Two open markets for GPUs.
One of them can't take your machine back mid-run.
The Core Economics
Both are open markets. Both beat a hyperscaler on price. The difference is what happens after you've got the GPU.
Vast.ai sells three tiers. On-demand holds a machine for your session. Interruptible is cheaper because the bidding never stops: a higher bidder or the host reclaiming capacity pauses or kills your instance. Reserved trades commitment for a discount.
Akash bids once, at placement. Providers compete for your deployment, you take a bid, and that lease runs until you close it or the escrow empties. Nobody outbids you for a workload that's already serving traffic. The competition happens before your container starts, not while it's running.
Here's what that costs you: nothing you can look up on a rate card. Prices move as providers compete and as more capacity joins the network. See what the market is charging right now on the GPU pricing page.
| Vast.ai | Akash | |
|---|---|---|
| Bidding | Continuous, on the interruptible tier | Once, at placement |
| Preemption | Interruptible can be outbid or reclaimed | None. The lease is yours until you close it |
| What you get | A machine, over SSH | A deployment, described in one file |
| Payment | Prepaid credits | Credit card into refundable escrow, or USDC and AKT |
| Who holds the rules | Vast.ai operates the platform | An open protocol. Anyone can run a provider |
Architectural Distinctions
A lease, not a standing bid
The cheapest capacity on Vast.ai is contingent capacity. That's the right trade for checkpointed batch work and the wrong one for anything with users in front of it. On Akash there's no equivalent exposure: once the lease opens, the provider is committed for as long as it's funded. Market pricing, without holding a position someone can outbid you out of.
A deployment, not a machine
Vast.ai hands you a container on a host with SSH — closer to renting a box. Akash takes a file that declares your services, resources, ports, and placement, then runs it. Endpoints, persistent volumes, and multi-service apps are part of the spec instead of something you configure after logging in. For a training run, either works. For something that has to stay up and be addressable, that's less assembly.
A protocol, not a platform
Vast.ai is a company that runs a marketplace, holds your credits, and sets the rules. Akash is open-source software with an on-chain market: anyone can run a provider without approval, settlement happens on the network rather than in a company ledger, and you can self-host the whole stack. Even if the dashboard went away, the network and your workloads stay live. That's also why hardware-attested confidential compute works here — the guarantee comes from the silicon, not from trusting an intermediary.
Service Parity Index
Both run standard containers. The differences are around them.
| Capability | Vast.ai | On Akash | How it works |
|---|---|---|---|
| Workload unit | Docker instance over SSH | One SDL deployment | Declares every service, its resources, and its exposed ports. |
| Price stability | Fixed on-demand, contingent on interruptible | Fixed for the life of the lease | Your accepted bid is the price until you close or update. |
| Selection | Browse listings, filter by reliability score | Read bids | Bids name the provider, region, and hardware. You choose from them. |
| Confidential compute | Not offered | One line in your SDL | tee: cpu or tee: cpu-gpu runs in an AMD SEV-SNP or Intel TDX enclave. |
| Automation | CLI, Python SDK, REST API | CLI, REST API, MCP, agent skill | An AI coding agent with an API key can deploy, inspect, and close workloads. |
| Payment | Prepaid credits | Refundable escrow per deployment | Unspent balance comes back when you close. |
What We Are. What We Aren't.
Being a market doesn't make two markets interchangeable. Vast.ai is the better call for some work, and pretending otherwise would waste your time.
The lowest possible rate on interruptible capacity
If your job restarts cleanly, preemption is a discount. Take it.
Interactive, exploratory work
SSH and a Jupyter notebook beat a declarative deployment file for poking at things.
Per-host reliability data
Vast.ai publishes a reliability score per host and lets you filter on it. Akash sorts providers into Verified Enterprise and Community Independent tiers, which is coarser.
What Akash is for: workloads that need to keep running at a price you agreed to once, on infrastructure nobody can quietly change the terms of.
FAQs
Start Building
Take a bid, hold the lease, and keep it for as long as the work takes.