Decentralized cloud marketplace on Cosmos where providers bid on Docker workloads (CPU + GPU) defined in SDL. Tenants create deployments, accept the cheapest matching bid, and pay providers in AKT escrow — useful for cheap GPU inference, training jobs, and self-hosted agent runtimes.
- 01decentralized GPU compute
- 02Docker-native AI inference workloads
- 03self-hosted LLM endpoints
- 04Cosmos-native deployments
- 05permissionless cloud hosting
- curl -sSfL https://raw.githubusercontent.com/akash-network/provider/main/install.sh | sh
- brew install akash
| Variable | Scope | Description |
|---|---|---|
| AKASH_KEY_NAME | Server | Local keyring entry holding the tenant wallet (created via `akash keys add`). |
| AKASH_NODE | Server | RPC endpoint for the Akash chain, e.g. https://rpc.akash.network:443. |
| AKASH_CHAIN_ID | Server | Chain id, typically `akashnet-2` for mainnet. |
| AKASH_KEYRING_BACKEND | Server | Keyring backend (`os`, `file`, or `test`). Use `file` or a HW wallet in production. |
Use Akash to deploy a workload by writing an SDL manifest (`deploy.yaml`) declaring `services` (Docker images + ports), `profiles.compute` (CPU/RAM/storage and optional `gpu: { units, attributes: { vendor: { nvidia: [{ model: a100 }] } } }`), `profiles.placement` (region/audited-by filters and pricing), and a `deployment` mapping. Submit with `akash tx deployment create deploy.yaml --from $AKASH_KEY_NAME`, wait for bids via `akash query market bid list --owner <addr>`, accept one with `akash tx market lease create`, then send the manifest to the provider with `akash provider send-manifest`. Read service URIs back via `akash provider lease-status`. For an agent stack, package the model server (e.g. vLLM, ollama) as a container, expose an HTTP port, and put a Cloudflare/Caddy front-end if you need TLS.
- ⚑GPU availability is bursty — specific models (H100, A100-80G) often have zero bids; fall back to a model list rather than pinning one SKU.
- ⚑Provider regions are self-declared; if data residency matters, filter `placement.attributes` by audited providers and verify location independently.
- ⚑AKT price volatility plus bidding means deploy cost can swing 20%+ within a day; quote tenants in USD and re-price escrow tops-up daily.
- ⚑Leases auto-close when escrow funds run out — fund the deployment generously or run a watcher that tops up via `tx deployment deposit`.
- ⚑Providers can shut down or upgrade with little notice; treat any single lease as ephemeral and design for redeploy on bid loss.
- ⚑Outbound network from a deployment is unrestricted but ingress requires `expose.global: true` plus a leased hostname; localhost-style services won't be reachable.
- ⚑Image inference on Akash works the same as text — but you must ship the model weights inside the image or pull them at start (cold-start can be tens of minutes for large weights).