New Trick Cuts Cloud Costs 56% Using 'Borrowed' Servers

Cloud computing has a bargain bin. Companies like Amazon and Google rent out spare servers at steep discounts — often 60-90% off — but with a catch: the server can be yanked back with little warning. Most businesses avoid them, because if a program is halfway through a job when the machine disappears, the work is gone. So they pay full price for reliability.

A team at EPFL in Switzerland has a fix. Their system, ThreadShift, sends only the busiest parts of a program to these cheap servers, and constantly saves its progress in a way that's so fast the program barely notices. When a server is snatched away, the work simply continues elsewhere. The old saving method, called CRIU, was so slow it ate up the savings. ThreadShift's approach is roughly a thousand times faster — it can hand off a piece of work in 164 microseconds, faster than a camera flash.

Crucially, ordinary software doesn't need to be rewritten. That matters because rewriting code is expensive and slow, which is exactly why most companies never bothered with cheap servers in the first place.

The researchers tested it on machine-learning inference (the kind of AI that answers your questions), encryption tasks, and data crunching. In all cases, costs dropped by up to 56%. For you, that's not abstract: when AI companies pay less to run their models, the price of AI tools, and the ads or subscriptions funding them, can fall too. Cloud bills are the single biggest expense for most AI startups.

The catch: this is a research prototype, not a product you can buy today. It only works on certain Linux servers with specific chips, so widespread adoption will take time. And cheaper infrastructure doesn't guarantee companies pass the savings along.