Build your own GPT.
From USD 0.08 an hour.
Train Llama, Qwen or Gemma on your own data, then serve it the same night. Nothing to buy and nothing to install — the GPU is yours over SSH or JupyterLab, billed by the second.
No card on file, no contract, and time a host spends offline is never billed.
Live prices by GPU
| GPU | Memory | From, per hour | Best host reliability | Hosts |
|---|---|---|---|---|
| /h | ||||
| No offers right now. The market refreshes every minute. | ||||
Lowest price per GPU type across every verified host on the board right now; the console lists each host with its own price, reliability score and location. Reliability is the host's historical uptime score, measured across every rental of that machine (new hosts start at 60% and are listed only above %). Storage and network transfer are billed separately at the rates shown in the console.
What will your job cost?
Priced from the board as it stands right now. Pick a job — the cheapest card is rarely much slower, and you can see it for yourself before you spend anything.
How it works
Add credits
Pay locally in your currency, send the reference, and we add the credits after a quick review.
Pick a GPU
Choose from the live board. The price you see is locked for the life of your instance; you pay only for the seconds it runs.
Connect
Your SSH key is installed on the machine. Open JupyterLab or SSH in — usually within a minute on a warm host, a few minutes on a cold one.
Ready-to-run software
Every instance starts from a maintained environment — no Docker, no drivers, no setup. Pick the software first and we show only the hosts that can run it.
Every GPU on the board
Grouped by generation. "From" is the lowest price on a verified host right now; the line is the last 30 days of that lowest price.
Hopper — datacenter5 types · 18 hosts
| H200 NVL | 140 GB | USD 6.62/h | 99.8% | 6 hosts |
| H100 SXM | 80 GB | USD 2.78/h | 99.9% | 4 hosts |
| H100 NVL | 94 GB | USD 4.62/h | 99.9% | 3 hosts |
| H200 | 140 GB | USD 8.01/h | 99.9% | 3 hosts |
| H100 PCIE | 80 GB | USD 3.51/h | 99.3% | 2 hosts |
Ampere — datacenter3 types · 15 hosts
| A100 SXM4 | 40 GB | USD 0.78/h | 99.9% | 11 hosts |
| A100 PCIE | 80 GB | USD 0.97/h | 99.9% | 3 hosts |
| A40 | 45 GB | USD 0.76/h | 98.7% | 1 host |
Ada — datacenter3 types · 10 hosts
| L4 | 22 GB | USD 0.44/h | 99.9% | 5 hosts |
| L40S | 45 GB | USD 0.86/h | 99.9% | 4 hosts |
| L40 | 45 GB | USD 0.80/h | 98.7% | 1 host |
Ada — workstation5 types · 12 hosts
| RTX 6000Ada | 48 GB | USD 0.87/h | 99.9% | 4 hosts |
| RTX 5880Ada | 48 GB | USD 1.29/h | 99.7% | 3 hosts |
| RTX 4500Ada | 24 GB | USD 0.63/h | 99.3% | 2 hosts |
| RTX 5000Ada | 32 GB | USD 0.63/h | 99.0% | 2 hosts |
| RTX 4000Ada | 20 GB | USD 0.28/h | 99.6% | 1 host |
Ada — consumer9 types · 58 hosts
| RTX 4090 | 24 GB | USD 0.63/h | 99.9% | 38 hosts |
| RTX 4060 Ti | 16 GB | USD 0.18/h | 99.8% | 8 hosts |
| RTX 4070S | 12 GB | USD 0.21/h | 99.1% | 3 hosts |
| RTX 4070S Ti | 16 GB | USD 0.33/h | 100.0% | 3 hosts |
| RTX 4070 | 12 GB | USD 0.23/h | 99.9% | 2 hosts |
| Q RTX 4000 | 8 GB | USD 0.10/h | 98.7% | 1 host |
| RTX 4060 | 8 GB | USD 0.13/h | 99.6% | 1 host |
| RTX 4080S | 16 GB | USD 0.43/h | 99.3% | 1 host |
| RTX 4080 | 16 GB | USD 0.44/h | 99.2% | 1 host |
Ampere — workstation4 types · 19 hosts
| RTX A4000 | 16 GB | USD 0.13/h | 99.9% | 7 hosts |
| RTX A6000 | 48 GB | USD 0.65/h | 99.9% | 6 hosts |
| RTX A2000 | 6 GB | USD 0.09/h | 99.9% | 3 hosts |
| RTX A5000 | 24 GB | USD 0.38/h | 99.8% | 3 hosts |
Ampere — consumer9 types · 87 hosts
| RTX 3090 | 24 GB | USD 0.23/h | 100.0% | 25 hosts |
| RTX 3060 | 12 GB | USD 0.09/h | 100.0% | 22 hosts |
| RTX 3080 | 10 GB | USD 0.22/h | 99.9% | 14 hosts |
| RTX 3070 | 8 GB | USD 0.14/h | 99.9% | 8 hosts |
| RTX 3080 Ti | 12 GB | USD 0.20/h | 99.8% | 6 hosts |
| RTX 3060 Ti | 8 GB | USD 0.18/h | 99.6% | 5 hosts |
| RTX 3060 laptop | 12 GB | USD 0.10/h | 99.9% | 3 hosts |
| RTX 3070 Ti | 8 GB | USD 0.20/h | 99.7% | 2 hosts |
| RTX 3090 Ti | 24 GB | USD 0.54/h | 99.1% | 2 hosts |
Turing8 types · 19 hosts
| RTX 2060 | 12 GB | USD 0.16/h | 99.8% | 4 hosts |
| RTX 2060S | 8 GB | USD 0.11/h | 99.7% | 3 hosts |
| RTX 2070 | 8 GB | USD 0.12/h | 99.5% | 3 hosts |
| Q RTX 8000 | 45 GB | USD 0.41/h | 99.9% | 3 hosts |
| RTX 2080 Ti | 22 GB | USD 0.16/h | 99.5% | 2 hosts |
| Titan RTX | 24 GB | USD 0.39/h | 99.9% | 2 hosts |
| RTX 2070S | 8 GB | USD 0.18/h | 99.8% | 1 host |
| Tesla T4 | 15 GB | USD 0.26/h | 98.5% | 1 host |
Volta2 types · 8 hosts
| Tesla V100 | 32 GB | USD 0.18/h | 100.0% | 6 hosts |
| Titan V | 12 GB | USD 0.24/h | 99.8% | 2 hosts |
Pascal7 types · 20 hosts
| GTX 1080 Ti | 11 GB | USD 0.11/h | 99.8% | 8 hosts |
| GTX 1080 | 8 GB | USD 0.09/h | 99.6% | 3 hosts |
| Titan Xp | 12 GB | USD 0.09/h | 99.6% | 3 hosts |
| Tesla P100 | 16 GB | USD 0.14/h | 99.6% | 3 hosts |
| GTX 1070 | 8 GB | USD 0.08/h | 98.9% | 1 host |
| GTX 1060 | 3 GB | USD 0.15/h | 99.9% | 1 host |
| Tesla P40 | 23 GB | USD 0.16/h | 99.5% | 1 host |
Maxwell1 type · 1 hosts
| GTX TITAN X | 12 GB | USD 0.09/h | 99.5% | 1 host |
Mining cards1 type · 2 hosts
| CMP 170HX | 64 GB | USD 0.65/h | 99.1% | 2 hosts |
Other3 types · 6 hosts
| GTX 1660 S | 6 GB | USD 0.10/h | 99.7% | 3 hosts |
| GTX 1660 Ti | 6 GB | USD 0.12/h | 99.8% | 2 hosts |
| GTX 1660 | 6 GB | USD 1.09/h | 99.2% | 1 host |
Built for scripts too
Everything the console does, your code can do with an API key from the Keys tab. Rent, connect, stop, destroy — in your currency, with the same limits and statements.
# find the cheapest RTX 4090 that can run vLLM, then rent it
curl -H "Authorization: Bearer tc_live_…" "https://powerpod.si/api/v1/offers?gpu=RTX%204090&sort=price"
curl -X POST -H "Authorization: Bearer tc_live_…" -H "content-type: application/json" \
-d '{"offer_id": 43681494, "template_id": 9, "disk_gb": 60, "ssh_key_ids": [7]}' https://powerpod.si/api/v1/instances
# the reply carries ssh_command, the Jupyter link and app links once the host is runningFull reference at /api/docs. Keys carry scopes (read / instances / billing) and every call is in your audit trail.
What we promise, and what we don't
We do
- Bill per second at the price locked in your currency, with every charge listed in your statement.
- Show the host's reliability score on every offer and every instance.
- Stop your instance before your credit goes negative, and keep the disk for 7 days so you can top up.
- Pause billing while a host is offline, and credit any time you lose to it.
- Offer scheduled backups of your workspace to cloud storage.
We don't
- Guarantee uptime. Hosts belong to a marketplace and can go offline without notice.
- Keep your data if a host disappears — turn on backups for anything you can't lose.
- Allow crypto-mining, attacks or anything illegal. Misuse closes the account.
Questions people ask first
How do I pay?
In your own currency, with the methods available where you are — in Nepal that is Fonepay QR, bank transfer, eSewa or Khalti; in India UPI or IMPS/NEFT; elsewhere PayPal, Wise or a card payment link. We detect your country automatically and you can change the currency yourself before the first payment. Enter the transaction reference in the console; credits appear after a short review. There is no card on file and nothing is charged automatically.
Do I need to send an ID?
No. Sign up with Google or an email address, add credits, and rent. Every payment carries the payer's name, and we check it before the credits land. We ask for a photo ID only in rare cases, such as a payment that does not match or a misuse report, and delete it within 90 days of the account closing.
Is the price on the board what I pay?
Yes — per second, at the price locked when you rent, in your account currency. Storage is billed per GB-month while the instance exists (also while stopped) and network transfer per GB; both rates are shown before you confirm. Every charge is itemised in your monthly statement.
What happens when my credit runs out?
You get warnings at 30 and 10 minutes of runtime left. Below five minutes the instance is stopped automatically so your balance never goes negative. The disk is kept for 7 days so you can top up and start again; after that it is destroyed.
Do you guarantee uptime?
No, and we don't hide it: the hosts belong to a marketplace and can go offline. Each offer shows the host's own reliability history. If a host drops, billing pauses and lost time is credited — but data on a host that disappears is gone, so turn on backups for anything you can't lose.
Can I run my own Docker image?
Not yet. You choose from maintained environments (PyTorch, TensorFlow, ComfyUI, Stable Diffusion, Ollama, vLLM and more) and install anything you like inside them with root access. It keeps starts fast and the network safe for everyone.
Can I automate this?
Yes. API keys with scopes let scripts browse offers, rent, read the SSH and Jupyter hand-off, stop and destroy — see /api/docs.
Who is behind this?
PowerPod is run by Talkc Ai Pvt. Ltd., Kathmandu. Write to compute@talkc.ai — a person answers.