Mine with your GPU.On Bittensor, or in a browser.
Mining here means your GPU answers people's AI questions and earns for the work. There are two ways in, and the GPU you have decides which.
- Mine on Bittensor, with data-center GPUsRun open models for Chutes on subnet 64, inside confidential-compute servers, and earn subnet 64's alpha token.Live on Bittensor. Needs eight data-center GPUs, and TAO to register.
- Mine on Inferno, right in your browserLoad an open model on your GPU and answer chats for people on Inferno. Any recent GPU with WebGPU, and nothing to install.Live beta. When the site has payments on, you earn 70% of each answer, paid out by hand in USDG.
Not sure what yours can do? The test runs right in this tab.
Mine on Bittensor
Bittensor is an open network of subnets, and each one pays miners for one kind of work. On subnet 64, called Chutes, miners run open AI models inside confidential-compute servers and answer requests from everyone who uses Chutes. They earn subnet 64's alpha token, which trades for TAO.
This is data-center work: eight-GPU servers, Linux, Kubernetes and TAO to register. None of it runs in a browser, and none of it is guaranteed to pay.
Can my hardware do it?
Not if it's a gaming PC, a laptop or a Mac. Subnet 64 only takes servers that prove, in hardware, that nobody can look inside: an Intel TDX confidential VM with NVIDIA confidential computing on. Consumer cards like the RTX 4090 don't have it. The GPU test below measures speed, so it can't change that answer, but those cards are a good fit for mining on Inferno, right in your browser.
8
data-center GPUs in one server, in every setup Chutes has tested for subnet 64.
Inferno, in your browser
- GPU
- Any recent GPU with WebGPU. 8 GB runs small models.
- Confidential computing
- Not needed.
- Machine
- Your own computer, with a browser tab open.
- To join
- A wallet on Robinhood Chain.
- Pays
- 70% of each answer, in USDG, when the site has payments on. Paid out by hand during the beta.
Bittensor, subnet 64
- GPU
- Eight H200, B200 or RTX Pro 6000 cards in one server, the setups Chutes has tested. The H100 isn't on that list.
- Confidential computing
- Required: an Intel TDX CPU and NVIDIA confidential computing.
- Machine
- Bare-metal GPU servers on Ubuntu 26.04 with fixed public IPs, plus a control server.
- To join
- A Bittensor wallet, and TAO for the registration price.
- Pays
- Subnet 64's alpha token, for the compute you serve.
Want Bittensor with a consumer card anyway? A few subnets take them. Lium, subnet 51, rents GPUs out, and its validator gives idle RTX 3090, 4090 and 5090 cards a share of emissions. IOTA, subnet 9, trains models and asks for a GPU with 16 GB or more, such as an RTX 4090. Each has its own setup and registration price, and many GPU subnets, subnet 64 included, want data-center cards.
The steps
In order, from Chutes' and Bittensor's own docs. Make the wallet on a computer you trust that isn't one of your servers, and keep the coldkey there.
Checked on 25 September 2026. Chutes changes its miner setup often, so when its docs disagree with this page, follow its docs.
Check the hardware
Chutes has tested three setups end to end: eight H200, eight B200 or eight RTX Pro 6000 cards in one server with an Intel TDX CPU, running Ubuntu 26.04. You also need one server without a GPU, with at least 4 cores and 32 GB of RAM, for the control plane.
Every server needs a fixed public IP, and the GPU servers must be bare metal. Runpod, Vast and similar rentals won't work.
Make a wallet and put TAO in it
pip install bittensor btcli wallet create -w miner -H default
That makes a coldkey, which holds your TAO, and a hotkey, which is your miner's identity on the subnet. btcli shows a recovery phrase for each only once: write them down offline, because they're the only way back in.
Then send TAO for the registration to your coldkey's address, which
btcli wallet listshows. Never send TAO to a hotkey address. Check it arrived:btcli wallet balance miner
Set up the control plane
The control server runs Chutes' miner API, database and scheduler. Ansible sets it up from your own computer. Get the miner repo and a copy of its inventory:
git clone https://github.com/chutesai/chutes-miner mkdir -p ~/chutes cp chutes-miner/ansible/k3s/inventory.yml ~/chutes/inventory.yml
Put your control server and
hotkey_path: ~/.bittensor/wallets/miner/hotkeys/defaultin that file, and your chart settings in~/chutes/values.yaml, as the README explains. Then install Ansible and its collections, and run:cd chutes-miner/ansible/k3s ansible-playbook -i ~/chutes/inventory.yml playbooks/site.yml
Register on subnet 64
The price floats, so check it first. The dry run shows exactly what you'd pay, without paying.
btcli query burn --netuid 64 btcli tx burned-register --netuid 64 -w miner -H default --dry-run
When the price looks right, register for real:
btcli tx burned-register --netuid 64 -w miner -H default
Register once: Chutes says a second slot only competes with your first, so add GPUs to one miner instead. Don't publish an axon either, because Chutes doesn't use one. A new slot is only protected for a while, so have your GPU servers ready to go. You can also watch the price on taostats.
Turn each GPU server into a confidential VM
Chutes' sek8s tools install the TDX kernel and attestation services on each GPU server, then boot the confidential VM that runs your share of Chutes. You'll need an API key from Intel's Trusted Services portal.
git clone https://github.com/chutesai/sek8s cd sek8s/ansible/host cp inventory/hosts.yml ~/chutes/my-inventory.yml
List your GPU servers in
~/chutes/my-inventory.ymland setchutes_hotkey_path: ~/.bittensor/wallets/miner/hotkeys/default. Then:ansible-playbook -i ~/chutes/my-inventory.yml playbooks/setup.yml \ -e pccs_api_key=<your-key> -e pccs_password=<your-password> ansible-playbook -i ~/chutes/my-inventory.yml playbooks/launch.yml
The VM has no SSH, and its disk only unlocks when the hardware proves it's running Chutes' approved image. So nobody, you included, can see the prompts it serves.
Add each GPU server to your miner
Once a VM reports ready, run this once per GPU server, from the computer that holds your hotkey. It checks the server's attestation and offers its GPUs to the validator.
pip install chutes-miner-cli chutes-miner add-node \ --name <gpu-server-name> \ --validator <validator-hotkey> \ --hourly-cost <usd-per-gpu-hour> \ --gpu-short-ref h200 \ --hotkey ~/.bittensor/wallets/miner/hotkeys/default \ --agent-api http://<gpu-server-ip>:32000 \ --miner-api http://<control-server-ip>:32000
The name is the server's name in your inventory. Use
b200orpro_6000for those cards. The README lists the validator's hotkey.Watch it earn
Subnet 64's validators score the compute you served over the last 7 days, so a new miner starts slow. Rewards arrive on your hotkey as subnet 64 stake every tempo, about 72 minutes by default, and you can unstake them into TAO.
btcli wallet overview -w miner
Costs and risks
Know these before you spend anything.
- The registration price
- Paid in TAO from your coldkey. It jumps with every registration and decays while nobody registers, so check it live, with btcli or on taostats. The burned part is never refunded, even if your miner is dropped later.
- Hardware and power
- Eight data-center GPUs in a TDX server and a control server, bought or rented, with power and bandwidth on top. They cost the same whether or not your miner earns.
- Earnings move
- Your share depends on how much compute you serve next to every other miner. It's paid in subnet 64's alpha, and what that's worth in TAO, or in dollars, changes all the time.
- You can be dropped
- Subnet 64 has a fixed number of slots. Once your immunity period ends, the lowest earner is the first pushed out when someone new registers, and coming back means paying the price again.
- No promises
- Inferno doesn't run subnet 64, doesn't pay Bittensor miners, and can't promise you'll earn anything.
How it ties to Inferno
- The same miners as Bittensor chat
- Pick Bittensor in Inferno's chat and Chutes answers on subnet 64 miners like these, inside the same confidential hardware.
- TAO on Robinhood Chain
- TAO reaches Robinhood Chain through Chainlink's CCIP bridge, so Inferno can take it for credits once the site's operator turns TAO payments on.
- Who pays whom
- The Bittensor network pays subnet 64 miners, in alpha. Inferno pays browser miners 70% of each answer, in USDG, when the site has payments on.
Sources
Everything above comes from these pages.
- Chutes miner README: TEE-only network, control plane, registration, adding nodeshttps://github.com/chutesai/chutes-miner
- sek8s host tools: tested GPU setups and Ubuntu 26.04https://github.com/chutesai/sek8s/tree/main/host-tools
- sek8s end-to-end miner guidehttps://github.com/chutesai/sek8s/blob/main/docs/end-to-end-miner.md
- GPU names Chutes acceptshttps://github.com/chutesai/chutes-api/blob/main/api/gpu.py
- Chutes docs: mining on Chutes (last updated in 2025)https://chutes.ai/docs/miner-resources/overview
- Bittensor docs: mininghttps://www.bittensor.com/docs/guides/mining
- Bittensor docs: the registration pricehttps://www.bittensor.com/docs/guides/mining/burn
- Bittensor docs: wallets and keyshttps://www.bittensor.com/docs/concepts/wallets
- Bittensor docs: emissionshttps://www.bittensor.com/docs/concepts/emissions
- btcli now ships in the bittensor packagehttps://github.com/opentensor/btcli
- Subnet 64 on taostatshttps://taostats.io/subnets/64
- Lium, subnet 51: the GPUs it takeshttps://docs.lium.io/providers/architecture
- Lium validator: which idle GPUs earn emissionshttps://github.com/Datura-ai/lium-io/blob/main/neurons/validators/src/services/const.py
- IOTA, subnet 9: compute requirementshttps://github.com/macrocosm-os/IOTA
- TAO on Robinhood Chain over Chainlink CCIPhttps://github.com/TalismanSociety/talisman/pull/2555
Test my GPU
To write each token, about three quarters of a word, a model reads all of its weights from GPU memory. So how fast your card moves memory sets how fast it can answer. This test measures that with WebGPU, right in this tab.
It takes a few seconds. Nothing downloads, and nothing leaves this page.
0GB/s
Memory bandwidth, estimated in this browser
Estimated speed
For the RTX 3060 12GB, with 4-bit models, one conversation at a time.
| Model | Size | Tokens a second |
|---|---|---|
| 1B class | 0.9 GB | About 220 |
| 8B class | 4.7 GB | About 42 |
| 70B class | 40 GB | Won't fit in 12 GB |
Speed is bandwidth × 0.55 ÷ model size, because answering never quite reaches a card's full bandwidth.
What it could earn
For the RTX 3060 12GB. Test your GPU or pick another card to change it.
The share of those hours your GPU spends answering requests.
76.44USDG a month
917.24USDG a year
42 tokens a second, 30% of the time, for 8 hours a day over 30 days comes to 10.9 million tokens a month. At 1 credit per 1,000 tokens, and $0.01 a credit, people would pay $109.19 for them. You keep 70%, which is 76.44 USDG.
Estimate from the card you picked. The small price is what the network charges today; the large price is planned. Real earnings depend on demand.
Serve from this browser
Load a model on your GPU and go online. This tab then answers chats for people on the network, who pick a lender by wallet address.
People's words pass through this tab's memory to your GPU and are never shown or saved here. You never see who is asking.
This runs on the beta relay.
Connect your wallet
Your address is how people pick you, and where payouts go.
Load a model
Sizes are the graphics memory each model needs. The first load downloads it from Hugging Face; after that it's cached in this browser.
Go online
Your wallet signs a note that links this tab's key to your address. Signing is free and moves no funds.
What you need to mine on Inferno
- A GPU with enough memory
- 8 GB or more for small models. Large models need at least 24 GB, and a whole 70B-class model at 4-bit takes about 40 GB.
- A steady connection
- Your computer stays online while you mine. Requests and answers are text, so an ordinary home connection is fine once the model has downloaded.
- A wallet on Robinhood Chain
- Payouts go there, in USDG. MetaMask, Robinhood Wallet or any wallet that can add Robinhood Chain will do.
- The worker
- You can serve from a browser tab today, on the beta relay, with nothing to install. A native app is planned.
How Inferno pays you, and keeps you honest
70%
of what each answer costs goes to the lender whose GPU served it, while the site has payments on.
- Paid per answer
- People pay for network answers with Inferno credits, at $0.01 per 1,000 tokens. You earn 70% of each answer your GPU serves, and the ledger records it.
- Paid out in USDG
- During the beta, payouts go out by hand, in USDG, to the wallet you serve with. There's no payment contract.
- Only while payments are on
- The site's operator turns payments on. While they're off, network chat is free and serving earns nothing.
- Vetted lenders
- Planned. Today any wallet can serve on the beta relay, and chat only lists lenders whose wallet signed their key.
- Spot checks
- Planned: prompts with known answers, to check you run the model you say you run.
- Faked work isn't paid
- You're paid for answers your GPU actually wrote. Once spot checks are live, an answer that fails one earns nothing.
Start with a wallet
Your wallet address is how people pick you, and where your USDG goes. Connecting adds Robinhood Chain to your wallet and shows this page your address. Nothing is signed until you go online.