In rat cortex the metabolic cost of synaptic plasticity is only 4.0–11.2% of fast excitatory synaptic transmission — per event, using a synapse costs far more than changing it

Karbowski's metabolic accounting of rat cerebral cortex estimates that "the energy cost of synaptic plasticity constitutes a small fraction of the energy used for fast excitatory synaptic transmission, typically 4.0 − 11.2%." In other words, changing a synapse (the biological act of learning) is metabolically cheap relative to using it (fast excitatory signaling): per event, transmission costs on the order of ten to twenty-five times more than plasticity.

This inverts the per-event asymmetry of artificial neural networks. In silicon, the expensive part is changing the weights: the backward pass of backpropagation costs roughly twice the forward pass, so a training token runs about three times an inference token — claim-training-inference-compute-asymmetry-mechanism. The brain shows the opposite sign at the per-event level — learning is the cheap operation, signaling the expensive one. The "training is the expensive phase" intuition is therefore a property of the backpropagation-based substrate (claim-cheap-gradient-bound-two-figures), not a general law of learning systems. It sits alongside the broader observation that the brain's learning process has no literal backpropagation analog (backpropagation-gap, claim-brain-approximates-backprop-core-principles-ngrad).

The estimate is per-event and local to fast excitatory synapses; it does not claim that learning dominates or fails to dominate the brain's total energy budget — that system-level question is treated separately in claim-brain-inference-bound-like-ai-at-system-level. Karbowski's cascade-model corollary — that longer-lasting memories require proportionally more energy to store — is a related thread not yet pursued in the vault.