CyberSwarm has built a neuromorphic processor called NPU One, working with researchers from Magurele after four or five years on memristors. The team turned those components into artificial synapses and then into neurons, aiming at a processor with plasticity: one that can learn, and also forget, over time.
A single neuromorphic chip can hold more than seventy thousand algorithms.
The energy comparison is the point. A biological brain runs on at most forty watts, and the artificial one is built to sit in the same range. A ChatGPT query, by contrast, consumes fifteen or sixteen times what a Google search does.
Binary computing exists for historical reasons rather than fundamental ones: it was what the available hardware could do. The architecture that followed has limits on how much can genuinely happen at once, which is what neuromorphic design sets out to escape.
The chips can be laid down on existing silicon or on materials close to those used for phone screens. That is what lets a device keep learning the way a child does, improving over time, with no connection to the internet.
An artificial brain is not the same thing as artificial intelligence. The mechanisms around the processor are what make the processing work.
“The light bulb did not emerge as a continuous improvement of the candle,” Raneti says. “For that reason, we have to dare more.” Romania, he adds, must look upward and stop saying that certain technologies are not ours.
Development at the time was spread across Ploiesti, Magurele, Taiwan and the United States.