AI’s energy crisis is reviving an old idea: computing with physics instead of digital logic. GPUs spend most of their power moving data, not performing arithmetic, driving data‑center demand to 290 gigawatts by 2030. Analog computing tackles this by keeping weights in a crossbar of phase‑change or resistive RAM, letting Ohm’s 경쟁 and Kirchhoff’s law perform matrix‑vector multiplication in one physical step. The IBM Research HERMES chip already delivers near‑software accuracy with 35 million cells and 17 million ॥

However, analog hardware suffers from device‑to‑device programming noise, read noise, finite ADC precision, and drift. A simulated 4‑bit ADC plus realistic noise yields over 8% relative error on a single layer. Even a conventionally trained network drops accuracy when its final layer is run through a noisy analog crossbar. Researchers patch this with periodic(runtime calibration, higher‑resolution ADCs, or hybrid digital‑analog architectures). The question remains: will the energy gains outweigh the noise‑induced accuracy loss? The field is poised for rapid progress, with startups like En Charge AI and Mythic pursuing edge inference.