AI’s Energy Crunch Revives Optical Computing as a Viable Alternative
Optical computing, driven by AI’s demand for lower‑energy inference, could slash multiplication energy by up to 100,000‑fold, but practical hurdles remain.
S
··2 min readSciTech Editorial Team
Newsroom

The relentless rise in energy consumption of AI inference – the process that powers chatbots, image recognizers and recommendation engines – has forced the industry to look beyond traditional silicon. Optical computing, a concept that has lingered in research labs for more than seven decades, is now being hailed as a practical answer. Market analysts note that the combined valuation of firms in this niche has jumped from US$0.4 billion in 2022 to US$2.22 billion today, and projections suggest exponential growth as data‑center operators scramble for greener solutions.
At the heart of the promise lies a simple physical principle: light can perform matrix‑vector multiplication – the workhorse operation of modern neural networks – with virtually no energy loss once the optical mask is fabricated. The idea was first sketched in 1978 by Joseph W. Goodman, who described a system where an array of LEDs encodes a vector, a cylindrical lens copies it, and a pixelated attenuator implements the matrix. Passing the light through the mask yields the product without active electronic multiplication. In 2022, Peter McMahon’s team at Cornell University demonstrated an optical neural network that consumed less than a single photon per multiplication (≈2.5 × 10⁻¹⁹ J), compared with the 10⁻¹⁴–10⁻¹² J typical of electronic GPUs – a theoretical reduction of up to 100,000‑fold.
Start‑ups are racing to embed this physics into real data‑center hardware. Optalysys, a UK‑based company, integrates specially engineered waveguides into the networking fabric, turning the data‑transfer path itself into a Fourier‑transform calculator. Similar ventures have sprouted across the United States, China, Germany and Israel, each hoping to complement or even supplant the dominant GPU manufacturers such as Nvidia. Researchers like Logan Wright (Yale University) argue that even if the first wave of AI‑driven photonic chips does not dominate the market, the technology will unlock specialised photonic systems for scientific computing and in‑network processing.
Nevertheless, the road to commercial viability is strewn with practical obstacles. McMahon cautions that the energy advantage assumes the matrix remains static; the cost of encoding each vector element into light can dwarf the passive multiplication gain. A back‑of‑the‑envelope calculation shows that if encoding costs 1,000 pJ per element while electronic multiplication costs 1 pJ, the optical route only wins when the vector length exceeds 1,000 elements – a size that many inference workloads do not yet reach.
If those scaling hurdles can be overcome, the impact could be transformative. Data‑center architects may redesign racks to route photons instead of electrons, dramatically lowering the power envelope of AI services. Ongoing research aims to shrink encoding circuitry, develop reconfigurable optical masks, and integrate photonic memory, all of which could push the break‑even point down to realistic model sizes. As AI workloads continue to balloon, the convergence of photonics and machine learning may finally give the long‑awaited second life to optical computing.



