Esta página está disponível apenas em inglês. A versão em inglês é a que prevalece.
Pricing
Setting your price per SKU, how live prices relate to reference prices, and how the scheduler ranks offers.
You set your own price, in MZG per GPU-hour, for each SKU you offer. Customers pay that price per second of use. You receive 93% of what is charged — 74.4% liquid and 18.6% added to your lock — and the 7% protocol fee goes to the treasury.
Setting prices
During mzgrid-provider init you are asked for a price for each detected SKU, pre-filled with the catalogue reference. You can change prices at any time through the API:
PUT /v1/providers/pricing
X-MZG-Provider: 0x8b3f...2e41:1762247400:0x5d0c...1a9b
Content-Type: application/json
{ "rtx4090": "160", "l40s": "305" }New prices apply to jobs created after the change. Jobs that are already funded keep the price recorded in escrow.
Reference and live prices
| SKU | Reference MZG/h | 0.85× | 1.20× |
|---|---|---|---|
b200 | 1,500 | 1,275 | 1,800 |
h200 | 1,000 | 850 | 1,200 |
h100-sxm | 750 | 637.50 | 900 |
h100-pcie | 640 | 544 | 768 |
a100-80 | 460 | 391 | 552 |
a100-40 | 320 | 272 | 384 |
l40s | 290 | 246.50 | 348 |
rtx6000-ada | 240 | 204 | 288 |
rtx5090 | 220 | 187 | 264 |
a6000 | 160 | 136 | 192 |
rtx4090 | 150 | 127.50 | 180 |
rtx4080s | 90 | 76.50 | 108 |
rtx3090 | 75 | 63.75 | 90 |
Live offers on the network typically sit between 0.85× and 1.20× the reference. New providers and community-tier machines tend to price toward the lower end while they build reputation; datacenter-tier providers with long track records can sustain prices toward the upper end.
How the scheduler ranks offers
When a customer uses autoSelect, the scheduler first filters offers to those that are eligible — VRAM at least the template minimum, provider active on-chain, and the job's deposit within your job cap — and then scores them:
score = 0.5 × normalizedPrice + 0.35 × (reliability / 100) + 0.15 × proximity- normalizedPrice (50%) — how cheap your offer is relative to the other eligible offers; the cheapest scores highest.
- reliability (35%) — your reputation score from 0 to 100.
- proximity (15%) — how close your region is to the customer.
Customers who browse the marketplace directly can also sort by price, reliability or availability.
Pricing strategy
- Start near the reference. You have no reputation yet; a price at or slightly below reference gets you into the scheduler's top results while you build a completion record.
- Reliability compounds. A provider with reliability 98 can charge noticeably more than one at 85 and still win the same
autoSelectjobs, because reliability is worth 35% of the score. - Mind your job cap. A low lock limits the size of jobs you are eligible for regardless of price. See Lock & exit.
- Price is per GPU. On multi-GPU machines each GPU serves its own job at the per-GPU price.
Estimating earnings
An RTX 4090 priced at 155 MZG/h that is rented 16 hours a day for 30 days (480 GPU-hours):
| Item | MZG |
|---|---|
| Charged to customers | 74,400.00 |
| Protocol fee (7%) | 5,208.00 |
| Liquid to your wallet (74.4%) | 55,353.60 |
| Added to your lock (18.6%) | 13,838.40 |
These are arithmetic illustrations of the fee and lock split, not forecasts: actual utilisation depends on demand, your price, your reliability and your region. The calculator on Become a provider lets you model your own hardware.