For demand
One request puts every source to work — your machines, community hosts, RunPod, and Alibaba Cloud compete on price.
The open compute network
Computing power,without the lock-in.
One open network connecting people who need compute with machines ready to work.
Need computing power?
Access more compute at a competitive price.
Have unused computing power?
Host it and earn from the work it completes.
Active leases
12
Checkpoint
5 / 7
Nodes lost
1
Goodput
90 / 100
State
RESUMED ELSEWHERE
The compute market
Today's GPU prices. Observed, not promised.
Every quote below carries when it was captured and whether it is stale — the same discipline the price-comparison page holds itself to.
| GPU | Price | Trend | Tier | Provenance |
|---|---|---|---|---|
| H100 80GB | USD 2.69/ gpu-hr | NEW | community | stale — observed 58d ago2026-08-12 03:55 UTC |
| A100 80GB | USD 1.39/ gpu-hr | NEW | community | stale — observed 58d ago2026-08-12 03:55 UTC |
| A100 40GB | USD 1.00/ gpu-hr | NEW | community | stale — observed 58d ago2026-08-12 03:55 UTC |
| RTX 5090 | USD 0.69/ gpu-hr | NEW | community | stale — observed 58d ago2026-08-12 03:55 UTC |
| RTX 4090 | USD 0.34/ gpu-hr | NEW | community | stale — observed 58d ago2026-08-12 03:55 UTC |
| RTX 3090 | USD 0.22/ gpu-hr | NEW | community | stale — observed 58d ago2026-08-12 03:55 UTC |
| L40S | USD 0.79/ gpu-hr | NEW | community | stale — observed 58d ago2026-08-12 03:55 UTC |
For supply
Idle machines already cost you. Connect them and earn when they complete useful work.
Early testing uses Zolli credits. Cash payout is not live.
How Zolli works
From a compute need to finished work.
01
Tell Zolli what you need.
Describe the work and what matters—price, speed, or hardware.
02
The network finds suitable machines.
Zolli matches it to owned, community, and rented capacity.
03
Your work continues as capacity changes.
Recorded progress lets supported work continue on another machine.
Host
Every machine answers to flashnode. Shared machines run only allowlisted Docker images, sandboxed from the host.
Runtime
Independent tasks lease across machines. Inside one machine, multi-GPU DDP and FSDP run as PyTorch intends.
Recovery proof
Machines disappear. Progress doesn't.
When capacity vanishes, Zolli resumes supported work from recorded progress on another compatible machine.
- 01RTX 4090 machine destroyed
- 02Resumed on an RTX 3090
- 0358 epochs preserved
01:28
The ledger records the exact protocol events behind the handoff, from heartbeat loss through the accepted replacement commit.
event ledgersample data
A worker claimed a time-limited lease.
LEASE_CLAIMED · +00:04
flashnode · node-7fb2 task-03
The catalog committed a verified checkpoint manifest.
CHECKPOINT_MANIFEST_COMMITTED · +01:01
checkpoint.catalog · step 1400 restore_verified
A machine stopped reporting heartbeats.
NODE_HEARTBEAT_LOST · +01:28
monitor · node-7fb2 silent 45s
The interrupted task returned to the queue.
TASK_REQUEUED · +01:32
coordinator · task-03 attempt 2/3
The coordinator accepted a validated result.
TASK_COMMIT_ACCEPTED · +02:28
coordinator · task-03 accepted
Every protocol name above is a real member of the runtime's event ledger. The displayed values are sample data until a captured run replaces them.
30
production attempts
Recorded across the first two contributing hosts.
2
proven architectures
macOS arm64 and Linux x86_64.
5
steps lost, not 35
Recovered from the last verified checkpoint.
1
accepted result per task
Idempotent commits reject duplicate outcomes.
Platform support
Bring the machines you already use.
macOS Apple siliconLinux x86_64Windows 11Preview
Python 3.11NumPypandasscikit-learnSciPyPyTorch CPUPyTorch CUDA 12.4DockerGitHub
macOS, Linux, and RunPod NVIDIA GPUs connect directly. Windows 11 connects through Docker Desktop and WSL2.
Professional services
Start with the machines and workloads you already have.
Zolli can help determine fit, connect a fleet, and adapt a divisible or checkpointable workload. Engagement scope is agreed directly with Zolli.
01
Architecture and workload assessment
Map the machines, workload boundaries, and recovery requirements before rollout.
02
Machine and GPU fleet onboarding
Bring local workstations, home rigs, and cloud GPU hosts into one operational workspace.
03
Runtime and job-spec integration
Shape existing Python entrypoints and flashml.yaml job definitions around bounded tasks.
04
Private deployment and recovery design
Define deployment boundaries, checkpoint flow, and deterministic recovery for your environment.
FAQ
Evaluate the fit. Know the boundaries.
Need to discuss a specific deployment? Email phongct1105@gmail.com.
01What is Zolli?
Zolli is an allocation network across owned, community, and rented compute. It coordinates supported work across compatible machines.
02Is Zolli another cloud provider?
No. Zolli is not another single cloud provider; it helps work move across capacity from more than one source.
03Can machine owners earn money today?
Early testing uses Zolli credits. Cash payout is not live.
04Will Zolli always be cheaper?
No. Competition creates choices, but Zolli cannot guarantee every job is cheaper.
05Which machines work?
Proven hosts are macOS Apple silicon, Linux x86_64, and tested RunPod NVIDIA GPUs. Windows 11 is in Preview.
06Which workloads fit?
Divisible or checkpointable work fits. Tightly synchronized multi-machine training is not the current target.
07What kind of training doesn't fit?
Zolli is best for work that can be divided or resumed. It is not currently designed for tightly synchronized training where every GPU must communicate continuously over a very fast network.
08What happens if a machine disappears?
Supported work can be reallocated and resumed from recorded progress on another compatible machine.
09How mature is the network?
Zolli is early, with verified cross-machine runs and a growing network.
Zolli Cloud
Join the open compute network.
Access compute from more sources, or make a machine available to the network. Zolli is currently in early access.
Host cash payout is not live; early testing uses Zolli credits.
TASK_COMMIT_ACCEPTED