Turn spare rigs
into a compute business.
Connect your GPUs, join or create a farm, run real workloads, and earn for verified compute.
Plug in a GPU. Put it to work. Compute runs offchain; escrow and receipts settle on Robinhood Chain.
- Benchmark-verified GPUs
- Sandboxed containers
- Escrow on Robinhood Chain
GPUs you can rent right now
Benchmark-verified, priced by their hosts, with a heartbeat in the last 45 seconds. Nothing here is a mockup.
Real GPU workloads, containerized
Every job is a container with a GPU count, minimum VRAM, max runtime and an output path. Hosts choose which workload types they accept.
- AI inferenceBatch or serve open models on 24–80 GB cards. Pay per GPU-second, not per reserved month.No capacity online yet
- RenderingBlender and path-tracing frames fan out across farms of RTX cards in one region.No capacity online yet
- VideoGeneration, upscaling and transcoding with outputs hashed and delivered by signed URL.No capacity online yet
- Fine-tuningLoRA and full fine-tunes on up to 16 GPUs inside one physical machine.No capacity online yet
- Custom computeAny containerized CUDA workload, sandboxed with no network unless you allowlist it.No capacity online yet
Farms compete in the open
No opaque “best farm” score. Compare price, uptime, completion rate, throughput and start latency, measured from real job history — and sort by what matters to you.
From idle rig to paid compute in five steps
- 1Connect GPUs
Install the host daemon, pair it to your wallet with a one-time code, and it detects every CUDA GPU over NVML.
- 2Create Farm
Group your machines into a public farm with a region, workload types and a price per GPU class.
- 3Accept jobs
The scheduler offers compatible jobs with a 10-second reservation. Your policy decides temperature caps, hours and minimum rates.
- 4Run compute
Workloads run in a locked-down container on the reserved GPU UUIDs. Logs, metrics and the result hash stream back.
- 5Get paid
Verified completion settles from escrow on Robinhood Chain. Earnings accrue by billable GPU-seconds; you claim them any time.
The chain settles. It doesn't compute.
Robinhood Chain is GPUFarm's settlement and trust layer. Payments, receipts and payouts are onchain and verifiable; the GPU work happens offchain where it belongs.
- JobContainer, GPUs, max runtime
- EscrowMax payment locked onchain
- Verified resultHash + telemetry checked
- ReceiptGPU-seconds, hashes, amount
- PayoutHost, farm and fee credited
- Job escrow: the customer's maximum payment, locked before any GPU is reserved
- Compute receipt: job hash, farm, workload, start/end, GPU-seconds, result hash, amount
- Payment release, refunds, disputes and their resolution
- Payouts credited per job to hosts, farm owners and the protocol fee
- Farm identity and host-key authorization (minimal registry)
- The compute itself — containers on host GPUs, never onchain
- Scheduling, heartbeats, telemetry, benchmarks and logs
- Input and output artifacts (signed URLs, per-job scope, automatic expiry)
- Verification: result hashing, telemetry checks, deterministic re-computation where possible
Never “verified” across the board.
Every completed job carries exactly one of three labels, and the receipt says which. The server can't release escrow on its own: release needs a verifier signature and a challenge window the customer can dispute, or the customer's own approval.
No token required — pay in ETH or USDG.
- Result verified
Output was deterministically checked: an independent re-run or verifier produced the same result hash.
- Execution verified
Telemetry from the assigned GPUs and the integrity of the uploaded result were confirmed. The output itself was not re-computed.
- Host reported
The host reports completion, but the workload is not independently reproducible and telemetry was insufficient to confirm execution.
Install the GPUFarm Host daemon
A Rust daemon for Linux machines with NVIDIA GPUs. It never runs in the browser and keeps its signing key in the OS keystore (or an encrypted file on headless servers) — no workload can reach it.
- Detects GPUs with nvidia-smi, cross-checked against NVML: UUID, VRAM, driver, CUDA, temperature, power
- Runs a server-issued benchmark challenge before any GPU is listed
- Executes jobs in Docker with the NVIDIA toolkit (gVisor where supported): no network, no privileges, no host mounts
- Heartbeat every 10 s; offers arrive over WebSocket or polling
curl … | sh installer. Every release binary ships with a SHA-256 and an Ed25519 signature over gpufarm-host <version> <sha256>; the installer, the self-updater and the steps here all check both. Signed releases aren't hosted on this deployment yet, so build from source for now.VERSION=0.1.0; TARGET=x86_64-unknown-linux-gnu
sha256sum -c gpufarm-host-$TARGET.sha256
# the signature covers the exact bytes "gpufarm-host <version> <sha256>";
# gpufarm-release.pem is the release key published at /docs/host#verify
printf 'gpufarm-host %s %s' "$VERSION" "$(sha256sum gpufarm-host-$TARGET | cut -d' ' -f1)" > msg
xxd -r -p gpufarm-host-$TARGET.sig > sig.bin
openssl pkeyutl -verify -pubin -inkey gpufarm-release.pem -rawin -in msg -sigfile sig.bin
sudo install -m 0755 gpufarm-host-$TARGET /usr/local/bin/gpufarm-hostcd host && cargo build --release --locked
sudo install -m 0755 target/release/gpufarm-host /usr/local/bin/gpufarm-hostgpufarm-host detect # GPUs + container runtime, sends nothing
gpufarm-host pair --server https://robingpu.farm # prints a code; approve it in the browser
gpufarm-host run # register, benchmark, heartbeat, run jobs
The network gets stronger every time somebody connects another machine.
Bring a rig online, or put the network to work on your next job.