New Scheduler Makes AI Data Centers Waste 30% Less Computing Power

Every time you use an AI chatbot, generate an image, or get a movie recommendation, your request lands on a specialized computer chip inside a data center. These chips are the expensive, hard-to-get engines of the AI boom. The problem: they sit idle far more often than you'd expect, even while companies are waiting in long queues for their turn. The reason is messy coordination — the rules for who gets priority, who can borrow extra capacity, and who has to wait are all managed separately, and they contradict each other.

A team of researchers from Chinese universities (with collaborators) built a fix called DeepShare. Their key idea is simple to describe: instead of five separate systems making five separate decisions, one shared signal called "tenant assurance" — basically, how well each customer's promises are currently being kept — guides everything. If a customer is comfortably on track, their spare chips get lent to someone else. If they start slipping, the borrowed capacity gets pulled back automatically. DeepShare also predicts how long each job will actually take, safely pauses less urgent work, and lets multiple jobs share a single chip without slowing each other down.

The results are striking. Testing on nearly 24,000 real jobs from a public dataset plus 3,200 internal jobs, DeepShare kept chips busy about 70.6% of the time — a 29.5% jump over the best existing method. Waiting time dropped 46%. On a live 16-chip test system, jobs finished 34% faster, and paying customers still got what they were promised 93% of the time.

Why should you care? AI companies spend enormous sums renting these chips, and those costs get passed to you as subscription fees. Wasted chips also mean wasted electricity. Squeezing more real work out of the same hardware is one of the few ways to make AI cheaper and greener without building anything new. This is a research paper, not a product you can buy today — but it's the kind of plumbing that quietly decides how much AI costs you in a year or two.