Buy a box, or rent one?

Prices checked

14 machines people buy to run open models, each against the cheapest VPS with as much usable memory, checked .

Every machine, against a VPS

all models

every model in the class fits some fit– none fit

Local machines, the models they fit, and the VPS months their price buys
MachinePriceMemory · usableBandwidthUp to 4B7–9B12–15B24–32B70B+Mixture-of-expertsVPS months
Raspberry Pi 5 (16 GB)~8 W$305approx.16 GB · 14.5 GB17 GB/s––34 moLeaseweb · $9.04/mo
GeForce RTX 5060 Ti 16 GB (card only)~180 W$456approx.16 GB · 15.5 GB448 GB/s––50 moLeaseweb · $9.04/mo
Mac mini (M6, 16 GB)~35 W$899vendor price16 GB · 12 GB153 GB/s–––99 moLeaseweb · $9.04/mo
GeForce RTX 5070 Ti 16 GB (card only)~300 W$1,190approx.16 GB · 15.5 GB896 GB/s––10+ yrsLeaseweb · $9.04/mo
Mac mini (M6, 32 GB)~35 W$1,299base + upgrade32 GB · 24 GB153 GB/s–92 moLeaseweb · $14.07/mo
GeForce RTX 5080 16 GB (card only)~360 W$1,650approx.16 GB · 15.5 GB960 GB/s––10+ yrsLeaseweb · $9.04/mo
Mac mini (M5 Pro, 24 GB)~50 W$1,699vendor price24 GB · 18 GB307 GB/s–10+ yrsLeaseweb · $14.07/mo
Framework Desktop (Ryzen AI Max+ 395, 64 GB)~120 W$1,959approx.64 GB · 48 GB256 GB/s75 moLeaseweb · $26.13/mo
Mac mini (M5 Pro, 64 GB)~55 W$2,299base + upgrade64 GB · 48 GB307 GB/s88 moLeaseweb · $26.13/mo
Mac Studio (M5 Max, 36 GB)~70 W$2,499vendor price36 GB · 27 GB460 GB/s–96 moLeaseweb · $26.13/mo
GeForce RTX 5090 32 GB (card only)~575 W$3,440approx.32 GB · 31.5 GB1792 GB/s–10+ yrsLeaseweb · $26.13/mo
Framework Desktop (Ryzen AI Max+ 395, 128 GB)~120 W$3,449approx.128 GB · 96 GB256 GB/s53 moLeaseweb · $65.69/mo
NVIDIA DGX Spark (128 GB)~240 W$4,699approx.128 GB · 96 GB273 GB/s72 moLeaseweb · $65.69/mo
Mac Studio (M5 Ultra, 96 GB)~110 W$5,499vendor price96 GB · 72 GB1200 GB/s84 moLeaseweb · $65.69/mo
How we count

Local hardware wins on speed, with GPU-class memory bandwidth; a VPS wins on upfront cost and uptime. “VPS months” is the machine's price divided by the cheapest VPS with at least its usable memory and a disk included.

Memory, bandwidth and power come from the vendors' spec pages. Hardware prices were last checked 26 Oct 2026 and refresh weekly. Apple prices are read from the Apple Store before tax, with a set amount added for memory upgrades (marked base + upgrade); vendors that block automated reads are looked up by search and marked approximate. Those follow street prices, which in 2026 sit far above the makers' own because of memory shortages.

Usable memory: unified-memory machines (Apple, Strix Halo, DGX Spark) give the GPU about 75% of RAM by default; discrete cards lose about 0.5 GB to the driver; CPU-only boards keep 1.5 GB for the OS. Weights and 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 about 100 times faster than a 4-vCPU VPS; a Mac mini sits in between. The speed for each model is on its model page.

Memory-driven price rises have pushed GPUs 40 to 135% above their makers' prices and doubled the Raspberry Pi 5. Every model page has a calculator, so you can plug in the price you can actually get.