👤 ✉

The End of the Electron Era for AI Hardware? (Part III of VI)

Eighty years after ENIAC, a Penn lab demonstrates all-light switching at four quadrillionths of a joule — and the redistribution-of-power case for photonic AI.

University of Pennsylvania · Physical Review Letters · 15–18 May 2026 · Reading time ~3 minutes

Eighty years after Penn researchers J. Presper Eckert and John Mauchly launched the electronic computer with ENIAC, a different Penn team has published the strongest case yet that electrons may not be the right medium for the next century of AI.

In Physical Review Letters, the lab of physicist Bo Zhen demonstrated all-light switching — the core logical operation of computing — using a hybrid quasiparticle called an exciton-polariton.

The energy cost is the number to remember: roughly four quadrillionths of a joule. That is, as the team noted, “far below the energy needed to briefly power a tiny LED light.” Photons normally make poor switches because they barely interact with their environment. Exciton-polaritons solve the problem by binding photons to electrons inside an atomically thin semiconductor, gaining matter’s interactive properties while keeping light’s speed.

Because they are charge-neutral and have zero rest mass, photons can carry information quickly over long distances with minimal loss. But that neutrality means they barely interact with their environment, making them bad at the sort of signal-switching logic that computers depend on.
— Li He — co-first author, formerly Zhen Lab

If the system scales, the implications run in two directions. First, AI hardware could become dramatically more energy-efficient — relevant in an industry where compute costs are now the dominant constraint on deployment. Second, photonic chips could process information directly from cameras and sensors without the energy-expensive conversion between light and electricity that today’s systems require.

HCI Reading. An ethical observation that the technical coverage tends to miss: the energy cost of AI is not only an environmental issue. It is a governance issue. Energy-hungry models can only be trained and operated by entities with vast power and capital. Energy-light models can run on devices owned by ordinary people. The exciton-polariton work, like the Stanford room-temperature device, is best understood not as a hardware breakthrough but as a redistribution-of-power breakthrough — if, and only if, it eventually leaves the lab.

Sources

  1. Wang, Z., Kim, B., Zhen, B., He, L. Strongly Nonlinear Nanocavity Exciton Polaritons in Gate-Tunable Monolayer Semiconductors. Physical Review Letters, 2026. DOI: 10.1103/gc15-qsvf.
  2. University of Pennsylvania / ScienceDaily, “Forget electrons, this breakthrough uses light-matter particles to power AI,” 18 May 2026. → https://www.sciencedaily.com/releases/2026/05/260518041341.htm

← Previous: Google’s $10 Million Bet on Biology
→ Next: Two Industrial Quantum Pilots, One Quiet Theme