Neuromorphic processors can perform an impressive trick: while nothing significant is happening around them, large portions of the circuit stay almost silent. When a signal arrives, the relevant artificial neurons activate, a brief stream of spikes passes through, and the computation ends. Compared with today’s AI accelerators that continuously crunch enormous arrays of numbers, the approach appears almost suspiciously efficient.
This efficiency raises an obvious question. If neuromorphic chips can save power, keep state next to the compute units, and even rewire connections during operation, why are data centers still filled with GPUs? Why can’t a large language model simply be moved to a brain-like processor and slash electricity bills?
The answer is straightforward. Neuromorphic architecture excels at a specific class of problems, but today’s generative AI has grown around an entirely different kind of mathematics. Large models favor dense matrix operations, enormous memory capacity, high numerical precision, and mature development tools. Neuromorphic systems must either alter their own architecture or fundamentally change the model itself. Researchers are currently pursuing both routes.
Why neuromorphic chips have not yet gone mainstream
The bottleneck is no longer the lack of working silicon. Intel has produced two generations of Loihi and the large-scale Hala Point system. BrainChip ships the Akida family. SynSense sells specialized processors for vision and temporal signals. SpiNNaker2 enables large event-driven systems. Yet a compelling advantage for mass-market workloads that would justify switching to an entirely new computing model is still missing.
GPUs have accumulated a vast software ecosystem over the past decade. A developer can take a ready model, library, driver, and profiler and run familiar code. For a neuromorphic processor, engineers often must select a different network architecture, convert the model, account for spike timing, and re-validate accuracy.
On-chip learning and synaptic plasticity
Neuromorphic processors can support local weight changes. The device receives new events and gradually adapts part of the network without shipping all data back to a central facility. Intel Loihi 2 was designed in part to study programmable on-chip learning. In 2025 researchers demonstrated contextual learning of a transformer on Loihi 2 using local plasticity for attention-like operations.
One well-known local rule is STDP (Spike-Timing-Dependent Plasticity). If one neuron fires just before another, the connection between them can strengthen; the reverse timing weakens it. This rule lets each synapse adapt using only nearby activity and precise timing, without waiting for a global gradient computation. However, simple local rules still fall short of the global optimization that gives today’s deep networks their accuracy.
Running large language models on neuromorphic hardware
At the 2025 ICML conference, researchers presented Sorbet, a spiking language model based on the transformer architecture. They replaced operations poorly suited to neuromorphic hardware, such as standard normalization and Softmax. Another 2025 effort converted a 1.5-billion-parameter model into spiking form, used sparse linear attention, and reported 161.8 tokens per second on an FPGA with markedly better energy efficiency than an A800 GPU. Direct comparison with production LLM stacks remains premature.
Future computing landscape
Complete replacement of GPUs by neuromorphic processors is considered unlikely because the two architectures optimize different properties. GPUs thrive on sustained high utilization of thousands of arithmetic units; neuromorphic systems gain advantage from sparsity. The most probable outcome is a heterogeneous platform in which neuromorphic blocks monitor rare events, matrix accelerators handle heavy generative workloads, in-memory and analog units reduce data movement, and photonic elements manage extreme parallelism. The CPU would continue to orchestrate tasks that do not benefit from specialized hardware.
