Training AI With Strangers Is Cheaper Than You'd Think, Study Finds

Here's a problem worth caring about: the best AI models need huge amounts of data, but most of that data lives in places that can't share it — hospitals, banks, your phone. A technique called federated learning (training AI across many devices without ever moving the raw data) solves the privacy part. The catch has always been trust and money: if strangers train a model together, who pays whom, and who decides? This new paper looks at using blockchain smart contracts (self-executing digital agreements) to run that marketplace automatically.

The authors built a test marketplace and measured every step of its life, from set-up to training rounds to storage. The headline number: each training task costs about 3.8 million "gas units" per hired participant. Gas is just the fee you pay a blockchain to do computation. That sounds enormous, but the interesting finding is the shape of the bill. Setting up the marketplace is the expensive part; the recurring rounds are 10 to 100 times cheaper. So costs spread out and shrink over time, much like a gym membership getting better value the more you show up.

How long until it pays off? About 20 training rounds — the point where the average cost per round drops to twice the rock-bottom ongoing rate. After that, you're mostly paying small change. Importantly, adding blockchain coordination and decentralized file storage didn't hurt the AI: accuracy was the same as a standard setup without any of it.

For everyday people, this points at a future where your data — health records, photos, driving patterns — can improve AI without ever leaving your hands, and where you might get paid for it. It also means companies that currently can't collaborate for legal reasons could share models instead of files. The honest caveat: this is one simulation with specific assumptions, and blockchain fees in the real world swing wildly with network traffic. Cheap in theory still has to survive a busy Tuesday on Ethereum.