A New Chip Uses Light Instead of Electricity to Power AI — Why Penn's Polariton Breakthrough Could Solve the Energy Crisis

Eighty years ago, in a basement at the University of Pennsylvania, two engineers named J. Presper Eckert and John Mauchly built ENIAC — the world's first general-purpose electronic computer. It worked by pushing streams of electrons through vacuum tubes, and in doing so it launched the era of electronic computing that has defined nearly every device humans have built since. Eighty years later, almost to the month, a different team of Penn researchers has published a discovery that could mark the beginning of computing's next era — one built not on electrons, but on light. In May 2026, physicists led by Professor Bo Zhen demonstrated a way to switch and process information using particles that combine the properties of photons and electrons, consuming an amount of energy so small it is difficult to express in familiar terms. The implications, if the discovery can be scaled from a laboratory bench to a manufacturable chip, reach directly into one of the most urgent infrastructure problems of the AI era: the staggering and rapidly growing amount of electricity that artificial intelligence consumes.

What Polaritons Actually Are

To understand what the Penn team built, it helps to start with why electrons have been the foundation of computing for eighty years, and why that foundation is now showing real strain. Electrons carry an electrical charge, and that charge is precisely what makes them useful for building circuits — but it also creates problems. As electrons move through a material, that movement generates resistance, and resistance generates heat. Every time your laptop or phone gets warm during a demanding task, you are feeling the byproduct of electrons losing energy as they push through circuitry. At the scale of a modern AI data centre, that heat becomes an enormous and expensive problem, requiring industrial cooling systems that themselves consume vast amounts of additional electricity.

Light, by contrast, has no charge and experiences essentially none of this resistance. Photons — the particles that make up light — can travel through certain materials with minimal energy loss, and they can carry information far faster than electrons can move through a wire. The challenge that has stumped physicists and engineers for decades is that photons, lacking charge, do not naturally interact with each other the way electrons do. Two beams of light can pass directly through one another without affecting each other at all. That is wonderful for transmitting information over long distances using fibre optic cables, but it is a serious obstacle if you want to use light to perform the kind of switching and logical operations that a computer chip needs to do — operations that, in a conventional computer, rely on electrons interacting with transistors to represent the ones and zeros of binary computation.

This is the specific problem that Bo Zhen's team, working with co-first author Li He, set out to solve. Their solution was to create a hybrid particle called an exciton-polariton. The particle forms when photons of light are strongly coupled — essentially forced into close, persistent interaction — with excitons, which are themselves bound pairs of an electron and the "hole" it leaves behind, existing inside an extremely thin semiconductor material. By coupling light into a nanoscale cavity and allowing it to interact with an atomically thin layer of a material called molybdenum diselenide, or MoSe₂, the researchers created particles that behave partly like light and partly like matter. As Li He, now an assistant professor at Montana State University, explained the underlying principle: "By forcing light to couple strongly with the matter in atom-thin MoSe₂ layers, we can effectively have photons interact and change the system's behaviour using very little optical energy." In plain terms, the polariton gives light a way to influence other light — solving the fundamental obstacle that has limited optical computing for decades, while inheriting almost none of the energy-wasting baggage that comes with using electrons.

Why This Uses So Astonishingly Little Energy

The headline number from the Penn study is almost difficult to process intuitively, which is part of what makes it remarkable. Using their exciton-polariton system, the researchers demonstrated all-light switching — meaning one beam of light controlling and changing the behaviour of another beam of light, the basic operation a computer chip needs to perform — using only about four quadrillionths of a joule of energy. To put that in context the research team itself used: that amount of energy is far below what is needed to briefly power a tiny LED light. A joule is already a small unit of energy in everyday terms — it takes roughly one joule to lift a small apple one metre off the ground. A quadrillionth of a joule is so far below that scale that it strains ordinary intuition.

The reason this matters goes beyond the satisfaction of an extreme engineering achievement. Every computation performed inside a conventional electronic chip — every bit flipped from a zero to a one, every signal routed from one part of a processor to another — costs energy, and that energy cost compounds across the billions of operations a modern AI system performs every second. A switching mechanism that requires dramatically less energy per operation does not just save a small amount of power. Multiplied across the scale at which AI systems now operate, even a modest reduction in per-operation energy cost translates into enormous absolute savings. And the Penn team's result was not a modest reduction — it demonstrated switching at an energy scale that is orders of magnitude below what conventional electronic switching requires.

How This Connects to the AI Electricity Crisis

The timing of this discovery is not coincidental to the broader conversation happening across the technology and energy sectors right now. Artificial intelligence has developed an electricity problem of genuinely historic proportions. According to the International Energy Agency, global data centre electricity consumption reached approximately 415 terawatt-hours in 2024 — about 1.5% of the world's total electricity use. That figure has been growing at a compound annual rate of around 12% since 2017, more than four times faster than overall global electricity demand. By the IEA's own projections, data centre electricity consumption could exceed 1,000 terawatt-hours by the end of 2026 — an amount roughly equivalent to the entire annual electricity consumption of Japan, the world's fourth-largest economy. If data centres were treated as a single country, that level of consumption would make them one of the five largest electricity consumers on Earth, sitting between Japan and Russia.

The local effects of this growth are already visible in household utility bills. In the PJM electricity market, which stretches from Illinois to North Carolina, data centres accounted for an estimated $9.3 billion increase in the regional capacity market for 2025 to 2026 alone, with the average residential electricity bill in parts of western Maryland projected to rise by $18 a month as a direct consequence. Some individual AI data centres under construction now draw as much electricity as 100,000 homes, and the largest facilities being planned can use twenty times more than that, putting them on par with major industrial plants like aluminium smelters. The fundamental driver of this crisis is the energy cost of electronic computation at scale — the same heat-generating, resistance-driven inefficiency that has been a feature of electron-based computing since ENIAC first hummed to life in 1945.

This is precisely the problem that light-based computing is positioned to address. If photonic chips built on technology like Penn's exciton-polariton system can eventually replace even a portion of the electronic switching happening inside AI data centres, the energy savings would not be a marginal improvement — they would represent a fundamentally different cost structure for the computation that AI depends on. The research published by Zhen's team explicitly identifies lowering the energy demands of large AI systems as one of the central motivations and potential payoffs of the work.

What Researchers See as the Path Forward

The Penn team is careful, in their published research, to frame this as a foundational physics result rather than a finished product — but the applications they envision are specific and significant. If the exciton-polariton switching mechanism can be successfully scaled from the nanoscale laboratory cavity used in this study to manufacturable chip architectures, the researchers see a path toward photonic chips capable of processing information directly from cameras and optical sensors, without the repeated and energy-costly conversions between light and electricity that current systems require every time an optical signal needs to be processed digitally. Eliminating those conversion steps — where light first has to be converted into an electrical signal before a conventional chip can process it, and often converted back to light afterward for transmission — removes a significant source of energy loss and processing delay in modern computing architecture.

Beyond reducing AI's energy footprint, the researchers also point to a second, more exploratory application: the potential to support basic quantum computing functions on future semiconductor chips. Quantum computing, which exploits the strange behaviour of particles at extremely small scales to perform certain calculations far beyond the reach of classical computers, has historically required enormously complex and expensive infrastructure to maintain the delicate quantum states involved. A semiconductor platform capable of generating and controlling polaritons reliably could, in principle, offer a more practical route toward integrating quantum-level effects into more conventional chip designs — though the researchers themselves describe this as an early-stage possibility rather than a near-term application. This work builds on a broader and increasingly active field of photonic computing research at Penn and elsewhere; a separate Penn Engineering team led by Professor Liang Feng has already developed a programmable chip that uses light to train nonlinear neural networks, demonstrating that the broader thesis — using photons rather than electrons for the mathematics underlying AI — is being pursued across multiple research groups simultaneously, each tackling different pieces of the same fundamental challenge.

A Realistic Timeline From Laboratory to Commercial Chips

It is worth being honest about where this technology currently stands, because the gap between a landmark physics result published in Physical Review Letters and a commercially available AI chip is substantial, and history offers useful guidance on how that gap typically closes. The Penn result is, at this stage, a laboratory demonstration of a physical principle: that exciton-polaritons can perform all-light switching at extraordinarily low energy using a specific, carefully engineered material system involving a nanoscale cavity and an atomically thin semiconductor layer. It has not yet been demonstrated in a fully integrated, multi-component chip capable of performing the complex range of operations a real AI processor needs to execute.

The realistic timeline for technology at this stage typically unfolds in stages spanning many years. The next phase of research will likely focus on demonstrating more complex circuits built from multiple polariton switching elements working together, proving that the technology can perform not just a single switching operation but the kind of cascading logical operations a genuine processor requires. Following that, researchers would need to address the manufacturing challenge of producing these nanoscale cavity and monolayer semiconductor structures reliably and at the scale required for commercial chip fabrication — a process that, for previous generations of novel semiconductor technology, has often taken five to ten years to mature from laboratory demonstration to pilot manufacturing. Photonic computing as a broader field is already at a more advanced stage than the specific polariton technology described here, with some photonic AI accelerators already being tested commercially for narrower applications, suggesting that the surrounding ecosystem and manufacturing expertise exists to potentially accelerate the path for the most promising laboratory breakthroughs. A reasonable estimate, based on the typical maturation timeline for comparably foundational semiconductor physics discoveries, would place commercially available chips incorporating this specific exciton-polariton technology somewhere in the 2032 to 2036 range, assuming continued research funding and successful resolution of the manufacturing challenges involved — though breakthroughs in adjacent areas of photonic computing could bring partial applications to market considerably sooner.

Conclusion

The discovery published by Bo Zhen, Li He, and their colleagues at the University of Pennsylvania is, at its core, a demonstration that one of the most stubborn obstacles in optical computing — getting light to meaningfully interact with itself — can be solved using hybrid light-matter particles operating at an energy scale that is almost impossible to intuitively grasp. Four quadrillionths of a joule, less energy than it takes to make a tiny LED flicker, is the cost of switching information using exciton-polaritons rather than electrons. That number matters enormously not because it will immediately appear in your next laptop or phone, but because it points toward a fundamentally different cost structure for the kind of computation that AI depends on, at a moment when AI's electricity consumption is on a trajectory to match the entire annual power usage of Japan within the next few years. The road from this laboratory result to a chip sitting inside a data centre is long, and the engineering challenges that remain are genuinely difficult. But eighty years after Penn researchers built the machine that started the electronic computing era using streams of electrons, a new generation of Penn physicists has shown a credible, physically grounded reason to believe that light may carry computing into its next one.

*This article is based on the peer-reviewed study "Strongly Nonlinear Nanocavity Exciton Polaritons in Gate-Tunable Monolayer Semiconductors" by Zhi Wang, Bumho Kim, Bo Zhen, and Li He, published in Physical Review Letters on April 8, 2026, and reporting from ScienceDaily, TechXplore, SciTechDaily, Dataconomy, the International Energy Agency, and Penn Today.*