AI systems gulp electricity at a pace that alarms data center operators and policy makers alike. One analysis from last year projects global data center power demand more than doubling by 2030 to around 945 terawatt-hours. That figure edges past Japan’s entire annual electricity use. Something has to give.
Enter neuromorphic computing. This approach copies the brain’s trick of mingling memory and processing in the same hardware. No more shuttling data back and forth across a bus. University of York professor Martin Trefzer laid out the case this month before British lawmakers. “Data movement is probably one of the fundamental things we can learn from the brain,” he said. “We don’t have a memory bank on one computer and a processor on the other; it’s all one system, and that is underpinning the efficiency.”
Trefzer spoke to the House of Commons Science, Innovation and Technology Committee on June 18. His testimony, reported by The Register, stressed that biological brains avoid the rigid clocked structure of digital systems. The result? Adaptable hardware that tunes itself for efficiency in specific situations. Yet he tempered expectations. Experimental systems face a mature rival in large language models running in data centers. Those models bring serious energy and sustainability headaches.
Short-term gains will come from hybrid setups. Think hearing aids or other wearables. A neuromorphic substrate could handle sound processing right at the sensor. “This is where there is significant potential to be much more energy efficient, by orders of magnitude,” Trefzer explained. Push the work off the main digital chip. Save power where it counts most — at the edge.
University of Manchester physics professor Caterina Doglioni injected a dose of realism. Building extra devices carries its own carbon cost. “I hate to be the person that breaks it, but you have to think about how much it costs you and the environment to build these devices,” she told the committee. Still, a break-even point exists. Studies must pinpoint it. Only then does the environmental math favor the new approach.
Recent lab results back the optimism. A April 2026 paper from University of Cambridge researchers described a memristive device that combines memory and compute. It reportedly slashes energy use by up to 70 percent compared with conventional chips that constantly shuffle data. The work appeared in ScienceDaily.
Numbers from other projects look even more striking. A framework called NeuEdge, tested on Intel’s Loihi 2 and IBM’s TrueNorth, delivered 847 GOp/s/W efficiency. That translated to 312 times less energy than GPU baselines for certain edge tasks. Inference latency clocked in at 2.3 milliseconds. Accuracy held between 91 and 96 percent across vision and audio benchmarks. Details surfaced in an arXiv preprint posted earlier this year.
IBM’s NorthPole chip stands out too. Released in 2023 and referenced in later analyses, it keeps weights on-chip. No off-chip memory fetches. The payoff appears in orders-of-magnitude better efficiency for targeted inference jobs. A PNAS article from late 2025 noted NorthPole outperforming conventional architectures on several metrics. PNAS explored whether such designs can dent AI’s overall energy appetite.
Commercial players push forward. BrainChip’s Akida processor shows up in radiation-hardened designs for space. A June 2026 video from Frontgrade Gaisler highlighted the integration. Intel continues refining Loihi 2. These chips favor event-driven, sparse computation. Neurons fire only when needed. Power scales with activity, not constant clock ticks. One X post from June 19 summarized the advantage: neuromorphic hardware suits always-on edge inference far better than training heavy models.
Yet challenges remain. Manufacturing at scale. Programming models that differ sharply from standard deep learning frameworks. And the sheer maturity gap versus GPUs and TPUs. A January 2026 project funded at €15 million aims to blend LED-based optical components with silicon for hybrid neuromorphic systems. Proponents claim simultaneous computation could cut energy while boosting speed. Coverage ran in Innovation News Network.
Academic efforts multiply. UT San Antonio secured NSF funding to launch THOR, an open-access neuromorphic hub. Director Dhireesha Kudithipudi called it a national resource to speed real-world applications. The February 2026 announcement signaled growing U.S. government interest. Meanwhile, cryogenic neuromorphic circuits tested near absolute zero could one day link with quantum processors. A June 12 paper in Nature Communications described silicon carbide hardware that functions at 10 millikelvin. ScienceDaily reported the Hong Kong university team’s breakthrough just days ago.
Power measurements tell a consistent story. Some neuromorphic designs reach thousands of giga synaptic operations per watt. Top GPUs hover in the hundreds of gigaflops per watt range for comparable work. The gap widens at low activity factors typical of edge sensing. Two orders of magnitude appear repeatedly in benchmarks. But real deployments must factor in fabrication energy, device lifetime, and integration costs. Doglioni’s caution still applies.
Industry voices on X this week echoed the tension. One developer noted neuromorphic chips won’t displace GPU training clusters anytime soon. They could, however, dominate inference at the edge where battery life and heat matter most. Another highlighted acoustic neuromorphic designs that use sound waves for even lower power. An IEEE Spectrum article from recent days explored the concept. Efficiency claims there reach intriguing levels for specialized tasks.
The path ahead mixes promise with pragmatism. Hybrid systems look likeliest for the next five years. A hearing aid that listens intelligently without draining its battery. Smart sensors that react to rare events without streaming video constantly. Data centers might adopt accelerator cards that offload specific sparse workloads. Full replacement of von Neumann architectures? That sits further out.
Trefzer put it plainly. Neuromorphic ideas should complement existing infrastructure rather than fight it head on. The brain offers lessons on adaptability and locality. Engineers now race to turn those lessons into silicon that scales. Success could ease the AI energy crunch before it becomes unmanageable. Failure would leave the sector doubling down on ever-larger GPU clusters and the power plants needed to feed them.
Either way, the conversation has shifted. Lawmakers hear the testimony. Researchers publish efficiency gains monthly. Companies ship chips into niche markets. The question is no longer whether brain-inspired hardware matters. It is how quickly the industry can make it practical at volume. And how soon those orders-of-magnitude savings reach the applications that consume the most electricity today.
WebProNews is an iEntry Publication