AI’s power problem is reopening the case for optical computing
Optical computing has spent decades in the shadow of conventional electronics. The basic proposition is compelling: use light to perform mathematical operations that would otherwise be handled by electronic circuits. Yet for more than 70 years, the technology has largely remained a research pursuit rather than a broad commercial computing platform.
That picture is beginning to change. A new wave of companies is pursuing photonic systems as artificial intelligence drives demand for more computing capacity and more electricity-efficient inference. The shift does not mean optical processors are about to displace graphics processors across data centers. It does mean that AI’s energy constraints have created a new commercial reason to develop hardware once regarded as technologically interesting but difficult to deploy.
IEEE Spectrum reports that the combined market valuation of companies in the optical-computing sector has reached US$2.22 billion, up from $0.4 billion in 2022. Companies are emerging in the United States, China, Germany and Israel, reflecting a wider effort to find computing architectures that can complement or, in selected workloads, challenge today’s electronic systems.
Why inference is the immediate target
The central driver is AI inference: the work of using trained models to answer prompts and make predictions. As more people and businesses use AI services, inference workloads are helping fuel the construction of large data centers. The result is a growing focus on the electricity needed not only to train models, but also to operate them continuously.
Peter McMahon, an associate professor of applied and engineering physics at Cornell University, told IEEE Spectrum that interest in photonics is being driven by the degree to which AI inference has become constrained by power. Optical systems offer, at least in principle, a path to carrying out relevant mathematical operations with far less power than conventional GPU-based hardware.
The promise is not simply that light moves quickly. It is that specially designed photonic hardware may execute particular calculations efficiently. That is a narrower and more practical proposition than replacing every component of a digital computer with optics. It also helps explain why developers are exploring several routes to commercialization rather than betting on one universal optical AI chip.
Computing where the data already travels
One approach comes from Optalysys, an England-based company that is placing optical-computing elements in the networking path. Data centers already use light for networking, so this strategy seeks to add computational capacity where optical signals are already moving. The company’s on-chip optical element is a Fourier transform calculator, according to the report.
That model illustrates an important distinction in the current revival. The near-term opportunity may not be a direct replacement for Nvidia-style accelerators. Instead, photonics can be introduced into parts of a computing system where it has a natural fit, such as the links that move data among machines or components. If useful calculations can occur while data is in transit, the technology could become part of data-center architecture without requiring every workload to be redesigned at once.
Other efforts are focusing on alternative AI algorithms, scientific computing and related photonic systems. The range of approaches suggests that the sector’s value may extend beyond whichever companies succeed in the most ambitious goal of competing directly with GPU makers.
A technology race with spillover value
Logan Wright, an assistant professor of applied physics at Yale University and a former postdoctoral researcher in McMahon’s laboratory, told IEEE Spectrum that a strong run by several companies could benefit the technology even if not all of them become major AI-computing providers. The development work could enable photonic systems that serve other specialized purposes.
That is a useful way to view the present moment. AI is supplying the urgency, investment and market attention that optical computing lacked for years. But the technical advances stimulated by that demand could have applications in scientific computing and network-integrated processing as well. The commercial outcome will depend on whether companies can convert laboratory advantages into systems that work reliably within the demanding environment of modern data centers.
For now, optical computing remains an emerging hardware field rather than an established replacement for electronic AI infrastructure. The renewed investment, however, signals that power efficiency has become a central competitive issue in AI. As inference expands, hardware developers are increasingly looking beyond faster versions of existing chips and toward architectures that change how and where computation is performed.
What to watch next
- Whether photonic systems demonstrate useful power advantages on real inference workloads.
- Whether in-network computing creates an easier entry point than full processor replacement.
- Whether optical-computing investment produces commercially useful components even where broader AI ambitions fall short.
This article is based on reporting by IEEE Spectrum. Read the original article.
Originally published on spectrum.ieee.org





