Why Light Moves Data and Electrons Still Compute

Scope

This article separates two claims that are constantly merged, that light is good at moving data and that light is good at computing, and assesses each on the evidence as of late 2026. It explains the appeal of optical matrix multiplication, the costs that appeal hides, why the interconnect case succeeded while the compute case has not, and where optical computation still earns its place. It is an argument, not a procedure, and it takes a position. Background: optical computing, Mach-Zehnder interferometer, co-packaged optics.

Two different bets on light

For fifty years light has been the way to send information over distance, and the reasons are not in dispute: enormous bandwidth, loss measured in fractions of a dB per kilometer, and no crosstalk between beams that cross. The open question was never whether light moves data well. It was whether light can also compute, and specifically whether the linear algebra at the heart of modern machine learning can be done in the optical domain more efficiently than in silicon transistors.

These are separate bets with separate physics, and treating them as one, under a single banner of photonic computing, is the source of most of the confusion. One bet has paid off decisively. The other has not, and the reason it has not is instructive.

Why optical matrix multiplication looks free

The appeal is real and worth stating at full strength. A mesh of interferometers performs a matrix-vector multiplication by interference: encode a vector on the amplitudes of several waveguides, set a matrix in the mesh's phase shifters, and the light emerging from the far side is the product. The computation happens during propagation, in the time light takes to cross a few millimeters of chip, and the arithmetic itself dissipates no switching energy. Matrix multiplication is the overwhelming majority of the work in neural-network inference, so a device that does it at the speed of light for no per-operation energy sounds like it should end the discussion.

It does not, and the gap between that promise and the system-level result is the whole story.

What the promise leaves out

Four costs sit between the free multiplication and a working computer, and together they reverse the verdict.

The conversions dominate. The input vector arrives as digital electronic data and must be turned into optical amplitudes, which means a digital-to-analog conversion and an electro-optic modulation per element; the output must be detected and digitized again. These conversions, repeated at the boundary of every optical layer, consume far more energy than the multiplication saved, and they scale with the data, not with the arithmetic, so they do not amortize away as the matrix grows.

The precision is low. Analog optical computation carries an effective precision of a handful of bits, limited by noise, detector resolution, and the accuracy with which phases can be set and held. Digital hardware supplies whatever precision is asked for. Many inference workloads tolerate low precision, but the ones that do are also the ones digital accelerators already run at very high efficiency in low-precision integer formats, so the comparison is against a moving and formidable target.

The matrix must be held and reloaded. The phase shifters that encode the weights drift with temperature and must be trimmed continuously and reloaded when the weights change, and a large model's weights do not fit in one mesh, so they stream through, which brings back the data-movement cost that optics was supposed to remove.

The rest of the network is still electronic. The nonlinear activation between layers and the memory that holds activations and weights remain electronic in every practical design, so a real system converts out of the optical domain and back at each layer boundary, paying the conversion cost again.

Count these honestly and general-purpose inference still belongs to digital electronics. The optical multiplication is genuinely free; everything around it is not, and everything around it is where the energy goes.

The pivot as evidence

The most telling evidence is what the companies did. Several of the best-funded optical-processor efforts, having built impressive photonic meshes, redirected their shipping products toward optical interconnect: moving data between electronic accelerators with co-packaged and near-package optics, which is a problem light unambiguously solves and a market the AI build-out made enormous. The compute continues as research; the revenue comes from interconnect. When the people with the deepest stake in optical computing ship interconnect instead, the market has rendered a verdict that no white paper overturns. Meanwhile the serious remaining attempts at compute by light have moved to platforms such as thin-film lithium niobate with native nonlinear functions, aimed at the specific workloads where the conversion overhead can be hidden, which is the right place to look for the exceptions.

Where compute by light still fits

The verdict is not that optical computing is impossible, only that it does not win the general case on today's terms. It fits where its weaknesses do not bind: fixed linear transforms that never reload their matrix, so the trimming cost is paid once; front-end signal processing where the data is already optical and analog, so the conversions are not added but avoided; and specialized low-precision or physically-inspired computations whose structure matches what a mesh does naturally. These are real and worth pursuing. They are not the datacenter inference market, and honest work in the field says so.

What this means for the reader

The practical guidance is to keep the two bets separate when reading a claim about photonic computing. If the claim is about moving data, light is the answer and the only remaining questions are packaging and cost; the co-packaged optics and architecture articles cover that winning case. If the claim is about doing arithmetic, ask where the digital-to-analog and analog-to-digital conversions are, what the effective precision is, and how the weights are held and reloaded, because that is where an optical processor's real energy budget lives, and it is usually not in the multiplication being advertised.

References: Shen et al., "Deep learning with coherent nanophotonic circuits," Nature Photonics 11, 441 (2017); Hamerly, Bernstein, Sludds, Soljacic & Englund, energy analyses of optical neural networks (2019 onward); Nahmias et al., photonic tensor processing reviews; industry product materials showing the optical-interconnect pivot (2025-2026). The interconnect side that succeeded is covered in Light sources for co-packaged optics and Pluggable, LRO, LPO, CPO.