Researchers Want a Million AI Chips to Act Like One Big Brain

A single researcher, Heng Liao, posted a seven-page paper this month laying out a new idea for how to build the enormous computers that train AI. It is theory only — no chip was built, no product was announced. The paper makes two linked proposals: a software rule for coordinating millions of processors, and a hardware blueprint for wiring them together through one shared pathway called a "Unified Bus." Think of it as a blueprint for a factory, not the factory itself.

The problem it tackles is simple to state. Companies like Nvidia and Huawei now glue thousands of chips together to train AI, but those chips still act like a crowd of strangers, not one machine. The paper argues that an old habit is to blame: almost every design assumes a boss chip giving orders and worker chips obeying — a CPU commanding an accelerator, a data center commanding distant devices. Networking many computers together, the author says, does not actually give you one larger computer.

The proposed fix is nesting, like Russian dolls. At every level — from a single package up to a whole "autonomous zone" of machines — the same four steps repeat: work in parallel, everyone pauses at a checkpoint to swap notes, combine results, then move to the next phase. That checkpoint idea (known as a barrier) is already familiar in supercomputing. The twist is making every node at every layer a peer, with no masters. The author calls it "physically sparse, logically tight." The paper also ties into Huawei's tau Scaling law, a research idea about how each layer folds time.

The honest catch: this is one author's proposal with three diagrams and no working hardware, no tests measuring real speed, and no proof it beats what already exists. Turning it into chips and software would take years and enormous money. Still, the direction matters. If approaches like this eventually work, training AI could get faster and cheaper — which means the AI tools you use might get better, or cost less, sooner.