Lumai
Last reviewed
Jun 3, 2026
Sources
12 citations
Review status
Source-backed
Revision
v1 · 1,668 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
12 citations
Review status
Source-backed
Revision
v1 · 1,668 words
Add missing citations, update stale details, or suggest a clearer explanation.
Lumai is a British optical-computing startup developing AI accelerator hardware that performs the matrix mathematics behind neural networks using beams of light rather than electric current. Spun out of physics research at the University of Oxford and based near Oxford, the company is trying to address one of the most stubborn problems in modern AI: the electrical power and cost of running large models at scale. Its pitch is that moving the heaviest arithmetic of AI inference out of silicon transistors and into photonics can deliver far more computation per watt than conventional electronic chips.[1][2]
Lumai sits in a small but growing field betting that optical computing is finally ready for commercial AI workloads. What sets its approach apart is the use of free-space optics, where light travels through open three-dimensional space rather than being confined to waveguides etched onto a chip, which the company argues sidesteps the scaling limits that have held back integrated silicon photonics.[3]
Lumai was incorporated in January 2022 as a University of Oxford spinout, building on optical-neural-network research carried out in the university's Department of Physics. Lumai's own materials trace the underlying research and spinout activity to 2021.[2][4]
The founding team brought together five people with backgrounds in optics, machine learning, and physics: Tim Weil, Xianxin Guo, Thomas Barrett, James Spall, and A. I. (Alex) Lvovsky, a professor in Oxford's Department of Physics known for work in quantum optics and optical machine learning.[1][5] Guo and Spall did much of the foundational research; their 2022 paper in the journal Optica, co-authored with Lvovsky, described a way to train optical neural networks in which the computations are carried out optically and the training gradients are approximated using passive optical components.[6] That work, along with a pending US patent titled "Optical multiplication system and optical multiplication method," underpins the company's technology.[3]
Tim Weil served as the founding chief executive. By the time Lumai launched its first commercial system in 2026, Xianxin Guo had taken over as chief executive and co-founder, while Phillip Burr led product.[2][7] The company completed Intel Ignite's London accelerator program and won "Best Overall Technology" at the Global OCP Future Technologies Symposium.[3]
Most of the computation in a transformer-based model is matrix multiplication: large grids of numbers multiplied and summed, over and over. On a GPU or other electronic accelerator, this work is done by transistors switching billions of times per second, and a large share of the energy goes into moving data rather than into the math itself.
Lumai's idea is to let physics do the arithmetic. Numbers are encoded onto light, for example in the brightness or amplitude of beams, and when those beams pass through optical elements and combine, the multiplication and accumulation happen as a consequence of how light behaves. The company calls the core of its processor an "optical matrix multiplier," and more precisely it performs optical matrix-vector multiplication, the dominant operation in inference.[1][3] Because many values can travel through the same optical volume in parallel, a single pass can carry out a large number of multiply-accumulate operations at once.
The distinctive part is that Lumai uses free-space optics in three dimensions. Instead of squeezing light through waveguides patterned on a two-dimensional chip, as integrated silicon photonics does, Lumai sends beams through a 3D volume. The company frames this as a way to "overcome the two-dimensional constraints of conventional chips" and to get past the scaling problems that have limited integrated photonics.[2][3] Working in three dimensions and across multiple wavelengths, in principle, gives more room to grow the size of the matrices being computed.
The systems are not purely optical. Lumai uses a hybrid electro-optical architecture: a digital electronic section handles system control, data movement, and the parts of the model that are not large matrix multiplications, while an optical tensor engine handles the bulk of the inference math.[2][8] This split is common in practical optical-computing designs, because tasks such as memory, control logic, and nonlinear activation functions are still easier to do electronically.
Lumai's headline claims are ambitious and should be read as company projections rather than independently audited benchmarks. The company says its design can deliver up to 50 times the performance of silicon-only accelerators while using about 10 percent of the power, and can cut the cost of inference to roughly one-tenth of current solutions.[1][2] It has also cited a longer-term technology speed limit of up to 10^17 (100 quadrillion) operations per second.[5]
Trade press has put more concrete numbers on the near-term product. According to EE Times, an early Lumai accelerator was planned as a PCIe card delivering around 8,000 INT8 TOPS (8 peta-operations per second) while drawing about 500 watts, roughly 16 TOPS per watt, with a 4U air-cooled rack shelf holding two optical processors for about 16,000 TOPS combined. Volume production was still a couple of years out, and future generations could improve through higher speeds, wider vectors, and wavelength multiplexing.[8]
In April 2026, Lumai announced its first commercial system, the Lumai Iris inference server, which it described as the first optical computing system to run billion-parameter large language models in real time. The company said the system ran real-time inference on Meta's Llama 3.1 8B and 70B models, with what it described as up to 90 percent lower energy consumption than conventional architectures.[7][9] The Iris line is planned as a family of servers named Nova, Aura, and Tetra; Iris Nova was the first to ship for evaluation by hyperscalers, neoclouds, enterprises, and research institutions.[7]
| Item | Detail | Source / status |
|---|---|---|
| Approach | Free-space 3D optical matrix-vector multiplication | Company / EE Times |
| Architecture | Hybrid electro-optical; optical tensor engine plus digital control | Company / The Register |
| Headline claim | ~50x performance, ~10% power vs silicon-only | Company projection |
| Early card (planned) | ~8,000 INT8 TOPS, ~500 W, ~16 TOPS/W, PCIe | EE Times reporting |
| First product (2026) | Lumai Iris (Nova, Aura, Tetra); ran Llama 3.1 8B and 70B | Company announcement |
| 2029 target | Iris Tetra at ~1 exaOPS within a 10 kW power budget | The Register |
For the long term, Lumai told The Register it expects its Iris Tetra systems to reach about one exaOPS within a 10 kilowatt power budget by 2029.[10] These roadmap figures depend on milestones the company has not yet demonstrated, so they remain targets rather than shipped specifications.
Lumai's growth has been backed by a series of investments. Early support came from IP Group and Runa Capital, and in February 2023 the company won a £1.1 million Innovate UK Smart Grant to commercialise its work on all-optical network training and "deep optics."[4][11]
In April 2025, at the Optical Fiber Communication conference in San Francisco, Lumai announced a round of more than $10 million led by Constructor Capital, with continued backing from IP Group and new investors PhotonVentures, Journey Ventures, LIFTT, Qubits Ventures, State Farm Ventures, and TIS Inc.[1][12] The company said it would use the money to advance product development, roughly double its headcount, and expand its presence in the United States.[1]
Lumai is entering a market dominated by Nvidia's electronic GPUs, where a wave of startups is competing specifically on inference rather than training. The Register has grouped Lumai among inference-focused challengers and noted that, while companies such as Groq (acquired by Nvidia), Cerebras, and SambaNova have concentrated on the fast token-by-token decode stage, Lumai is initially positioning its optical accelerators for compute-bound work such as batch processing, with longer-term plans to serve as prefill processors in disaggregated inference setups.[10]
Optical computing has a long history of promise and disappointment. The appeal is real: light can carry many signals in parallel, and operations such as multiplication and summation can happen passively as beams combine, which can be far more energy-efficient than switching transistors. The difficulties are also real. Converting data between the electronic and optical domains costs power and time, analog optical systems are sensitive to noise and calibration, and earlier integrated-photonics efforts struggled to scale to useful matrix sizes. Lumai's bet is that free-space 3D optics gets around the scaling wall that limited chip-based photonics, but it still has to prove its systems are reliable, manufacturable, and competitive against improving electronic accelerators at volume.[3][8]
Whether optical inference becomes a mainstream part of data center infrastructure or stays a niche depends on results that are not yet public. What is clear is that Lumai has moved from a physics-lab idea to a funded company with hardware in the hands of potential customers.