The write-cost bottleneck: energy constraints on continual learning in brains and machines

In recent decades, neural energy-budget research has extensively quantified signaling costs-the "read" side of computation-including action potentials, synaptic transmission, and ionic homeostasis. Yet the metabolic cost of learning itself remains far less studied. The "write" side-encompassing synaptic plasticity (LTP, LTD, and homeostatic scaling), receptor trafficking, protein synthesis, spine remodeling, and long-term consolidation-is still poorly understood. We argue that these write costs, although smaller than signaling costs in their time-averaged share of the mature brain's energy budget, constitute a real peak-throughput bottleneck for any intelligent system that must learn continuously within fixed energy constraints. A durable synaptic update aggregates an order-of-magnitude estimate of ~ 10⁴-10⁶ ATP per synapse across multiple coincident processes (receptor trafficking, protein turnover, local translation, actin remodeling, Ca²⁺ handling, and concurrent ionic recovery), with the range reflecting both natural variability and a 1-2 order-of-magnitude uncertainty across independent estimates (Karbowski, 2019). This per-synapse multi-process sum is larger than the protein-synthesis component alone-a term that standard energy budgets typically fold into housekeeping rather than treat as a distinct plasticity cost. Dense simultaneous writes are therefore unlikely to be sustainable at scale, because active cortical and hippocampal tissue can operate close to its oxidative ceiling during high-demand states, suggesting that many mature cortical circuits operate with limited additional metabolic headroom during high-demand states, plausibly on the order of tens of percent rather than several-fold (Hyder et al., 2013; Yu et al., 2023; Watts et al., 2018). The brain manages this constraint across multiple regimes: awake activity is characterized by sparse population coding and asynchronous irregular firing that preserves metabolic headroom; durable plasticity also occurs during waking but is sparse, neuromodulator- and attention-gated, and supplemented by brief awake replay events; and sleep provides additional write windows-slow-wave sleep offers a low-cost offline window, while REM sleep provides a selective plasticity window in which restricted circuits sustain patterned activity supporting durable plasticity. We recast learning as the selective allocation of scarce writes-metabolic investments that reorganize circuit structure and can reduce future signaling costs at the population level. Gated plasticity, synaptic caching, and sleep-dependent consolidation are, under this view, write-management strategies. At the systems level, selective attention allows the brain to concentrate plasticity on the circuits engaged in a current task while other regions idle, further constraining writes in space as well as time. By contrast, backpropagation in modern machine learning tightly couples reads and writes, which increases update traffic, amplifies interference, and raises energy consumption. We argue that imposing biologically inspired write-cost constraints during training, rather than relying on post-hoc compression, offers a plausible direction for energy-efficient continual learning in machines. Our claim is not that write costs dominate the brain's average energy budget, but that durable plasticity imposes a locally clustered peak-throughput demand that must be scheduled within the limited metabolic headroom left by ongoing signaling.