io.net is a company that rents out AI computing power. Its promise is easy to like: the same computing power that costs a fortune at Amazon, Google, or Microsoft, for up to 90 percent less. It pulls this off by borrowing chips that sit unused all over the world and stitching them together, over the internet, into one Frankenstein supercomputer of sorts.

I wanted this one to be true the first time I read about it, because cheap computing power for everyone sounds like exactly what the AI era needs. So this issue asks the two questions that matter: is it legit, and what is the catch?

To see why a 90 percent discount would be a big deal, you need to see the full price. Here is the receipt for one AI training cluster, the kind of machine that teaches an AI model everything it knows.

Add it up and one machine costs billions, and the chips inside it become outdated within a few years. When Sam Altman, the head of OpenAI, was asked in 2023 whether training GPT-4 had cost $100 million, his answer was "It's more than that." That was several chip generations ago, and prices have gone up since.

These economics are also why data centers turned into national politics. The United States announced a project called Stargate from the White House: a plan backed by OpenAI, Oracle, and SoftBank to spend $500 billion building AI data centers in the United States, including one Texas site designed to hold over 450,000 chips. Washington also bans the sale of top AI chips to China, the way it once restricted weapons technology, because whoever owns these machines decides what AI gets built.

Microsoft, Google, Amazon, Meta, the AI labs they fund, and governments. They can pay for it, and they can reserve future versions of it, because next year's chips are sold before they are even manufactured.

Everyone else waits in line or pays prices designed for giants. A medical-imaging startup, a university lab, or a solo researcher with a good idea builds on the giants' terms or does not build at all. This is the market io.net walked into, because everyone who needs the machine and cannot afford the receipt is a potential customer.

If big tech bought up the supply, where does io.net get chips? From three places where powerful chips sit around doing nothing.

The biggest pile came from crypto itself. In September 2022, Ethereum switched off mining overnight, and warehouses full of GPUs that had been earning money by the hour suddenly earned nothing. Their owners were stuck with the hardware, the cheap electricity contracts, and no work. The second pile is gaming PCs, because the same chips that render video games happen to be good at AI math, and they sit unused most of the day. The third pile is independent data centers with idle capacity between their customers' busy hours.

io.net's software finds these chips, verifies them, and synchronizes them over the internet so they can work on one job together. That software is a modified version of Ray, an open-source tool that splits a big computing job into thousands of small pieces, hands the pieces out to many machines, collects the answers, and recovers when any machine drops out. Ray is respected in the AI world; it helped train GPT-3 and GPT-4.

Partly yes, and both sides of the answer matter.

The product is real. Customers, mostly AI startups, rent clusters and run real work on them. Analysts at Nansen, a blockchain research firm, measured the network earning more than $1 million a month by early 2025.

The company's history with its own numbers is the problem. In April 2024, io.net ran a token giveaway that rewarded people for the hardware they contributed. At the exact moment it was measuring contributions, attackers exploited a security hole and registered roughly 1.8 million fake GPUs to collect rewards. Around the same time, outside observers caught io.net publishing four different chip counts at once. The company admitted the failure and fixed the hole, and its founder resigned two days before the token launched. Ever since, every number io.net publishes gets read with suspicion, and the suspicion is earned.

That history is why the newest change matters. Since June 11 of this year, when customers pay io.net, the company takes at least half of that revenue in its own IO tokens and destroys those tokens permanently, a practice crypto calls a burn. Destroying tokens shrinks the total supply, which makes the tokens that remain scarcer, similar to the way a company buying back its own shares makes each remaining share worth a bigger slice. And because the destruction happens on the Solana blockchain, which is a public ledger, anyone can look up the transactions and add them up. You do not have to take io.net's word for it. Is revenue harder to fake than a chip count? Yes, because faking it costs real money: a company pretending to have customers would be paying itself and then destroying its own cash. It is not impossible, but it is expensive, which makes this the first io.net number an outsider can actually check. The burn started seven weeks ago, so it cannot prove a trend yet.

How big is the discount really? io.net's ads say up to 90 percent off the big clouds. Independent reviews agree it is much cheaper; the exact savings depend on which chip you rent and what job you run.

So the first question gets a qualified yes. The product works, the customers are real, and the company earned the suspicion it now operates under. That makes the second question the important one.

What do you trade for the low price? There are three tradeoffs.

The first tradeoff is speed. Chips inside one supercomputer building talk to each other over connections about a thousand times faster than the internet, and the biggest training jobs need that constant conversation. io.net's chips talk over the internet, so those jobs are out of reach. So what kind of jobs can it do? The everyday ones: running trained models, which is the work that powers an app or a chatbot; tuning a model on your own data; and training smaller models. That is a real market. It is also a smaller promise than the word supercomputer suggests.

The second tradeoff is capacity. How much computing power can you actually get? The claims vary wildly, which is part of io.net's reputation problem: marketing has said one million GPUs, about 327,000 have passed a verification check at some point, and the number actually ready to work on a given day has measured five to seven thousand. Now remember the receipt. Training a serious model takes ten thousand chips or more, so on a typical day io.net does not have enough ready chips to assemble even one full training cluster.

The third tradeoff is reliability. These are borrowed machines. A gaming PC leaves the network when its owner turns it off, and a data center pulls its spare chips back when its own customers need them. The software recovers when a machine drops out, but io.net does not offer the uptime guarantees the big clouds sell. It fits work that can tolerate a hiccup, and it is a hard sell for work that cannot.

The technology is real, the customers are real, and the discount is real. What you trade for the discount is speed, capacity, and reliability.

The Gate is my first filter for any crypto project, and here is how it scores, straight from the Vault Letter that introduced it. Eight questions, each scored green, yellow, or red. Green means clear evidence and real traction. Yellow means some promise with gaps or open questions, and unknowns default to yellow. Red means weak: no real need, no users, or obvious risk. A project passes with five or more greens and no more than one red, which earns it the deeper evaluation I call the Key. Three or four greens puts it on the watchlist to revisit later. Three or more reds means stop spending time on it.

One green, six yellows, one red. A pass takes five greens and io.net has one, so it does not pass the Gate and does not earn the deeper Key evaluation. It is also short of a hard fail, which takes three reds. It lands on the watchlist, to revisit if the numbers change.

The technology is legit, and it works today for the everyday jobs: running trained models, tuning them on your own data, and training smaller models. It is slower than a true supercomputer. Its capacity on a given day is too small to train a big model like the ones behind ChatGPT. And its reliability makes it a fit for work that can tolerate interruptions, at a price that makes those tradeoffs worth considering.

Three things, and all of them are watchable.

More chips. The capacity problem is a supply problem. More owners adding their machines to the network would grow the number of chips actually available each day, and if that number climbed from five thousand toward the ten thousand a training cluster needs, io.net could start selling the jobs it currently cannot.

A better answer to the speed problem. The internet is the bottleneck. Researchers are already working on training methods that need far less conversation between chips; the best known is DeepMind's DiLoCo, which is short for distributed low-communication training. If methods like that mature, a scattered network like io.net's becomes capable of much bigger jobs, and the tradeoff that limits it today gets smaller.

A more reliable system. The fix for machines that drop out is redundancy: give the same piece of work to more than one machine, save progress constantly, and promise customers uptime the way the big clouds do. If io.net can sell a guarantee and not just a discount, it can sell to businesses and not just startups.

Until then, io.net stays on my watchlist. When I check back, I will look at two public numbers: how many chips are actually available each day, and how much revenue is being destroyed each month.

This is structural analysis. Nothing here is financial advice.

io.net is a network that rents computing power for AI by pooling idle GPUs from around the world, including former crypto-mining hardware, data-center spares, and gaming PCs, and wiring them into coordinated clusters. It is built on Ray, the same orchestration software that helped train GPT-3 and GPT-4, and settles payments on Solana with its IO token.

The 90 percent figure is io.net's own marketing number, and outside reviews describe a steep discount that varies by chip and job. The bigger caveat is availability: about 327,000 GPUs are listed as verified, while the number ready to join a working cluster on a given day has measured five to seven thousand.

In April 2024, during the snapshot for io.net's token giveaway, attackers exploited a security vulnerability to register roughly 1.8 million fake GPUs and farm rewards. The company acknowledged the failure publicly, fixed the vulnerability, and its founder resigned two days before the token launched.

IO has a fixed maximum supply of 800 million. Customers pay for compute in IO, hardware suppliers earn IO for keeping verified chips online, and holders can stake it. Since June 2026, at least half of the network's revenue is permanently destroyed, tying the token's scarcity to actual paying demand.

Next up: tokenomics — what makes a token worth anything, and how to tell a design connected to the business from one that runs on excitement.