New System Turns Ordinary Office PCs Into an AI Supercomputer

Running an AI chatbot normally means buying racks of expensive, power-hungry servers, plus a central piece of software that acts like an air-traffic controller deciding which machine handles each request. A team of researchers just published a paper describing Cascadia, a system that throws out the air-traffic controller entirely. Instead, the work runs across a fleet of ordinary "AI PCs" — everyday computers that come with a built-in AI chip (called an NPU, a small processor designed for AI tasks). Every machine handles its own scheduling, so there's no single point of failure and no costly central hub.

The machines introduce themselves using digital ID certificates — like showing a badge at the door — then tell each other what they can do and how busy they are. When a request comes in, it goes straight to whichever machine has room. Users can run a model on one computer, spread copies across many, or slice the model into pieces across several machines. There's also a paper trail: every answer comes with a signed receipt and tamper-evident logs, so a business can prove what the AI said and when — handy for audits or legal disputes.

The results are promising but modest in scope. In one test, three mini-PCs working together handled 3.1 times as many simultaneous requests as a single machine. A separate four-machine setup hit 4.06 times the throughput of one machine. That's good, but not a perfect multiplier — machines spend some time coordinating. The researchers also compared Cascadia against commercial platforms from IBM, Nutanix, VMware, and HPE on cost, hardware needs, licensing, and trust, arguing it needs far less specialized gear.

The honest catch: all of this was tested on a small open-source model (Phi-3.5-mini) on clusters of three or four machines. That's nowhere near what a real business runs. Bigger, smarter models may not fit on office PCs at all, and nobody has shown this working across hundreds of machines. Still, the idea is compelling: instead of buying a specialized AI server, you add capacity by buying more of the computers you already own.