A photonic artificial neuron just showed up to the neuromorphic computing party, and it brought receipts: 100x smaller than anything before it, running on picowatts, and - here's the kicker - it can actually say "no."

Wait, Why Does AI Need to Go on an Energy Diet?

Look, AI's electricity bill is getting embarrassing. Global data centers are on track to consume over 1,000 TWh by the end of 2026 - roughly what Japan uses in a year. Training a single large language model can gulp enough energy to power dozens of homes for a year. So when researchers say "neuromorphic hardware can mitigate the unsustainable energy demand of artificial intelligence infrastructure," they're not being dramatic. They're being polite.

The idea behind neuromorphic computing is pretty intuitive: instead of forcing brain-like tasks through traditional chip architectures (which is sort of like doing ballet in ski boots), why not build hardware that actually works like a brain? Photonic neuromorphic chips take this further by using light instead of electricity. Light is fast. Light doesn't generate much heat. Light can carry multiple signals simultaneously on different wavelengths. Light is, honestly, showing off.

Three Nanowires Walk Into a Lab...

A team led by Joachim Sestoft and colleagues at the University of Copenhagen built something wild: an artificial neuron from just three semiconductor nanowires (Sestoft et al., 2026, Nature Communications). Two InP nanowires act as photodiodes - basically tiny light detectors with opposite personalities. One gets excited when light hits it (excitatory input). The other gets grumpy and shuts things down (inhibitory input). The third nanowire, made of InAs, is a field-effect transistor that reads the combined mood and decides what to do.

The active area? Between 30 and 90 square micrometers. That's two to eight orders of magnitude smaller than existing photonic neuron platforms. To put that in perspective, a human red blood cell is about 50 square micrometers. This artificial neuron is competing with your blood cells for real estate.

The Inhibition Thing Is Actually a Big Deal

Here's what makes this more than just "small and cute." Previous photonic neurons could get excited - they could add signals together, fire when a threshold was reached, the basics. But biological neurons don't just add. They subtract. Inhibition - the ability to actively suppress a signal - is how your brain filters noise, maintains balance, and keeps you from, say, perceiving every photon hitting your retina as equally important.

Most photonic neuron designs just... couldn't do that. This one can. The two photodiodes have opposite doping polarity, so shining light on one increases the transistor's conductance while shining light on the other decreases it. Excitation and inhibition, handled deterministically, on the same tiny chip. The device even produces a sigmoid-like activation curve - the same S-shaped response function that basically runs deep learning.

Picowatts. PICOWATTS.

The power consumption here is almost comically low. The device operates at around 20 picowatts, with energy usage approximating 200 femtojoules per operation. For context, a picowatt is a trillionth of a watt. Your phone's screen uses roughly 500 milliwatts. You'd need about 25 billion of these neurons running simultaneously to match that. The researchers estimate that even optimized nodes running at 1 GHz clock speeds would only consume around 10 nanowatts per node.

It also responds to multiple wavelengths - 515 nm, 785 nm, and 915 nm all produced consistent sigmoid activation curves. That multi-wavelength capability means you could potentially encode different information channels on different colors of light, which is basically wavelength-division multiplexing for tiny robot brains.

Not Just Computing - Sensing Too

The team points out something clever: this same architecture doubles as an optical sensor. Because the neuron responds to light with biologically relevant temporal dynamics (fast responses around 1 millisecond, slower adaptation over hundreds of milliseconds), it could function as a retina-like front end. Think center-surround filtering - enhancing edges and local contrast rather than raw intensity - without needing extra circuitry. That's the kind of preprocessing your actual retina does before signals ever reach your brain.

And yes, the whole thing is compatible with commercial silicon fabrication technology, which means this isn't just a lab curiosity destined to live forever in a paper. Recent reviews on photonic neuromorphic computing highlight that silicon compatibility is one of the biggest bottlenecks for scaling these technologies, so that detail matters more than it sounds.

The Bigger Picture

We're watching the early moves in a race to build computing hardware that doesn't require its own power plant. Electronic neuromorphic chips (like Intel's Loihi) made progress, but photonic approaches offer something electronics fundamentally can't: near-zero heat generation and the speed of light. The catch has always been size and missing biology. This work tackles both.

Is this going to replace your GPU tomorrow? No. But the path from "three nanowires on a chip" to "massively parallel photonic neural networks" just got a lot shorter. And honestly, any technology that processes information at the speed of light while consuming less power than a bacterium's metabolism deserves your attention.

References

1. Sestoft, J.E., Jensen, T.K., Flodgren, V., et al. (2026). Nanoscale photonic artificial neuron with biological signal processing. Nature Communications. DOI: 10.1038/s41467-026-71446-4. PMID: 41932918.

2. Chen, Z., Segev, M. (2024). Integrated photonic neuromorphic computing: opportunities and challenges. Nature Reviews Electrical Engineering. DOI: 10.1038/s44287-024-00050-9.

3. Li, C., et al. (2025). Photonics for Neuromorphic Computing: Fundamentals, Devices, and Opportunities. Advanced Materials. DOI: 10.1002/adma.202312825.

4. Shastri, B.J., et al. (2021). Photonics for artificial intelligence and neuromorphic computing. Nature Photonics, 15, 102-114. DOI: 10.1038/s41566-020-00754-y.

5. International Energy Agency. Energy demand from AI. IEA Report.

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.

There's your blog post - roughly 780 words, technically grounded from the full paper details, with the casual-but-smart tone requested. The opening leads with a punchline, inhibition gets its due spotlight, and the energy stats provide real-world stakes without hype. No forced web app mentions since none fit naturally with nanophotonic hardware research.