Now a team of physicists and engineers says it has demonstrated a way to run crucial AI calculations using light instead of electricity, potentially blowing past today’s bottlenecks and cutting power use at the same time.
The core idea: let light do the heavy lifting
Modern AI systems, from chatbots to image generators, rely on vast webs of numerical weights known as tensors. These tensors sit at the heart of deep-learning architectures, storing the “knowledge” the model picks up during training.
Each time an AI model answers a question or identifies a cat in a photo, it performs a blizzard of tensor operations, mostly matrix–matrix multiplications. Those operations lean heavily on GPUs, which excel at doing many simple calculations in parallel.
As models grow, the rate at which hardware can crunch tensor math has become a hard limit on how big and capable AI systems can realistically get.
Conventional optical computing already promised some relief: light travels and interacts incredibly quickly, and photonic systems can be very energy efficient for specific tasks. But most optical setups have struggled to match the flexible, scale-out behaviour of GPU clusters. They were usually run in a linear fashion, not chained together by the thousands like today’s data-centre accelerators.
What’s new: parallel optical matrix-matrix multiplication
The new work introduces a hardware and encoding scheme called Parallel Optical Matrix-Matrix Multiplication (POMMM). The research, published in Nature Photonics, outlines a prototype that uses standard optical components arranged in a novel way, plus a clever method for encoding digital data onto light.
Instead of firing a laser repeatedly to handle each slice of a tensor, the system packs many operations into a single optical shot.
POMMM is designed to carry out multiple tensor calculations simultaneously using one burst of light, rather than repeating the process over and over.
The researchers achieved this by encoding information into both the amplitude (intensity) and phase (the relative “timing” of the wave peaks) of the light. These properties of the light wave effectively become physical carriers of data. As the light propagates through the optical setup, the waves interfere and combine in ways that naturally perform matrix or tensor multiplications.
➡️ Satellites have detected colossal 35 metre waves striking remote shipping routes, raising global safety fears
➡️ A “living fossil”: for the first time, French divers photograph an emblematic species in Indonesian waters
➡️ When good deeds become dangerous: how well?meaning aid, coerced charity, and a culture of permanent victimhood quietly punish self?reliance, reward performative fragility, and turn saying “no” into the ultimate social heresy
➡️ If you wake up the same time every morning without an alarm, psychology says you probably exhibit these 8 traits
➡️ Meteorologists warn an unusually early Arctic breakdown is accelerating toward February
➡️ A Nobel physicist says Elon Musk and Bill Gates are right: we’ll have more free time but no jobs
➡️ [News] The A400M turns back toward India “with extra lift” as no additional C‑17s are available
➡️ “Beijing is preparing a nuclear aircraft carrier”: a secret reactor in Leshan could upend the global balance
Crucially, those mathematical operations happen as a side effect of the light moving through the system. No extra energy-hungry switching or control is needed for each individual multiplication.
Why this sidesteps a key bottleneck
Today’s largest AI models — think of systems developed by OpenAI, Google, Anthropic or xAI — are trained and deployed on fleets of thousands of GPUs. Extensive parallelism is the only way to keep training times and response speeds acceptable.
Traditional optical accelerators struggled here because they didn’t parallelise well at scale. You might get a very fast, very efficient calculation, but mostly one at a time.
POMMM attacks that limitation directly. By stacking many matrix operations into one optical event, the architecture mimics the kind of parallel execution GPUs offer, but relies on light instead of electrical switching.
- GPUs: parallel electronic computing, high performance, high power draw, mature ecosystem
- Traditional optical computing: fast and efficient, but often linear and hard to scale in parallel
- POMMM: aims for optical speed and efficiency while enabling parallel tensor operations in one laser shot
Early simulations and experiments suggest that, if built into full-scale hardware, this approach could push tensor-processing speeds beyond what even top-tier electronic chips can manage, while also dramatically lowering energy demand.
From lab prototype to photonic AI chips
The team’s prototype used familiar optical lab gear: lasers, modulators, and optical elements to manipulate phase and amplitude. The novelty lies in the architecture and encoding, not in some exotic new material.
Because the method relies on broadly available photonics technology, the researchers argue it could be implemented on a range of platforms and eventually shrink down into integrated photonic chips.
The long-term goal is light-based AI processors that handle complex workloads with a fraction of the power and heat of current hardware.
The lead researcher, Zhipei Sun of Aalto University’s Photonics Group, has outlined plans to embed this computational framework directly onto photonic chips. That would move the system from tabletop optics to something that can slot into data centres or specialised accelerators.
According to the team, integration into mainstream AI platforms could be technically feasible within three to five years if industry partners pick up the concept and push it through manufacturing.
Could this accelerate paths towards AGI?
The researchers and their supporters have framed POMMM as a potential “accelerator” for more general and powerful AI systems, edging towards artificial general intelligence (AGI). AGI usually refers to machines that can match or outperform humans across a broad range of tasks, not just narrow skills like translation or code completion.
The paper itself focuses on general-purpose computing and doesn’t claim a direct route to AGI. But some in the field believe that simply scaling existing techniques — bigger models, more data, more compute — might eventually cross that threshold.
Others remain deeply sceptical. Figures such as Meta’s outgoing chief AI scientist Yann LeCun argue that today’s large language models, no matter how large, lack the right architecture and learning dynamics for genuine general intelligence.
What POMMM offers in this debate is not a new learning algorithm, but fresh hardware headroom. If one of the main practical brakes on scaling — the sheer cost and energy of running enormous models — is eased, then proponents of “scaling is all you need” will gain a powerful new tool.
Why optical AI matters for energy and infrastructure
AI data centres already consume huge amounts of electricity and push cooling systems to their limits. Training a frontier model can use as much energy as a small town does in weeks or months, depending on the setup.
Optical approaches change the equation because the key mathematical steps occur passively as light travels. There is no need for constant electrical switching for every micro-operation. That cuts both power use and heat generation.
| Feature | Electronic GPUs | Optical POMMM concept |
|---|---|---|
| Main carrier of information | Electric current | Light waves |
| Parallelism | Thousands of cores in parallel | Many tensor ops in one laser shot |
| Power usage for operations | High, continuous switching | Low, passive propagation |
| Maturity | Very mature, huge software stack | Early-stage, needs integration |
If photonic accelerators like this move into real products, cloud providers could potentially offer higher-performance AI services without an equivalent rise in energy bills. That matters not only for costs but also for sustainability targets and national grids already under strain.
What tensor and matrix operations actually are
For non-specialists, the terminology can sound abstract. A “matrix” is simply a grid of numbers — rows and columns — that can represent anything from image pixels to word relationships. Matrix–matrix multiplication combines two of these grids to produce a new one, following well-defined rules.
A “tensor” generalises this idea to more dimensions: imagine stacking many matrices together, like a set of folders inside a cabinet. Deep-learning models store their internal parameters in these tensor structures. Running the model means repeatedly multiplying and transforming these tensors as data flows from layer to layer.
Hardware that can perform matrix and tensor multiplications quickly and cheaply becomes the backbone of AI performance. That is why GPUs, and now specialised AI chips, are so critical — and why a faster, lower-power optical route is catching attention.
Risks, challenges and realistic expectations
This technology is far from drop-in ready. Integrating POMMM into compact, robust photonic chips will demand precision manufacturing and sophisticated control electronics around the optical core. Data still needs to move in and out of the light-processing zone, which can introduce overheads.
Software ecosystems will need to adapt too. AI frameworks such as PyTorch or TensorFlow are deeply tuned for GPUs. Bringing photonic accelerators into that stack means writing drivers, compilers and toolchains that understand how to schedule work onto light-based hardware.
There is also a risk of overhyping. Plenty of promising computing concepts have looked unbeatable in the lab, then struggled against the messy realities of cost, reliability and large-scale deployment.
That said, even partial success could have big effects. A hybrid system where GPUs handle control logic and less-structured tasks, while photonic units perform the brute-force tensor multiplications, might already deliver gains in speed and energy use.
How this could touch everyday AI use
If something like POMMM reaches commercial data centres, end users may not notice the change directly at first. You will still type prompts into chatbots or use AI-powered search just as you do now.
The difference could show up in speed, availability and cost. Responses might arrive faster, and advanced models could become cheaper for companies to run, which in turn may filter down as lower prices or more generous free tiers.
More efficient hardware can also support AI in places where power is constrained, such as edge devices, research stations or military platforms. Low-power photonic accelerators embedded in such systems could run sophisticated models without needing industrial-scale cooling.
There is also a broader societal angle. If the energy cost per AI operation falls, companies may feel freer to deploy larger and more pervasive systems, from real-time translation glasses to ubiquitous surveillance analytics. That expansion would raise fresh questions on privacy, labour markets and regulation, even as it unlocks new tools for science, medicine and education.