SWNFT
AI Crypto

What Is Decentralized AI? Who Runs the Computers

What is decentralized AI? A plain-English guide to who runs the computers, how the pieces fit, what's proven so far, and what's still hard. Not advice.

James Killick6 min read

TL;DR

  • Decentralized AI means no single company runs the data, the computers and the model. Many people share those jobs.
  • Big AI today leans on a few cloud firms. A US regulator's staff flagged concerns about that in January 2025.
  • Real pieces exist: private training on phones (federated learning), a 10-billion-parameter model trained across three continents, and crypto marketplaces for data and computing power.
  • The hard parts are proving the work was done honestly and keeping it fast.
  • Many crypto tokens tied to this idea lose value. This is education, not advice.

Decentralized AI means spreading the building and running of AI across many people and computers, not one company. Different groups can supply the data, the computing power and the testing. The goal is that no single firm controls the whole thing. If you're curious about crypto and AI, this is the idea behind a lot of the hype.

Risk Notice: Many tokens tied to AI ideas lose value. Scams that borrow AI buzzwords are common. This guide explains the ideas and what's been shown so far. It doesn't tell you to buy any token. Our guide to how crypto scams work is a good first read.

Why do people want decentralized AI?

Many leading AI developers lean on a few large cloud companies for computing power. That worries some people.

In January 2025, the US Federal Trade Commission's staff published a report on AI partnerships. It looked at three deals: Microsoft with OpenAI, and Amazon and Google with Anthropic. It flagged things to watch. These were effects on access to computing resources, higher costs to switch cloud providers, and cloud firms seeing sensitive information.

Be fair about what it said. The report doesn't find that the deals broke the law or caused harm. It's a staff report, and two commissioners disagreed with parts of it. But it shows why people ask the question. Who controls AI, and does it matter?

That's the pitch for the decentralised idea. More players, less reliance on a few.

What does decentralized AI actually mean?

There are three layers. Each one can be shared out on its own.

| Layer | The question | Example approach | |---|---|---| | Data | Who holds the training data? | Keep it with the owner, like in federated learning | | Compute | Who owns the computers? | Spare machines rented through a network | | Model | Who judges the output, and who owns it? | Open networks that score and reward models |

Decentralised doesn't mean "open source". An open model can still be trained and hosted by one company. And a decentralised network doesn't always use a blockchain. That's the next bit.

How does federated learning keep data private?

This is the oldest real piece. In 2016, researchers from Google proposed federated learning. The training data stays on phones. It doesn't go to a data centre. Each device works out a small update. A central system averages those updates into one shared model.

It's decentralised on the data side. The raw data never leaves the device. It doesn't need crypto at all.

Ocean Protocol takes a different route to a similar goal. Owners keep their data. They rent out the right to run approved programs on it. Our Ocean Protocol explainer walks through that.

Can you train an AI model on computers all over the world?

A team called Prime Intellect gave one answer. In December 2024 they posted a technical report on INTELLECT-1. It was a 10-billion-parameter language model trained on 1 trillion tokens, across three continents. The report says up to 14 machines worked at once. About 30 independent computing providers joined and left during the run.

It's a proof of concept. It shows large training can run on scattered, unreliable machines. It doesn't show that this beats a big data centre on cost or speed. Be careful with anyone who says it does.

The compute side of crypto uses a similar idea. Projects like Akash let providers bid to rent out computers, including GPUs. See our guide to Akash Network.

Where does crypto come into decentralized AI?

Crypto adds two things. It's a way to pay strangers. And it's a shared record of who did what.

Bittensor is one example. It's a network where subnets reward people for useful AI work, and validators score that work. Read how Bittensor works for the full picture.

Notice what the token does here. It's a payment and incentive tool for strangers who don't trust each other. It's not proof that the AI is any good. A token can rise or fall with no link to how well the AI works. That gap is where most of the hype lives.

What about AI agents?

AI agents are programs that work toward a goal in steps, using tools. Some projects imagine agents that buy computing power or data from these networks on their own. That's early and mostly untested. Want a plain explainer on agents themselves? See this simple guide to what AI agents are.

What's still hard about decentralized AI?

Two big problems stand out.

Proving honest work. A stranger says they trained your model or ran your job. How do you know? In 2021, researchers introduced proof-of-learning. It's a way for a party to show that a model really came from the claimed training. The paper looks at ownership disputes. It also looks at protecting distributed training from untrusted workers. It's a research idea. Verifying work at scale is still an open problem.

Speed. Home internet is far slower than the links inside a data centre. Teams like Prime Intellect use tricks to cut how much data has to move. That helps, but it adds complexity.

Then add the money risk. Many tokens tied to hype cycles have lost most of their value. A good idea doesn't make a good investment.

Frequently Asked Questions

What is decentralized AI in simple terms?

It's AI built and run by many people, not one company. Data, computing power and testing can all be shared out.

Who runs the computers in decentralized AI?

Independent owners. They might be data centres, small operators or individuals with spare GPUs. Networks like Akash and Bittensor connect them to people who need AI work done.

Is decentralized AI the same as blockchain AI?

Not always. Federated learning is decentralised and uses no blockchain. Crypto networks add payments and a shared record on top.

Is decentralized AI better than big-company AI?

It's not proven. It can offer more choice and privacy options. I think big data centres are still faster and simpler for most jobs. But that's a judgement call.

Not financial advice

This is general education. It isn't financial advice. It doesn't recommend any token or project. Most tokens lose value. Only risk money you can afford to lose.

What to do next

Read our AI crypto guide. It maps the projects in this space, so you can tell the real ones from the noise.

About the Author

James Killick

AI Product, Web3 & Sales Automation Specialist

James Killick runs this site →

Digital Product Development

Devwiz is a Sydney-based digital product development agency specialising in launching and scaling SaaS products, web apps and mobile apps. James is a shareholder and advisor, shaping product strategy and PLG approach.

Visit Devwiz

// Related

Keep Reading