Most discussions about the future of artificial intelligence converge on the same bottleneck: energy. Training a large language model consumes enough electricity for a small city for days. Running inference at scale isn't much different. Meanwhile, the human brain processes complex language, vision and decision patterns while consuming about 20 watts — less than an incandescent light bulb. This discrepancy is not academic curiosity; is the most important engineering problem facing the hardware industry.
Neuromorphic computing is one of the most serious bets to address it. Not because it will replace the GPUs that run current models, but because it solves a different category of problem, with an efficiency that the von Neumann architecture — the one that has governed practically every computer since the 1940s — cannot achieve.
What makes the brain different from a GPU
A GPU processes in a dense and synchronous way: all cores work in parallel, the clock pulses continuously, energy is consumed regardless of whether there is useful computation happening or not. This is great for batch linear algebra — which is exactly what deep neural networks do during training.
The brain works in a radically different way. Neurons fire in response to events, not clock cycles. Most of the time, most neurons are silent. When a stimulus arrives, only the relevant subset activates — sparsely, asynchronously, and with an energy consumption proportional to the work done, not the time elapsed.
Neuromorphic chips attempt to reproduce this architecture in silicon. Intel Loihi 2, for example, implements artificial neurons and synapses that behave in a similar way: event-based firing, local memory processing and computation in the same element, energy consumption only when there is activity. IBM TrueNorth followed a similar principle, with 4096 neuromorphic cores running at a few milliwatts. They are not fast chips in the conventional sense; They're efficient chips in a way that CPUs and GPUs simply aren't.
Where this technology makes sense today
In the real world, neuromorphic chips now outperform the GPU in edge continuous sensing problems, with low latency and minimal power.
Think about industrial sensors that need to detect anomalies in data streams in real time, without sending everything to the cloud. Or IoT devices that need to recognize audio or vibration patterns for months on a small battery. Or vision systems embedded in drones that need to react in milliseconds without a power cord nearby. In these scenarios, energy efficiency is the variable that governs viability — and neuromorphic architectures win by a margin that improves with each chip generation.
What these chips don't do well, for now, is train large models or run dense language inference. For this, the GPU still reigns. The category is complementary, not a substitute.
DNA computing and the biological frontier
Beyond silicon, there is an even more radical frontier: using biological molecules as a computing substrate. DNA computing is not a metaphor — it is literal computing with strands of DNA, where data is encoded in sequences of bases and molecular reactions perform logical operations.
The motivation is absurd storage density: one gram of DNA can theoretically store exabytes of data. And durability: Intact DNA has been recovered from samples tens of thousands of years old. Microsoft and others have already demonstrated functional data storage and retrieval in synthetic DNA. It's not a product yet, but it's already an advanced laboratory.
Analog computing based on molecular circuits and memristors — components that combine memory and processing in the same physical element — is also advancing at a relevant pace. The pattern that emerges is that the rigid division between where data is stored and where it is processed, which governs all conventional architecture, begins to be questioned at the physical level.
How a leader should look at this
The easiest strategic mistake to make here is to treat neuromorphic computing as a research curiosity that will only matter in ten years. The horizon of operational relevance is shorter than it seems — especially for companies with a presence in edge, manufacturing, logistics, healthcare, or any domain where embedded sensors and devices make local decisions.
Adopting neuromorphics now is not the point. The work is to understand when the equation changes for your use case. If you're paying for cloud inference to process data that a sensor could analyze locally with a specialized chip, the conversation about energy efficiency and latency is already relevant. If you are building a product for markets where energy consumption and connectivity costs are real restrictions — and in Brazil this is a large part of the industrial and agricultural market — this technology enters the radar sooner than you think.
What leaders must do today is maintain an intelligence channel about the sector without compromising production budgets on it. Keep up with what Intel, IBM, and startups like BrainChip and Innatera are launching. Maintain an internal engineer or researcher who monitors the area. When the tipping point arrives—and it will, because physics favors efficiency—you want to be one learning cycle ahead, not starting from scratch.
Also read
- Inference chips and ASIC: when specialized beats generic
- Energy efficient AI: Smaller and cheaper beats bigger
- Specialized chips and the end of the generic CPU era
- AI PCs and NPU: what changes when the device gains AI accelerator
- How to build a business case for AI in industrial operations
- AI Energy and Sovereignty: What Governments Need to Plan Now