. Every job must be safe to re-run.

Bill of materials

USD, US list as of October 2026. Landed cost in Rwanda adds freight, duty and VAT; get a quote from an Apple Authorised Reseller in Kigali before committing, and compare it with buying in Dubai.

Tiers and triggers

The Drive settles it: cash flow first. A workstation that eats the runway before there is revenue is the failure mode, not the plan.

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.

The roads not taken

NVIDIA DGX Spark: rejected

128 GB unified at ~$4,000–4,700, with CUDA, vLLM and DGX OS (Ubuntu). It was the previous plan's box. Against the default M5 Max it holds 3.5× the memory at under two-thirds the bandwidth (273 vs 460 GB/s): bigger models, slower decode. It would be a second, headless machine beside the desk rather than the desk itself. If 80B-class models become the need, the +$2,600 upgrade at checkout keeps them in one box. Revisit only if a workload appears that genuinely needs CUDA (video generation is the likely one).

M5 Ultra now: watch

+$3,000 over the default Max for 2.6× the bandwidth and 96 GB (from $5,499; up to 512 GB, with that option slipping). 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 for mini-LED, HDR and 120 Hz on the same 27″ 5K canvas. Code, docs and dashboards gain nothing from HDR, and the $1,700 difference pays for the fibre line for two years.

Risks

Open decisions

1. Memory: the default 36 GB, or the +$2,600 128 GB build. It can only be decided at checkout.
2. Fibre provider: get written quotes from CanalBox Business, Liquid and MTN.
3. Where to buy: a Kigali Apple Authorised Reseller vs Dubai (price, warranty, duty).
4. Nano-texture glass: depends on the desk's position relative to windows.
5. Whose desk the box sits on, and who holds physical access.

Sources

The box: Apple Mac Studio technical specifications · Macworld — 2026 Mac Studio M5 Max and M5 Ultra specs and prices · MacRumors — Mac Studio roundup · Engadget — M5 Ultra review · Engadget — why the Mac Studio got expensive · Apple Newsroom — Mac Studio with M5 Max and M5 Ultra · Macworld — M5 Max Mac Studio review (base config, 460 GB/s) · M5 Max 128 GB retail listing — Abt · M5 Ultra vs DGX Spark on local LLMs The display: Apple Newsroom — new Studio Display and Studio Display XDR · 9to5Mac — new Studio Display · EveryMac — Studio Display 2026 specs · Macworld — Studio Display 2026 review · iGeeksBlog — Studio Display XDR Connectivity: Liquid Home fibre in Kigali · More Kigali households get fibre Serving: vLLM GPU installation (Linux + CUDA) · Prefill is compute-bound, decode is memory-bound Open-weight models: Best open-weight coding models to self-host, 2026

See also

• /hardware: product photos and the live cost model. Toggle line items or the memory upgrade, watch capex, the monthly bill and UPS runtime move
• Delivery: the on-prem tier this workstation is the reference build for
• CRM · Hermes: the first two services to move onto it
• Stack: what runs, and where