Neuromorphic Computing in 2026: What Intel and IBM Have Built, and Why It Still Isn’t Mainstream

3 min read

Neuromorphic computing has been “five years away” from commercial relevance for approximately fifteen years. The hardware has advanced substantially in that time — Intel’s Loihi 2 and IBM’s NorthPole represent genuine engineering achievements, not research demos. What has not advanced at the same pace is the software ecosystem, the programming model, and the market pull required to justify the investment in porting existing AI workloads to fundamentally different hardware architectures. In 2026, the gap between neuromorphic hardware capability and neuromorphic commercial reality remains wide — but the trajectory matters for anyone planning AI infrastructure beyond a three-year horizon.

What Intel’s Loihi 2 and Hala Point Actually Do

Intel’s Loihi 2, released in 2021, is a research chip containing 1 million programmable neurons and 120 million synapses. It processes information using spiking neural networks — a paradigm that mimics biological neural activity, where neurons fire only when inputs cross a threshold rather than continuously. Hala Point, a 2024 research system built from 1,152 Loihi 2 chips, scales this to 1.15 billion neurons and 128 billion synapses. Intel’s benchmark figures show Hala Point solving optimisation problems 50x faster than GPU clusters while consuming a fraction of the energy.

The energy efficiency numbers are real — for the specific workloads where sparse, event-driven computation maps naturally onto the hardware. Continuous auditory processing, sparse temporal data, certain classes of optimisation problems, and some robotics control tasks fit this profile. Standard image classification, transformer-based language models, and most enterprise AI workloads do not. The mismatch is fundamental: neuromorphic chips are designed for sparse, event-driven computation; the dominant AI paradigm runs dense matrix operations on regular data streams.

IBM NorthPole: A Different Architectural Bet

IBM’s NorthPole, detailed in a Science paper in late 2023 and refined through 2024, takes a different approach. Rather than mimicking biological spiking dynamics, NorthPole focuses on eliminating the energy cost of memory access — the dominant power consumer in conventional GPU inference. By distributing compute and memory across a highly parallel on-chip structure with no off-chip memory access during inference, NorthPole achieves 22x better energy efficiency than a comparable GPU on ResNet-50 image classification.

NorthPole is more conventionally programmable than Loihi — it runs quantised neural network inference using standard frameworks with some adaptation. The limitation is size: current NorthPole implementations support models up to a few hundred million parameters. Running a modern large language model, which may have 7–671 billion parameters, is not feasible on current NorthPole architecture. The chip is genuinely compelling for edge inference on medium-sized models, and IBM is pursuing embedded and defence applications accordingly.

Why It Keeps Missing Its Commercial Window

The neuromorphic deployment gap is primarily a software and ecosystem problem, not a hardware problem. Writing applications for spiking neural networks requires learning a programming paradigm with no widely used standard, minimal tooling, and a talent pool measured in the hundreds globally. Intel’s Lava framework and IBM’s toolchain are functional but immature relative to PyTorch and CUDA, which have a decade of investment and millions of trained practitioners.

The second barrier is task compatibility. The workloads where neuromorphic hardware offers genuine advantages — sparse temporal signals, event-driven sensor data, certain optimisation problems — are real but niche relative to the transformer-dominated workloads that drive AI infrastructure investment. Organisations building AI strategy around language models, image generation, or RAG pipelines have little immediate incentive to develop neuromorphic expertise.

Where It Is Actually Being Deployed

Current neuromorphic deployments cluster in three areas: defence and government research (DARPA programs, NATO technology assessments), academic neuroscience and AI safety research, and edge robotics applications where power constraints make energy efficiency the primary design criterion. Intel’s research partners include Accenture, Ericsson, and several national laboratories. IBM’s NorthPole work has attracted defence contractor interest for embedded inference applications.

None of these represent commercial scale adoption. The question for technology planning purposes is not whether neuromorphic chips are commercially relevant today — they are not, in any material sense — but whether the current trajectory suggests relevance within a 5–10 year planning horizon. The hardware is improving faster than the software ecosystem. If Intel or IBM achieves a programming model breakthrough that makes neuromorphic hardware accessible to standard AI practitioners, the energy efficiency advantage becomes a genuine competitive factor as inference costs scale.