| Product | Public description and pricing model | Source |
|---|---|---|
| Proxies.sx Mac Compute | A whole Apple Silicon Mac rented to one customer for 30 days, serving a catalog model under MLX behind /chat/completions. Admin-controlled 30-day defaults as of 11 September 2026: Starter $249 (24 GB), Pro $449 (48 GB), Max $799 (96 GB). Individual prices can vary. No per-token billing; request limits apply. No network token. | this site |
| Nosana | "An open-source GPU cloud built for AI and high-performance workloads"; per-GPU hourly USD pricing, e.g. NVIDIA H100 listed at $1.500/h and RTX 4090 at $0.320/h; the NOS token powers the network | as published on nosana.com |
| io.net | "The Open Source AI Infrastructure Platform"; its homepage did not state GPU prices when we checked, so we quote none | as published on io.net |
| Akash | A marketplace for provider compute capacity using competitive bids. Check current GPU offers and deployment terms on its site. | as published on akash.network |
| Whole machine, one tenant | Priced in | Unlimited tokens for the term | OpenAI-compatible endpoint | Network token | |
|---|---|---|---|---|---|
| Proxies.sx Mac Compute | Yes - one Mac, one customer, 30 days | USD, flat per 30 days | Yes - no per-token meter | Yes - /chat/completions, streaming | None |
| Nosana | GPU jobs; check the chosen host and deployment isolation | Per GPU-hour | Your own serving stack | Your own serving stack | NOS |
| io.net | GPU capacity; check deployment isolation | Not stated on homepage | Your own serving stack | Your own serving stack | IO |
| Akash | Provider container leases; check isolation terms | Per hour, reverse auction | Your own serving stack | Your own serving stack | AKT |
| Your workload | Better fit | Why |
|---|---|---|
| One open model, steady traffic all month, agent or app back-end | Mac Compute | Flat price, no token meter, managed serving on ready marketplace stock; check availability first. |
| Bursty jobs, a few hours at a time | GPU-hour network | Hourly billing stops when the job stops; a 30-day Mac would sit idle. |
| Fine-tuning or training | GPU-hour network | Mac Compute is inference only. |
| Custom serving stack, custom weights, CUDA kernels | GPU-hour network | You get a container there; on Mac Compute you get a managed endpoint with pinned catalog models. |
| Private endpoint, one tenant, no shared hardware | Mac Compute | The machine is exclusively yours for the term; suppliers still control their Macs; relay retention applies. |
| Paying in fiat or from an existing prepaid balance, no wallet | Mac Compute | USD from the Proxies.sx balance (card or crypto top-up); no token to acquire. |
| Models outside the current three-entry catalog | A provider supporting your model | Mac Compute currently offers only the pinned Qwen3.8 27B catalog. There is no Ultra tier or cluster service. |
io.net's public homepage did not show GPU prices when we fetched it, and Nosana's and Akash's per-hour rates move with supply, so treat their figures as a snapshot from 2026-09-11. We do not claim any of them requires holding its token to buy; check each project's docs for current settlement rules. If you find a figure on this page that is stale, email maya@proxies.sx and we will correct it.
It is a different unit. A Max tier Mac is $799 for 30 days of unlimited use; an H100 at the hourly rates Nosana ($1.50/h) and Akash ($1.33/h) publish comes to roughly $960-$1,080 for a 720-hour month, about 1.2-1.4x the Mac, but an H100 is also far faster per token and can train. Compare on the workload: steady inference on one open model favours the flat Mac price; bursty or training work favours GPU-hours.
Each of those networks has its own token (NOS, IO, AKT); how much of the purchase must be settled in it varies by network and changes over time, so check their docs. Proxies.sx Mac Compute has no token: you pay USD from your account balance.
No. You rent a managed inference endpoint, not a shell. The node serves a catalog model under MLX and exposes chat/completions. If you need arbitrary containers or CUDA, use a GPU marketplace.