Buy a box, or rent one?

14 reference machines people actually buy to run open models, against the cheapest VPS in the index with the same usable memory and a disk included. Local hardware wins on speed (GPU-class bandwidth); a VPS wins on upfront cost and uptime. Prices were last checked on 7 Sept 2026 and refresh weekly (September 2026 edition).

Reference machines for running open models locally. Memory, bandwidth and power come from vendor spec pages. Prices are refreshed weekly by scripts/update-hardware.ts: Apple prices are read from the Apple Store page (list price before tax; BTO memory upgrades add a curated delta, marked 'derived'); vendors that block automated fetches are checked through a web-search lookup and marked 'approximate'. Those follow street prices, which in 2026 sit far above MSRP because of memory shortages. Prices are editable on the page, so treat them as starting points.

MachinePriceMemory · usableBandwidthUp to 4B7–9B12–15B24–32B70B+Mixture-of-experts= VPS months
Raspberry Pi 5 (16 GB)

Raspberry Pi · sbc · ~8 W

$305street, approx. · 7 Sept 202616 GB · 14.5 GB17 GB/s19 moHOSTKEY 16 GB · $16.15/mo
GeForce RTX 5060 Ti 16 GB (card only)

NVIDIA · card only · ~180 W

$800street, approx. · 7 Sept 202616 GB · 15.5 GB448 GB/s50 moHOSTKEY 16 GB · $16.15/mo
Mac mini (M6, 16 GB)

Apple · desktop · ~35 W

$899vendor list price · 7 Sept 202616 GB · 12 GB153 GB/s77 moHOSTKEY 12 GB · $11.70/mo
GeForce RTX 5070 Ti 16 GB (card only)

NVIDIA · card only · ~300 W

$1,200street, approx. · 7 Sept 202616 GB · 15.5 GB896 GB/s74 moHOSTKEY 16 GB · $16.15/mo
Mac mini (M6, 32 GB)

Apple · desktop · ~35 W

$1,299vendor base + upgrade · 7 Sept 202632 GB · 24 GB153 GB/s48 moOVHcloud 24 GB · $27.11/mo
GeForce RTX 5080 16 GB (card only)

NVIDIA · card only · ~360 W

$1,600street, approx. · 7 Sept 202616 GB · 15.5 GB960 GB/s99 moHOSTKEY 16 GB · $16.15/mo
Framework Desktop (Ryzen AI Max+ 395, 64 GB)

Framework · desktop · ~120 W

$1,659street, approx. · 7 Sept 202664 GB · 48 GB256 GB/s36 monetcup 64 GB · $46.49/mo
Mac mini (M5 Pro, 24 GB)

Apple · desktop · ~50 W

$1,699vendor list price · 7 Sept 202624 GB · 18 GB307 GB/s63 moOVHcloud 24 GB · $27.11/mo
Mac mini (M5 Pro, 64 GB)

Apple · desktop · ~55 W

$2,299vendor base + upgrade · 7 Sept 202664 GB · 48 GB307 GB/s50 monetcup 64 GB · $46.49/mo
Mac Studio (M5 Max, 36 GB)

Apple · desktop · ~70 W

$2,499vendor list price · 7 Sept 202636 GB · 27 GB460 GB/s80 monetcup 32 GB · $31.43/mo
Framework Desktop (Ryzen AI Max+ 395, 128 GB)

Framework · desktop · ~120 W

$3,149street, approx. · 7 Sept 2026128 GB · 96 GB256 GB/s28 moContabo 96 GB · $114.24/mochecked 10 days ago
GeForce RTX 5090 32 GB (card only)

NVIDIA · card only · ~575 W

$4,300street, approx. · 7 Sept 202632 GB · 31.5 GB1792 GB/s10+ yrsnetcup 32 GB · $31.43/mo
NVIDIA DGX Spark (128 GB)

NVIDIA · desktop · ~240 W

$4,699street, approx. · 7 Sept 2026128 GB · 96 GB273 GB/s41 moContabo 96 GB · $114.24/mochecked 10 days ago
Mac Studio (M5 Ultra, 96 GB)

Apple · desktop · ~110 W

$5,499vendor list price · 7 Sept 202696 GB · 72 GB1200 GB/s44 moCherry Servers 72 GB · $125.78/mo
every model in the class fits some fit– none fit · “= VPS months”: hardware price ÷ the cheapest VPS with at least the machine's usable memory and a disk included

Usable memory

Unified-memory machines (Apple, Strix Halo, DGX Spark) give the GPU about 75% of RAM by default; discrete cards lose ~0.5 GB to the driver; CPU-only boards keep 1.5 GB for the OS. Weights + KV cache must fit in what is left.

Why bandwidth decides speed

Generating a token reads every active weight once. An RTX 5090 at 1.8 TB/s streams an 8B Q4 model ~100× faster than a 4-vCPU VPS; a Mac mini sits in between. Per-model bands are on each model page.

2026 caveat

Memory-driven price rises have pushed GPUs 40–135% above MSRP and doubled the Raspberry Pi 5. Every model page has an editable calculator, so you can plug in the price you can actually get.