Fly is retiring its GPUs.
Your app can stay where it is. Your model needs somewhere to go.
The Core Economics
Fly built something genuinely good: hardware-virtualized Machines you can place close to your users, billed by the second, with an API that treats infrastructure as something you script rather than click. For latency-sensitive apps in many regions, that's the right tool and we're not going to pretend otherwise.
What Fly is not is a GPU platform any more. Their own documentation now carries a deprecation notice on every GPU page. If you were running inference on Fly A100s or L40S Machines, that workload needs a new home — and the question is whether you move it to another company's rate card or to an open market where providers compete for it.
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.
| Fly.io | Akash | |
|---|---|---|
| GPUs | Deprecated | H100s, A100s, and L40s, by the hour |
| Price | Fly's published usage rates | Set by competitive bids, per deployment |
| Hardware | Fly's fleet | A global network of competing providers |
| Egress | Metered | Zero |
| Where it runs | Fly's regions | Any provider on the network, or your own hardware |
Architectural Distinctions
Somewhere for the GPU workload to land
A deprecation is a deadline. The work is the same either way — containerize the model server, point your app at it — so the only real decision is what you're moving onto. On Akash, GPU capacity comes from many independent providers rather than one fleet with a roadmap, and no single company's product decision can retire it out from under you again.
Nobody can deprecate an open market
Fly's GPUs went away because Fly decided they should. That's a normal thing for a company to do, and it's the structural risk of building on one. Akash is an open protocol: the marketplace is software anyone can run, settlement happens on the network rather than in a company ledger, and providers join because there's demand. Even if the dashboard went away, the network and your workloads stay live.
Your containers stay standard
If it runs in Docker, it runs on Akash. No proprietary runtime, no rewrite, no SDK you've never heard of. The deployment file is portable YAML that reads like a Docker Compose file, and the same file deploys to any provider on the network, or to your own hardware if you run the provider software.
Service Parity Index
| Feature layer | Fly.io | On Akash | How it works |
|---|---|---|---|
| Compute runtime | Fly Machines | Standard containers | Your image runs directly on provider hardware. |
| GPUs | Deprecated | Requested in the same SDL | Add a GPU block naming vendor and model to any service. |
| Data persistence | Fly Volumes, managed Postgres | Persistent volumes, or an outside managed database | Run it as a container with a volume, or point at a managed service. |
| Configuration | fly.toml | SDL | One portable YAML file covering services, resources, ports, and placement. |
| Deployment control | flyctl, Machines API | CLI, REST API, MCP, agent skill | An AI coding agent with an API key can deploy, inspect, and close workloads. |
| Confidential compute | Not offered | One line in your SDL | tee: cpu or tee: cpu-gpu runs your workload in an AMD SEV-SNP or Intel TDX enclave. |
What We Are. What We Aren't.
Akash doesn't have a centralized SLA, because we aren't a centralized provider. We're a marketplace. You choose your level of reliability by choosing your provider — Verified Enterprise data centers with redundant power and networking, or Community Independent operators optimized for price.
Two things Fly does that we don't. Global edge placement: Fly will run your app in many regions behind anycast routing and put it physically near your users. Akash providers are distributed, but placement is something you request from bids, not an edge network with automatic routing. If sub-50ms response times in a dozen countries are the product, Fly is built for that and we aren't. Scale-to-zero and a Machines API: starting and stopping Machines programmatically is a first-class primitive on Fly. Akash leases run continuously until you close them.
For GPU workloads that need somewhere to go, and for teams who'd rather have providers compete than accept a rate card, that trade is worth making. For a latency-critical edge app, keep it on Fly.
FAQs
Start Building
Move the GPU workload to a market that can't retire it, and keep the rest of your stack wherever it's happiest.