Rent the GPU.
Keep the change.
GPUs by the second and AI models by the output, on Robinhood Chain. Your balance sits in a vault you can empty at any time. A rental locks one hour, counts every second, and hands the rest back when it stops.
From 0.288 USDG an hour. Images from 0.0036 USDG.
Instance
H100 SXM
80 GB VRAM
Elapsed
00:12:34
Rate
3.094/ hr
Cost so far
0.648022
Locked for this hour
3.094000
Your money stays
where you can reach it.
Deposits land in a vault with one balance per wallet. Only you can withdraw it. Coldplate takes payment in four ways, each one bounded by something you signed: outputs within the daily limit you set, a rental lock that settles by the second, an agent voucher within that agent's limit, or an x402 payment for one exact amount.
Withdraw any time
Your available balance leaves the vault on your word alone, whenever you ask.
A lock you can reclaim
If a rental is not settled within a day of its deadline, anyone can send its whole lock back to you.
Limits held onchain
The app and each agent spend inside a daily cap that lives in the vault, and you can end any of them at once.
Say how much memory.
Get the GPU that fits.
Ask for gpu: "auto" with the VRAM your job needs and what matters to you. Coldplate takes the GPU that fits, starts it with SSH and Jupyter, and stops it when it sits idle.
| GPU | VRAM | Generation | USDG / hr | per second |
|---|---|---|---|---|
| RTX A4000 | 16 GB | Ampere | 0.288 | 0.000080 |
| RTX A5000picked | 24 GB | Ampere | 0.311 | 0.000086 |
| RTX 3090 | 24 GB | Ampere | 0.495 | 0.000138 |
| L4 | 24 GB | Ada | 0.495 | 0.000138 |
| RTX A6000 | 48 GB | Ampere | 0.564 | 0.000157 |
| RTX 4090 | 24 GB | Ada | 0.794 | 0.000221 |
| L40S | 48 GB | Ada | 0.989 | 0.000275 |
| RTX 5090 | 32 GB | Blackwell | 1.081 | 0.000300 |
| A100 PCIe | 80 GB | Ampere | 1.886 | 0.000524 |
| A100 SXM | 80 GB | Ampere | 2.001 | 0.000556 |
| H100 PCIe | 80 GB | Hopper | 2.749 | 0.000764 |
| H100 SXM | 80 GB | Hopper | 3.094 | 0.000859 |
| H200 | 141 GB | Hopper | 4.129 | 0.001147 |
| B200 | 180 GB | Blackwell | 6.889 | 0.001914 |
One prompt in.
One price out.
Ready models for pictures, clips, voice and 3D. Each output has its own price, taken from your balance once the output is ready.
FLUX.1 schnell
Image by Black Forest Labs
0.0036
per image
FLUX.1 dev
Image by Black Forest Labs
0.03
per image
FLUX1.1 pro
Image by Black Forest Labs
0.048
per image
Seedream 4
Image by ByteDance
0.036
per image
Qwen Image
Image by Alibaba Qwen
0.03
per image
Seedance 1 lite
Video by ByteDance
0.108
per 5 second clip
Wan 2.2 fast
Video by Wan
0.06
per 5 second clip
Speech 02 turbo
Voice by MiniMax
0.072
per 1,000 characters
Chatterbox
Voice by Resemble AI
0.03
per 1,000 characters
TRELLIS
3D by Microsoft
0.048
per 3D asset
Agents get a budget.
The vault keeps it.
Link an agent's address and give it a daily limit with one signature. It pays from your balance with vouchers it signs, and the vault settles them only up to that limit. Agents without a link pay each request with x402 in USDG. Both work for outputs and for whole GPUs, and the MCP server gives any model the tools to use them.
// 1. ask POST /api/v1/outputs { model: "flux-schnell", prompt } 402 PAYMENT-REQUIRED exact · coldplate-voucher // 2. pay from the wallet's balance sign AgentVoucher { owner, agent, cumulative, deadline } POST /api/v1/outputs + PAYMENT-SIGNATURE 200 output ready · settles in the next batch // 3. or rent a whole GPU POST /api/v1/instances { gpu: "auto", vramGb: 48, hours: 2 } sign RentalOpen { id, lock, deadline } 202 starting · ssh root@host -p port
Every charge
is an event.
Outputs are charged in groups per wallet, and each group carries the Merkle root of its outputs. Your receipts page shows the proof that ties each output you made to the charge on the chain.