Research

Notes from the lab: compression theory, model behaviour, and what we are learning as we build.

A logarithmic energy ladder where the operations AI runs today, from a 640 pJ DRAM read down to a 0.03 pJ addition, cluster near the top while the Landauer thermodynamic floor of 2.9 zeptojoules sits far below, the vast gap between them labelled as 10 thousand to 1 million times of unclaimed headroom.

Research

There's Plenty of Room at the Bottom of the Energy Budget

Training GPT-3 consumed about 1,287 MWh; thermodynamics says the same bit-operations could in principle have cost about eight watt-hours. A full accounting of that gap, what closes it, and what we tested that failed.

Gaurav Gandhi Aug 10, 2026 41 min read 9,064 words