io.net

io.net is a GPU cloud for AI teams that runs its payments on Solana. It gathers spare graphics chips (GPUs) from data centers, crypto miners and other networks, then rents them out by the hour for training and running AI models. People who supply the chips are paid in the network's token, IO. That makes io.net a DePIN, a network where a crypto token pays people to supply real hardware. io.net sells three things: rented GPU machines (IO Cloud), pay-per-use access to open AI models (IO Intelligence), and a way for AI agents to rent machines themselves (Agent Cloud). The io.net website advertises prices about 70% below AWS.

How does io.net work?

io.net works as a matchmaker between people who own GPUs and people who need them. Owners install the IO Worker software and list their machines. Customers pick a chip type and a location, pay, and get a working machine or cluster in minutes. io.net's software handles the scheduling, networking and billing in between. Bookings and payments are recorded on Solana.

The io.net docs list four jobs the network is built for: serving models, training in parallel, tuning models, and reinforcement learning. All four can be split across many machines at once, which is why scattered hardware can do them.

You can pay with a card, USDC or IO. Whatever you pay with, the IO coin docs say the payment is converted to IO behind the scenes, and suppliers are paid in IO. That conversion is the link between customer demand and the token.

The idea came from the founders' own bill. The team was building trading systems and needed more than 50 NVIDIA A100 cards, which the company history page prices at over $80 per card per day. In June 2022 they switched to building the GPU network instead.

IO Cloud: renting GPU clusters, containers and virtual machines

IO Cloud is where you rent GPU machines on io.net. You can launch a container or a virtual machine yourself, from the dashboard or through the API. For bigger jobs, the io.net team will set up bare metal servers, Kubernetes clusters or Ray clusters for you. Ray is open-source software that spreads one AI job across many machines.

You pay from a prepaid balance called IO Credits, priced in dollars. Adding credits with USDC or IO is free. Adding them by card costs a 5% fee, the IO Cloud payments page says.

io.net also rents machines for private data, which it calls confidential compute. These machines keep your data locked while they work on it. They use Intel TDX or AMD SEV-SNP chips next to NVIDIA GPUs. You have to ask for them by email, the confidential compute docs say, so you cannot launch one from the dashboard.

On 5 October 2026, io.net said the network had delivered more than 34 million compute hours.

IO Intelligence and the OpenRouter inference business

IO Intelligence lets you use open AI models on io.net without renting a machine. You send a prompt to an API and pay for what you use. The service also has tools to build AI agents and to fine-tune models. A "confidential" mode adds a proof that your prompt ran on a secure chip.

This side of io.net has grown fastest in 2026. io.net sells its hosted models through OpenRouter, a marketplace that sends developer requests to model providers. io.net said its daily volume there rose from 500 million tokens in March 2026 to 15 billion in late September. A token here is a chunk of text a model reads or writes. It has nothing to do with crypto tokens.

In early October 2026 io.net reported more than 20 billion tokens in a single day. These are io.net's own figures. They show where the business is heading: selling cheap use of open models as well as renting raw chips.

Agent Cloud: GPUs that AI agents can rent and pay for themselves

Agent Cloud lets an AI coding agent rent and manage io.net machines without a person clicking through a dashboard. It is a Model Context Protocol (MCP) server, a standard plug that tools such as Claude Code, Cursor and Windsurf can connect to. Once connected, the agent can launch containers or virtual machines, scale them and shut them down.

Payment is the interesting part. If the account runs short of credits, the server answers with a payment request using x402, a standard for paying over the web. The agent can then pay in USDC on Solana, with a $1 minimum top-up, per the Agent Cloud docs. Each deployment is limited to one location.

IO coin tokenomics: supply, emissions and the Incentive Dynamic Engine

IO has a hard cap of 800 million coins. Of those, 500 million were created when the token launched in June 2024. The rest were set aside as rewards for GPU suppliers and stakers, to be paid out over 20 years. The emission schedule pays out a little less every month.

On 11 June 2026 io.net announced a new system on top of that schedule, called the Incentive Dynamic Engine (IDE). Under the IDE, supplier pay is set in dollar terms. At least half of the revenue left after paying suppliers is used to buy IO and burn it, which means destroying it for good.

In its press release, io.net said it expects to burn at least 12 million IO in the first year. The same release claimed an $8 million contract with one large customer.

An independent newsletter, Own Your Mind, checked the on-chain data in a review published on 22 June 2026. It confirmed the burn is live and can be tracked on io.net's explorer. It also found that customer purchases recorded on-chain were roughly 3% of what suppliers were paid. Its view was that the burn was being paid for mostly by newly issued tokens, with revenue covering little of it.

The tokenomics page in io.net's docs still describes the older fixed schedule.

Staking IO and worker earnings for GPU suppliers

To earn on io.net as a supplier, you connect a machine with IO Worker and stake IO against it. Staking means locking tokens as a deposit. The base requirement is 200 IO per graphics card, multiplied up for stronger chips. The staking docs give the example of a server with eight H100 cards needing 16,000 IO.

Suppliers earn in two ways: pay from customers who rent the machine, and "block rewards" from the IO reward pool. To qualify for rewards, a machine must pass an hourly test that proves the chip is real and as fast as claimed. The proof-of-work page says the test loads the machine for about 15 minutes of each hour.

There are two catches. Unstaking takes a 14-day cooldown. And stake can be slashed, meaning partly taken away, if a machine fakes its hardware or serves customers badly.

If you hold IO but own no hardware, you can co-stake on someone else's machine and share its rewards at a rate the owner offers. The co-staking docs warn that slashing can hit your principal too, and each account can take up to two offers.

Who io.net is for

io.net suits AI startups and research teams that need GPUs fast, by the hour, with no long contract. It fits best when the work can be split across machines, such as fine-tuning a model or running a large batch of prompts. Teams that only want cheap access to open models can use IO Intelligence and never touch a server.

io.net is a weaker fit if you need one uptime promise that covers every machine. The hardware comes from many owners, so quality varies from one cluster to the next. Check the exact chips, region and network speed before you move a live product onto it.

For suppliers, io.net is aimed at data centers and miners with modern NVIDIA cards. The staking deposit and the hourly test make it a poor fit for a casual gaming PC.

Is io.net legit? Risks, incidents and centralization

io.net is a real business with paying customers. There are still four risks to know before you use it or hold IO.

The network has been gamed before. On 25 April 2024, attackers used leaked user ID tokens to change the details of GPUs listed on the network. After the fix, active GPU connections fell from about 600,000 to about 10,000, Cointelegraph reported. The drop showed how many listed machines were fake. It is the reason the hourly hardware test exists.

io.net is run by a company. The company runs the scheduling, the billing and the model hosting itself. Own Your Mind's June 2026 review said the core platform is closed and that token holders have no vote. Your balance is a prepaid credit held with io.net. The on-chain transaction docs do say that bookings are held in escrow by Solana smart contracts.

Most growth figures come from io.net itself. The contract size, token volumes and compute hours on this page have not been checked by an outside auditor.

We found no audit report for the staking contracts in io.net's list of docs pages. Solana Compass has no program addresses on file for io.net, so we could not check who is able to change those contracts.

io.net vs Akash, Render and Aethir

io.net is the most AI-focused of the well-known crypto GPU networks. It is also the one most clearly run by a single company.

Akash is a general market for containers, where providers bid for your job. Akash is more open, but you manage more of the setup yourself. Render Network RENDER$1.61+5.1%, which also uses Solana, was built for 3D rendering, and it supplies some of io.net's hardware too. Aethir sells reserved GPU capacity to larger business buyers.

Pick io.net if you want ready-made AI clusters or a cheap model API. Pick Akash if open control and portable containers matter more to you. Outside crypto, the main choice is a big cloud such as AWS. AWS costs more, but you get one contract and one party to hold to account.

Who runs io.net: team and history

io.net began as a trading-systems company and switched to GPU compute in June 2022. The IO token launched on 11 June 2024.

io.net has had three chief executives. Founder Ahmad Shadid stepped down in June 2024, two days before the token launch, and co-founder Tory Green took over. In April 2025 Gaurav Sharma became CEO. Sharma had joined as chief technology officer after working as an engineering executive at Binance. Green moved to chair the io.net Foundation.

Suppliers can also connect an Aptos wallet, though block rewards are paid to Solana wallets.