The idea of computing with light has been around for over 70 years, but until recently it has remained largely in the realm of academic research. Over the past few years, however, optical computing has emerged as a fast-growing commercial industry. Market valuation of companies in this sector now totals US $2.22 billion, up sharply from $0.4 billion in 2022, with industry projections suggesting that the market will continue to grow exponentially in the coming years.
The reason for this sudden surge will surprise no one: artificial intelligence. In particular, AI inference—the process of using AI models to respond to user queries—has driven much of the construction of huge, power-hungry data centers. Relatively energy-efficient optical systems are increasingly seen as a potential way forward to mitigate this staggering energy demand.
"Most of the interest in photonics is driven by the fact that AI inference has literally become limited by power, and in a rather striking manner," says Peter McMahon, associate professor of applied and engineering physics at Cornell University in Ithaca, N.Y., who has researched optical computing for over a decade.
Optics, at least on paper, carries the promise of executing the same heavy mathematical workloads as today’s GPUs at a drastically lower level of power consumption. Switching from electronics to optics would be a radical architectural solution. Nevertheless, like mushrooms after a rainstorm, optical-computing startups are popping up across the United States, China, Germany, and Israel. And while competing directly against established GPU makers like Nvidia is an ambitious goal, capital investment in optical computing has spurred the development of complementary approaches. These include designing alternative AI algorithms, taking on demanding scientific computing tasks, and enhancing in-network computing architectures.
"I think it would be good for the technology if many of these companies made a good run at it," says Logan Wright, assistant professor of applied physics at Yale University in New Haven, Conn., and a former postdoc in McMahon’s lab. "I think that the technology that will be developed through this would be really, really useful, and enable photonic systems that maybe don’t get on this wave of AI but would be capable of doing really quite special things."
Optics to Inference’s Rescue?
Inference makes up the vast majority of AI operational energy use. The tech industry is already spending billions of dollars designing advanced GPUs and other specialized silicon chips to run inference as efficiently as possible. At the heart of inference computation lies a single operation, repeated countless times every second: matrix-vector multiplication.
Back in 1978, optical-computing pioneer Joseph W. Goodman and his coworkers proposed a method for performing matrix-vector multiplication using light. The setup was relatively straightforward. Imagine a vector of length $N$ that needs to be multiplied by a matrix of size $N$ by $N$. A row of $N$ light-emitting diodes encodes the elements of the vector into the intensity of light—the brighter the light from an individual diode, the greater the numerical value that diode encodes. Then, a cylindrical lens spreads out the light vertically, essentially copying the vector $N$ times. The light subsequently passes through an attenuator containing $N$ by $N$ pixels. At each pixel, the light is attenuated according to the corresponding value at that location in the matrix. Passing the light through this attenuator performs the multiplication. Finally, a cylindrical lens oriented perpendicular to the first lens re-collects the light from each row, summing it up, and the final output vector is measured by a row of detectors.
The argument for why optics might hold a profound advantage over electronics goes like this: Multiplying a matrix of size $N$ by $N$ by a vector of length $N$ takes $N^2$ multiplication steps. In traditional electronics, each individual multiplication consumes energy, so the total energy used scales proportionally to $N^2$. In optics, however, the multiplication step is passive. As long as the matrix stays the same, passing light through the attenuator costs almost zero energy. The only energy cost is incurred when encoding the vector into light and reading out the final answer. Because the vector is of length $N$ rather than $N^2$, the overall energy used scales as $N$ rather than $N^2$ as increasingly large vectors and matrices are multiplied.

Since Goodman’s foundational 1978 proposal, the quality of available optical elements has improved dramatically. In 2022, a team of researchers at Cornell led by McMahon demonstrated an optical neural network that consumed less than one photon of energy—approximately $2.5 times 10^-19$ joules—per multiplication operation. In a standard electronic computer, a single multiply operation can take between $10^-14$ and $10^-12$ joules, suggesting that optics could theoretically reduce energy use by a factor of 100,000.
However, McMahon cautions against over-extrapolating from these figures. First, the energy cost of encoding data into optics was not included in that calculation. Even if the scaling is favorable with light, in practice it is not just the scaling that matters, but also the baseline coefficients. Suppose the energy cost of doing one multiplication with electronics is 1 picojoule, while the cost to encode a single vector element with optics is 1,000 picojoules. The electronic approach still wins if the length $N$ of the vector is fewer than 1,000 elements.
McMahon notes that this hypothetical scenario is close to reality. "One of the key questions is, How big do you have to make the system before you start winning? Lots of people have done this calculation for various kinds of neural networks. The very rough answer is somewhere between $N$ equals 1,000 and $N$ equals 10,000."
It is also crucial to keep in mind a fundamental computing principle known as Amdahl’s law.
"Amdahl’s law shouldn’t even be called a law. It’s a simple piece of common sense," McMahon says. "If you have something that’s going to accelerate a computation, and it can accelerate 90 percent of the code, but there’s 10 percent of the code that it can’t accelerate, then the maximum possible speedup you can get is 10x."
With AI inferencing, the primary goal is often to save energy rather than raw time, but Amdahl’s observation still applies. Determining what fraction of today’s AI inferencing code can be better handled with optics is a complicated question that varies by architectural approach. Crucially, it is not 100 percent. Matrix-vector multiplication, or matrix-matrix multiplication, is only one part of the inference workload; there are other nonlinear operations that must be performed on the data.
Taking all of these engineering caveats into account, optical-computing companies today are realistically aiming for a 10- to 80-fold improvement in energy efficiency over current commercial GPUs. The central question is whether they can deliver on that promise at commercial scale before traditional electronics catches up to those efficiency gains.
Bridging the Gap in Free Space
The startup Lumai, based in Oxford, England, has taken perhaps the most direct inspiration from Goodman and his team. Lumai’s head of product, Phillip Burr, declares that a free-space approach, where light travels through air rather than through confined on-chip waveguides, is what will ultimately enable the company’s technology to outperform GPUs.

Instead of the fixed attenuator used in Goodman’s original proposal, Lumai utilizes a modern programmable device called a spatial light modulator—a pixelated membrane akin to a sophisticated liquid-crystal display capable of modulating both the intensity and phase of light. The rest of the design is more sophisticated in practice but shares the same fundamental principles as Goodman’s layout: A vector is encoded into light, copied $N$ times via a lens, passed through the spatial light modulator, collected with another lens, and converted back to digital electronics. The electronics then handle any necessary nonlinear functions. "A full vector-matrix multiply completes in a single optical cycle," Burr says.
To accurately compare the energy expenditures of optical and GPU approaches, engineers must specify which part of the AI workload they are considering. Modern inference is typically split into two distinct phases: prefill, where the input query is processed and propagated through the model, and decode, where the output is generated one word, or token, at a time. Prefill is by far the more computationally intensive phase, involving heavy vector-matrix multiplications for the same matrix repeated over and over. Decode also involves vector-matrix multiplies, but less intensively; the primary bottleneck during decode is the time-consuming process of reading the necessary matrices in and out of system memory.
Lumai, along with several other startups in the space, is focusing heavily on the compute-intensive prefill phase of inference. Lumai claims that its Iris Nova optical server, made available for evaluation, serves as a proof of concept, and its next-generation Iris Aura server will perform prefill operations using roughly one-tenth of the energy consumed by a traditional GPU.
Another startup, Neurophos, based in Austin, Texas, is putting a different spin on the classic Goodman computer design. To pack more computational power into a smaller device, the company is leveraging its deep expertise in metamaterials—complex engineered structures capable of manipulating light at scales smaller than a single wavelength. The structures Neurophos has designed are slightly larger than a single wavelength of light but still remarkably small, measuring just several micrometers per side.
"The equivalent of the optical transistor that you can get from standard silicon photonics today is like 2 millimeters long," says Patrick Bowen, CEO of Neurophos. "You just can’t fit enough of them on-chip to get a compute density that remotely competes with the latest and greatest that TSMC [Taiwan Semiconductor Manufacturing Co.] does."
Bowen’s team designed metamaterial-based modulators that can be programmed to manipulate the phase and intensity of light at the micrometer scale, resulting in a dense array of 1,000 by 1,000 modulators housed on a square chip with 5-millimeter sides. According to Bowen, Neurophos’s modulators switch over 10,000 times as fast as standard liquid-crystal spatial light modulators.
Bowen also claims a significant performance advantage over Nvidia’s Blackwell GPU architecture: "We’re targeting 50 times Blackwell B200’s speed and 35 times its energy efficiency, and have a road map to reach 250 times the speed and roughly 90 times the energy efficiency in subsequent generations."
The approach is fully CMOS-compatible, Bowen notes, which should allow Neurophos to manufacture its chips on the semiconductor industry’s existing high-volume production lines. The company is currently selecting a site for a high-volume production facility ahead of its planned rack deliveries to enterprise customers.

Both Lumai and Neurophos are betting heavily that they can bring their products to market before traditional CMOS technology catches up to their projected energy savings. "Silicon will keep improving, but digital efficiency gains now come slowly and at rising cost," Burr says. "Offers a different scaling relationship, in which efficiency improves as matrices grow. Larger models, longer context windows, and agentic workloads are all pushing prefill in that direction."
Bridging the Gap On-Chip
Free-space optics carries the alluring promise of almost energy-free multiplication, as demonstrated in Goodman’s classic design. However, another massive advance in optics since the 1970s is integrated photonics, which makes it possible to guide and manipulate light entirely on-chip.
"The name of the game is to do as large a vector size as possible to maximize the benefit of using optics, and for that reason the free-space approaches are promising," says Yale’s Logan Wright. "But it’s still fundamentally harder" to use the free-space approach, "because there’s just so much less infrastructure than in silicon photonics."
Other companies—such as Shanghai-based Lightelligence and Q.ANT, based in Stuttgart, Germany—are trading the three-dimensional advantage of free-space optics for the advanced manufacturability of photonics integrated directly onto a standard silicon microchip.
"There is a good reason that free-space optics are usually limited to 1,000 units a year," says Q.ANT CEO and founder Michael Förtsch. "The reason for this is lens inspection, assembly strategy, and so on. Semiconductor-manufacturable integrated photonic circuits are at this point way more promising to reach volume production."
In Q.ANT’s approach, input data is encoded into the phase and amplitude of light pulses via on-chip modulators. A vector in this scheme is represented as a sequence of such pulses, with each pulse representing one vector element. The light is guided around the microchip using microscopic waveguides. The basic computational element is an interferometer: A waveguide splits into two distinct paths, and the light beam traveling down one path is subjected to a phase shift controlled by a set voltage. When the two paths recombine, the resulting optical interference performs a mathematical computation. Subsequently, photodetectors and analog-to-digital converters translate the result back into digital data.
This serialized vector-matrix multiplication approach is inherently slower than Goodman’s parallel computer setup, where an entire calculation happens instantly. However, that slower speed can be mitigated by utilizing multiple wavelengths of light simultaneously, effectively encoding multiple vectors into light streams at once.
Q.ANT’s Native Processing Server is already being utilized in research institutions such as the Leibniz Supercomputing Centre in Munich and the Jülich Supercomputing Centre in Germany. The company also announced a commercial sale to cloud provider Ionos. Similarly, Lightelligence’s PACE2 processor is capable of multiplying 128-by-128-element matrices. In a research paper released by the company, Lightelligence described vector-matrix multiplication operating at a 1-gigahertz frequency with 3-nanosecond latency, claiming two orders of magnitude lower latency than traditional electronic GPUs.

The net energy advantage, however, is yet to be conclusively demonstrated at scale, and Q.ANT’s Förtsch admits that hurdles remain. The primary challenge is the energy cost of performing analog-to-digital conversion and back. As long as a fast, ultra-low-energy converter does not exist, Förtsch says, "we don’t see realistic chances that a general vector-matrix multiplication machine in optics makes sense."
Beyond Matrix-Vector Multiplication
Not everyone in the optical computing field is fixated on matrix-vector multiplication, despite it dominating the modern AI inference conversation. While optical vector-matrix multiplication dates back to the 1970s, other mathematical computations using optics have been investigated as far back as the 1950s. For instance, a Fourier transform—a common mathematical operation that decomposes a complex wave or signal into its constituent frequencies—can be executed optically using a single lens.
"A digital chip has to approximate [Fourier transforms] in many small steps, using hundreds to thousands of transistors and more compute cycles," says Q.ANT’s Förtsch. "Models that use such nonlinear functions can reach the same or better accuracy with fewer parameters and fewer operations."
Even though Q.ANT’s optical processor has been deployed for traditional vector-matrix multiplication, the company believes its greatest advantages lie in alternative AI architectures that leverage more complex mathematical functions. "Tasks such as image classification and segmentation rely heavily on operations like Fourier series, transformations, and other nonlinear functions, which photonics handles natively. We see similar potential in generative AI, especially image and video generation," Förtsch notes.
The Israel-based startup LightSolver is also targeting algorithms that represent a more natural fit for its specialized device. "If you look at propagating electromagnetic waves, essentially, they are running differential equations in their nature," says Ruti Ben-Shlomi, CEO and cofounder of LightSolver. She explains that differential equations can be solved much faster and with significantly higher energy efficiency on the company’s laser processing unit (LPU) than on a traditional digital computer.
LightSolver’s LPU introduces an interesting twist on the Goodman optical-computer design. It features input light, a programmable membrane in the form of a spatial light modulator, and output light, but the entire assembly is placed inside an optical cavity where light bounces back and forth many times before reaching a steady state. Additionally, the team places a gain medium inside the cavity, which functions much like the gain medium inside a traditional laser, causing the bouncing light to build up in intensity.
Once a particular computational problem is encoded into the modulator, the gain medium is stimulated to emit random photons. These photons bounce back and forth, increasing in intensity as they pass through the gain medium, experiencing preprogrammed changes through the modulator, and interacting with one another until they reach a steady state representing the solution. A simple camera then captures a picture of the steady-state result. Using hardware emulation, the LightSolver team demonstrated that their LPU can solve differential equations 40 to 80,000 times as fast as a GPU, depending on the workflow, and they are currently working on explicit energy benchmarking.
LightSolver’s team is also developing AI applications for its LPU, utilizing the company’s proprietary algorithms to handle image generation directly from text prompts. "We’re not doing vector-matrix multiplication very fast. We are doing something different. We need to assume that the way you train the model, and you then do inference, is completely different. But if you do it differently, you’ll be able to do image generation with dramatic energy-consumption decrease and speedup," Ben-Shlomi says. LightSolver currently has a working prototype of its LPU and is developing a larger, rack-mountable commercial unit.

Computing in Transit
Yet another innovative way to leverage light’s unique physical strengths is being pursued by Optalysys, a startup based in Leeds, England. Nick New, founder and chief product officer, has a background in free-space optics, which led him to explore how Fourier transforms can be executed efficiently with light.
The core idea behind Optalysys’s technology is taking that Fourier processing capability of optics in free space and translating it to work on-chip. "Rather than using a lens these days, we have our own component, which is a passive component in the plane of the silicon that does that same job," New says. By cleverly designing that component with multiple distinct outputs, the team figured out how to perform other operations as well. "Because of how the Fourier transform response is generated, you can actually pull out different mathematical operations from it," New explains. "We can produce very large, high-precision digital functions at superhigh efficiencies, using our analog photonics approach."
Instead of redesigning a standalone AI accelerator from scratch, Optalysys is targeting a segment of the AI workload that is already physically located in the optical domain: networking. Inside modern data centers, racks of computers are interconnected using fiber optics, and those optical connections are increasingly moving closer to the GPUs and CPUs themselves.
Certain computations already take place while data is actively in transit. These computations, which include combining results from multiple GPUs, are executed inside network switch cards, which consume power to perform that math. By employing passive optical components, those computations could theoretically be performed for free.
New points out that in addition to facilitating communication between GPUs, Optalysys’s technology could be used to assist a GPU’s communication with system memory by executing compression and format conversion on the fly. Returning to AI inference, this approach could significantly improve the decode phase of the workload. Unlike prefill, decode is strictly memory-bound, meaning the most costly part of the operation is shuffling model weights in and out of memory.
"We’re more focused on the memory-bound applications, which fits very well with this principle of adding processing into the communication path," New says. "We’re effectively adding processing to where there is none right now."
The Likely Future of Optical Computing
How likely are optical computers to completely replace electronic GPUs in the data center? Yale’s Wright remains cautious. "The direction of algorithm development is not going to be a direction that is automatically favorable for optics," he says. "In fact, it’s most likely to go in the opposite direction, because [AI model developers] want it to work better on GPUs."
However, even if standard AI queries are not computed entirely in light over the next few years, the heavy industry investment flowing into the technology will not be wasted. As Wright notes, "The reason that I and many other people in the field are at least somewhat optimistic is that there’s real investment from companies like Nvidia and TSMC in developing co-packaged silicon photonics—not for computing, but for circuit switching and interconnects, and that just kind of opens the door to do a little bit more in the future."

Whether that technology ultimately ends up crunching vector-matrix multiplications, executing bespoke scientific simulations, or performing complex in-network computations, the broader computing ecosystem appears increasingly ripe for a diverse array of approaches.
"We see the data center of the future as heterogeneous: CPUs, GPUs and photonic NPUs [native processing units] each take on the tasks they are best suited for," says Q.ANT’s Förtsch. "Running each task on the most efficient hardware can significantly reduce energy use and operating costs."
