Bu sayfa yalnızca İngilizce olarak mevcuttur. Bağlayıcı olan İngilizce sürümdür.
Introduction
What MZGrid is, who uses it, and how a GPU job moves from a funded escrow to a settled payment on Base.
MZGrid is a decentralized marketplace for GPU compute. Independent providers — from single-GPU workstations to datacenter racks — list their hardware, and customers rent it by the second to run AI inference endpoints, notebooks, image generation and batch workloads. Every payment is made in MZG, the network's ERC-20 token on Base, and every job is backed by an on-chain escrow. For the full protocol design, token model and roadmap, read the MZGrid white paper.
- Sağlayıcılar
- Çevrimiçi GPU’lar
- Kullanım oranı
- İşler · 24 sa
- Mutabakatı yapılan MZG · 24 sa
Why MZGrid
- One unit of account. Offers are priced in MZG per GPU-hour, deposits are in MZG, settlement is in MZG. There are no price oracles, second payment assets or conversions anywhere in the protocol.
- Pay for what you use. You deposit
pricePerHour × maxDurationup front, you are billed per second of actual use, and the unused remainder is refunded automatically at settlement. - Providers have skin in the game. There is no separate staking step. Instead, 20% of every provider payout is locked in the
ProviderRegistrycontract and serves as collateral that can be slashed if a provider abandons a job, goes offline or fails verification. - Production-ready templates. Launch an OpenAI-compatible vLLM endpoint, ComfyUI, Whisper or JupyterLab with one call, or bring your own Docker image.
Roles
| Role | What they do | Where to start |
|---|---|---|
| Customer | Picks an offer (or lets the scheduler pick), funds a job in escrow, uses the endpoint, gets refunded for unused time. | Quickstart, Deploying workloads |
| Provider | Runs the mzgrid-provider agent on NVIDIA hardware, sets prices per GPU model, executes jobs in isolated containers, earns MZG. | Installing the provider agent |
| Coordinator | The MZGrid API and scheduler. Matches jobs to offers, verifies benchmarks and usage, and submits settlement to the escrow contract with its oracle role. | API reference |
| Contracts | MZGToken, MZGridEscrow and ProviderRegistry on Base hold funds, enforce fees, locks, job caps and slashing. | Smart contracts |
How a job flows
- Create. The customer calls
POST /v1/jobs(or clicks Deploy in the marketplace) with an offer id orsku+autoSelect, a template and a maximum duration. The coordinator returns the job and the exact escrow parameters to sign. - Fund. The customer's wallet calls
createJob(after anapprove) orcreateJobWithPermitonMZGridEscrow. The deposit —pricePerHour × maxDuration— moves into escrow. The job is nowfunded. - Accept. The provider agent picks up the job and calls
acceptJobwithin the 10-minute accept window. The escrow sets a deadline ofnow + maxDurationand the job becomesrunningonce the container is up. - Run. The provider's agent pulls the template image, starts the container on the assigned GPU, publishes the endpoint and reports GPU utilisation every 60 seconds.
- Settle. When the workload finishes, the duration expires or the customer stops early, the coordinator calls
completeJob(jobId, usedSeconds, resultHash). The escrow computes the charged amount and splits it in one transaction:
| Recipient | Share of the charged amount |
|---|---|
| Protocol treasury (fee) | 7% |
| Provider, liquid | 74.4% (80% of the 93% provider share) |
Provider, locked in ProviderRegistry | 18.6% (20% of the 93% provider share) |
Whatever was not charged goes straight back to the customer in the same transaction.
Not
The 7% protocol fee is already part of the offer price. A customer paying 750 MZG/h pays exactly 750 MZG per hour of use — the fee is taken out of that amount, not added on top.
What happens when something goes wrong
- The provider never accepts → after 10 minutes the customer can call
cancelJobfor a full refund. - The job fails or the provider is caught misbehaving → the coordinator calls
failJoband the customer receives a full refund; a slash may additionally send part of the provider's locked MZG to the customer. - The coordinator never settles → once
deadline + 1 hourhas passed, the customer can callreclaimTimeoutdirectly on the contract for a full refund, no API involved.
Next steps
- Quickstart — deploy a Llama 3.1 8B endpoint in a few minutes.
- Core concepts — offers, SKUs, escrow, lock, reputation and job caps.
- Become a provider — earnings, requirements and rules.