Light can carry information, but it can also perform selected calculations inside a purpose-built circuit. Photonic computing is moving from isolated optical operations towards experimental systems that run neural networks. This review examines what recent demonstrations actually establish, from programmable tensor processors to tiny task-specific optical structures and computing-centre trials.
Its central question is what these advances could change for artificial intelligence: useful computing capacity, response time, energy demand and cost. The answer depends on the whole system, including memory, conversion, precision, calibration and data movement. Published classification results are distinguished from general AI capability; projected power budgets are separated from measured consumption. The article explains why a fast optical operation does not guarantee a faster application, and why silicon can remain part of a photonic future. It identifies credible opportunities for hybrid and specialised accelerators while setting out the evidence still needed before broad efficiency or economic claims can be accepted.
Why light enters the calculation
Many neural-network layers repeatedly multiply matrices and vectors. Optical circuits can represent numerical values through properties of light and combine signals to perform these linear operations. Different designs use interference, controlled transmission or the accumulation of optical intensities. Several wavelengths can also carry parallel streams. This is classical physical computation: it does not require a quantum computer. The growing ability to integrate, configure and control such circuits is surveyed by Pérez-López and Torrijos-Morán. [1]
Two developments must be kept distinct. Photonic interconnects move information between components; photonic processors perform calculations on it. Both can matter to AI, but an improvement in a communication link is not evidence that a neural network has been computed optically. Nor does photonics necessarily mean abandoning silicon: silicon-based optical circuits remain an important implementation route. [1]
This simplified path explains both the promise and the difficulty. Accelerating its middle stage helps only if the surrounding stages can keep up.
What has actually been demonstrated?
Beyond a single optical operation
Lightmatter’s 2025 Nature paper reports a processor executing ResNet and BERT models and an Atari reinforcement-learning workload, with precision approaching electronic implementations on many of the tested workloads. This matters because it connects optical computation to functioning AI tasks rather than stopping at an isolated multiplication. The paper also links its data and analysis materials. These are results from the developing team, published in a peer-reviewed journal; they are not measurements conducted by this magazine. They establish an experimental milestone, not that every model or training job will outperform a current GPU. [2]
Programmability brings a precision trade-off
A 2026 Nature Communications study by Meyer and colleagues demonstrates a reconfigurable, rack-mounted photonic tensor processor with electronic interfaces. It reports 98.1% accuracy on MNIST and 72.0% on CIFAR-10; the latter result uses a subset of 400 images. These are different classification tasks, not a ranking of general intelligence. The networks receive hardware-aware fine-tuning. Averaging repeated operations reduces noise but adds work, illustrating a direct tension between precision and speed. The paper’s energy-efficiency discussion uses a projected power budget: it must not be presented as a measured whole-system saving. [3]
A smaller optical region is not a smaller complete computer
Sved and colleagues report inverse-designed optical structures with footprints of 20 × 20 and 30 × 20 micrometres, and experimental classification results of 89% on MNIST and 90% on MedNIST. The physical structure is designed to transform light for its task. This is a compelling route to compact, specialised computing, but those dimensions describe the optical region, not the light sources, input preparation, detectors and complete system. A fabricated, task-specific transform is also different from a freely reprogrammable accelerator. MedNIST performance is a research benchmark, not clinical validation. [4]
From a paper to operation in a computing centre
In April 2026, the Leibniz Supercomputing Centre reported testing Q.ANT photonic computers in everyday operation. Its account describes hybrid systems that exchange data with conventional computers and reports server-and-card consumption in the 350–420 W range. That is useful operational evidence, but power alone cannot establish efficiency: a fair comparison also needs the amount and quality of completed work. The centre’s tests show that integration is being investigated beyond the optical laboratory; they do not establish a universal replacement for CPUs or GPUs. [5]
Which limits could this move for AI?
A 2026 perspective by Nechipurenko, Gautam and Shcherbakov separates cloud-oriented, general-purpose accelerators from specialised edge systems. Its central constraints include optical losses and electronic–optical interfaces in large systems, and reconfigurability, nonlinear operations and footprint at the edge. This is an assessment of the field, not another experimental performance record. [6] The practical implications below are conditional deductions from these constraints and the demonstrations above.
| Limit | Possible advance | Decisive condition |
|---|---|---|
| Energy and operating cost | More useful inference within a fixed electrical budget. | Lower energy per completed task at equivalent accuracy, including lasers, conversion, memory and control. |
| Response time | Faster repeated linear operations in tightly timed workloads. | Data transfer, batching and the remaining operations must not consume the gain. |
| Computing capacity | Parallel optical channels could increase the work processed by a suitable accelerator. | Enough data must reach those channels, with acceptable noise and sustained utilisation. |
| AI near sensors | Specialised optical processing could reduce the work sent to a distant server. | The task, packaging, environmental stability and ability to update the model must fit the device. |
| Hardware cost | A favourable system design could lower the cost of useful computation. | Manufacturing yield, packaging, lifetime, maintenance and software support must justify the investment. |
None of these changes automatically gives a model better reasoning, more reliable knowledge or safer behaviour. Hardware can enlarge the feasible budget for a model or reduce the cost of using it. Whether the model deserves that budget remains a separate scientific question.
The part cannot outrun the whole
Consider an illustrative workload in which 80% of execution time can be accelerated tenfold, while the other 20% is unchanged. The total speed-up is 1 ÷ (0.20 + 0.80 ÷ 10), or approximately 3.57 times, even before adding conversion or transfer overhead. This is a transparent arithmetic example, not a measurement or forecast for any processor discussed here.
What would count as a convincing next step?
A useful comparison would run the same task to the same quality target on the optical system and a well-configured electronic alternative. It would disclose numerical precision, batch size, latency and sustained throughput, and measure energy over a clearly stated boundary. Reprogramming, calibration and data movement belong in that account whenever the application needs them. A component record, a simulated future system and a measured server result answer different questions.
The September 2026 review by Chen and colleagues places practical optical computing at the intersection of materials, devices, architecture and integration. [7] That is why further progress cannot be inferred from an attractive chip image alone. The device must survive the demands of an actual workload, a maintained system and repeatable measurements.
The evidence supports a serious research direction: photonic hardware can perform AI-relevant calculations, and several demonstrations now connect those calculations to complete inference tasks. The strongest near-term case is for hybrid or specialised systems whose workload fits the hardware. A general reduction in the cost of AI, or the displacement of electronic accelerators across the board, has not been established by the evidence reviewed here.
Sources and scope
A selective state-of-the-art review, current to 13 November 2026. Experimental papers, reviews, a perspective and a computing-centre report are identified separately. Their results are not pooled into a performance league table. No hardware measurements were performed for this article.
- Pérez-López, D. & Torrijos-Morán, L. Large-scale photonic processors and their applications. npj Nanophotonics 2, 32 (2025). Review.
- Universal photonic artificial intelligence acceleration. Nature 640, 368–374 (2025). Experimental article. Linked data and analysis materials.
- Meyer, L. et al. Deep neural network inference on an integrated, reconfigurable photonic tensor processor. Nature Communications 17, 3396 (2026). Experimental article.
- Sved, J. et al. Inverse-designed nanophotonic neural network accelerators for ultra-compact optical computing. Nature Communications 17, 1059 (2026). Experimental article.
- Leibniz Supercomputing Centre. LRZ Puts Photonic Computers to Real World Test (21 April 2026). Computing-centre report.
- Nechipurenko, D., Gautam, A. & Shcherbakov, M. R. Photonic accelerators for AI across cloud and edge platforms. Nature Photonics 20, 1131–1138 (2026). Perspective.
- Chen, Y. et al. How nanophotonics can drive optical computing toward practical applications. Nature Nanotechnology 21, 1518–1530 (2026). Review.
Liked this post? Follow this blog to get more.