AI's Power Hunger Is Reviving Optical Computing—70 Years After the Idea
Optical computing companies are now worth $2.22 billion, up from $0.4 billion in 2022, as AI inference power costs push startups to compute with light instead of electrons.

Why it matters
- The optical computing sector's market valuation hit $2.22 billion, up from $0.4 billion in 2022, driven by AI inference power demand.
- A 2022 Cornell team demonstrated an optical neural network using less than one photon (~2.5 x 10^-19 joules) per multiplication, suggesting up to 100,000x energy savings over electronics.
- Neurophos targets 50x the speed and 35x the energy efficiency of Nvidia's Blackwell B200, with rack deliveries planned for 2028.
- Lumai's Iris Nova optical server opened for evaluation in April 2026; the follow-on Iris Aura aims to do inference prefill at one-tenth of GPU energy.
- LightSolver's laser processing unit solved differential equations 40 to 80,000 times faster than a GPU in emulation tests, depending on the workflow.
By Sophie Lindqvist
Optical computing companies are now worth a combined $2.22 billion, up from $0.4 billion in 2022 — and the reason is AI. Inference, the process of running models to answer user queries, has pushed data center power consumption to the point where researchers say the industry is literally constrained by electricity. That constraint has resurrected a 70-year-old idea: do the math with light instead of electrons.
"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, who has researched optical computing for over a decade.
Startups are now appearing across the United States, China, Germany and Israel, in McMahon's words, "like mushrooms after a rainstorm." Most want to compete with Nvidia's GPUs. Others are building complementary plays — alternative AI algorithms, scientific computing, and in-network processing. One projection cited in the market analysis suggests the sector will keep growing exponentially.
Why the idea works on paper
The heart of AI inference is one operation repeated billions of times: matrix-vector multiplication. Back in 1978, optical pioneer Joseph W. Goodman and his coworkers proposed doing it with light. N light-emitting diodes encode a vector's values into light intensity. A cylindrical lens copies the vector N times. The light passes through an N-by-N pixel attenuator — each pixel dims the light according to a matrix value, performing the multiplication. A second lens collects and sums the result, and a row of detectors reads out the answer.
The energy argument is simple. In electronics, multiplying an N-by-N matrix by a length-N vector takes N² multiplication steps, each costing energy, so total energy scales as N². In optics, the multiplication is passive. If the matrix stays fixed, passing light through the attenuator costs almost nothing; you only pay to encode the vector and read the result, which scales as N. That's the whole pitch.
The hardware has also improved dramatically since 1978. In 2022, a Cornell team led by McMahon demonstrated an optical neural network that used less than one photon of energy — roughly 2.5 x 10⁻¹⁹ joules — per multiplication. A standard computer consumes between 10⁻¹⁴ and 10⁻¹² joules per multiply, implying a possible 100,000-fold efficiency gain.
Why the reality is messier
McMahon himself cautions against that extrapolation. The calculation excluded the cost of encoding data into light. And even with favorable scaling, the constants matter. If one electronic multiply costs 1 picojoule but encoding one optical vector element costs 1,000 picojoules, electronics still wins for vectors shorter than 1,000 elements.
"One of the key questions is, How big do you have to make the system before you start winning?" McMahon says. "The very rough answer is somewhere between N equals 1,000 and N equals 10,000."
Then there's Amdahl's law. "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," McMahon says. Optical systems accelerate matrix multiplication, but inference also includes nonlinear operations that must happen elsewhere.
Factor all that in, and today's optical-computing companies are targeting a 10- to 80-fold energy-efficiency improvement over current GPUs. Whether they can deliver at scale before silicon catches up is the open question.
What does free-space optics buy you?
Lumai, based in Oxford, England, follows Goodman's blueprint most directly — with light traveling through air rather than on-chip waveguides. Instead of a fixed attenuator, Lumai uses a programmable spatial light modulator, a pixelated membrane that modulates both the intensity and phase of light. Electronics handle the nonlinear functions. "A full vector-matrix multiply completes in a single optical cycle," says Phillip Burr, Lumai's head of product.
The company targets the prefill phase of inference — processing the input query, the more compute-intensive half versus token-by-token decode. Lumai says its Iris Nova optical server, available for evaluation in April 2026, is a proof of concept, and its next-generation Iris Aura server will do prefill using roughly one-tenth of a GPU's energy.
"Silicon will keep improving, but digital efficiency gains now come slowly and at rising cost," Burr says. "Optics 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."
Neurophos, in Austin, Texas, attacks the density problem with metamaterials — structures that manipulate light at sub-wavelength scales. "The equivalent of the optical transistor that you can get from standard silicon photonics today is like 2 millimeters long," says CEO Patrick Bowen. "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 does."
Bowen's team built micrometer-scale programmable modulators — an array of 1,000 by 1,000 on a 5-mm-square chip. He claims they switch over 10,000 times faster than standard liquid-crystal spatial light modulators, and he's aiming squarely at Nvidia: "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 CMOS-compatible, and Neurophos is choosing a production-facility site ahead of rack deliveries to customers in 2028.
What about putting light on a chip?
Integrated photonics trades free-space's 3D advantages for silicon-style manufacturing. "There is a good reason that free-space optics are usually limited to 1,000 units a year," says Michael Förtsch, CEO and founder of Q.ANT in Stuttgart. "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."
Q.ANT encodes data into the phase and amplitude of light pulses in on-chip waveguides. The core computing element is an interferometer: a waveguide splits, one path gets a voltage-controlled phase shift, and recombination performs the multiplication. Photodetectors and analog-to-digital converters return the result to digital. It's slower than Goodman's design, but multiple wavelengths can carry several vectors at once. Q.ANT's Native Processing Server already runs at the Leibniz Supercomputing Centre in Munich and the Jülich Supercomputing Centre, and the company announced a commercial sale to cloud provider Ionos earlier this year.
Shanghai-based Lightelligence, meanwhile, sells the PACE2 processor, which multiplies 128-by-128-element matrices. In a 2025 paper, the company described vector-matrix multiplication at 1-gigahertz frequency and 3-nanosecond latency, claiming two orders of magnitude lower latency than traditional GPUs.
The energy advantage isn't proven yet. The bottleneck is analog-to-digital conversion. Until a fast, low-energy converter exists, Förtsch says, "we don't see realistic chances that…a general vector-matrix multiplication machine in optics makes sense."
What else can light compute?
Other players avoid the matrix race entirely. Q.ANT argues photonics natively handles Fourier transforms and other nonlinear functions. "A digital chip has to approximate [Fourier transforms] in many small steps, using hundreds to thousands of transistors and more compute cycles," Förtsch says. "Models that use such nonlinear functions can reach the same or better accuracy with fewer parameters and fewer operations." He sees potential in image classification, segmentation, and generative AI, especially image and video generation.
Israel's LightSolver built a laser processing unit — light, a spatial light modulator and a gain medium inside a cavity where photons bounce until they reach a steady-state solution, photographed by a simple camera. "If you look at propagating electromagnetic waves, essentially, they are running differential equations in their nature," says CEO and cofounder Ruti Ben-Shlomi. Via emulation, the team showed the LPU solves differential equations 40 to 80,000 times faster than a GPU depending on the workflow; energy benchmarks are in progress. The company has a working prototype, a rack-mountable box in development, and its own text-to-image algorithm. "We're not doing vector-matrix multiplication very fast. We are doing something different," Ben-Shlomi says.
Optalysys in Leeds, England, puts its Fourier-transform chip where data centers already use light: the network. Rack-to-rack connections run on fiber, and some computations already happen in switch cards, which burn power. Passive optics could do them essentially for free — including compression and format conversion in GPU-memory traffic, easing the memory-bound decode phase. "We're effectively adding processing to where there is none right now," says founder and chief product officer Nick New.
Will light actually replace GPUs?
Probably not soon, says Yale's Logan Wright. "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."
Still, Wright sees a foot in the door: "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."
"We see the data center of the future as heterogeneous: CPUs, GPUs and photonic NPUs 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." That heterogeneity, not outright GPU replacement, is the realistic prize the optical-computing wave is chasing.
Original: lumai.ai
More from Sophie Lindqvist
Show full bio
Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.
749 articles
Related articles
- NVIDIA to invest up to $100 billion in OpenAI for 10 gigawatts of AI infrastructure
- Microsoft Shifts AI Strategy Toward Homegrown Chips to Cut Nvidia Reliance
- Alphabet's Real AI Edge Is Locked-In Power, Not Chips
- Nvidia Donates $250 Million in Computing Power to UK AI Safety Research
- Nvidia’s AI-defined model targets semiconductor process simulation