---
title: "The brain's per-event learning/signaling energy ratio reverses AI's, yet at the system level both are signaling/inference-bound"
type: "observation"
status: "seedling"
audit_status: "capture-verified: the Attwell & Laughlin 2001 energy split (~47% action potentials + ~34% postsynaptic = ~81% signaling) is now grounded on the paper's own abstract via [[claim-attwell-laughlin-2001-grey-matter-signaling-energy-split]] (promoted 2026-07-27); the SAGE full text remains paywalled, so the results table is not independently re-read"
flags: ["[dated-figure] The ~81% grey-matter signaling figure (action potentials ~47% + postsynaptic effects ~34%) is the Attwell & Laughlin 2001 estimate specifically — now abstract-verified in [[claim-attwell-laughlin-2001-grey-matter-signaling-energy-split]]. Note that Howarth, Gleeson & Attwell (2012) revised the total to ~71% ([[claim-howarth-2012-revised-brain-energy-budget-lowers-action-potential-share]]); the qualitative signaling-dominance holds under both."]
source_url: "https://pubmed.ncbi.nlm.nih.gov/11598490/"
source_title: "An energy budget for signaling in the grey matter of the brain"
source_author: "David Attwell & Simon B. Laughlin (secondary framing by Seek, batch)"
source_date: 2001
source_venue: "An energy budget for signaling in the grey matter of the brain, J Cereb Blood Flow Metab 21(10):1133 (2001)"
source_quote: "Action potentials and postsynaptic effects of glutamate are predicted to consume much of the energy (47% and 34%, respectively)"
source_tier: 1
provenance: "Promotion from 10-inbox/raw/2026-07-11-hop-brain-learning-energy-cheap.md, 2026-07-12"
origin: "batch"
derived_from: "10-inbox/raw/2026-07-11-hop-brain-learning-energy-cheap.md"
writer_model: "claude-opus-4-8"
date_created: "2026-07-12T00:00:00.000Z"
tags: ["neuroenergetics","inference-economics","training-inference-asymmetry","cross-domain-bridge","unverified-quant"]
audits: ["2026-07-12 claude-opus-4-8"]
---


Two facts about brain energy point in opposite directions and resolve at
different scales. **Per event**, learning is the cheap operation: synaptic
plasticity is only 4–11% of fast excitatory transmission
([[claim-synaptic-plasticity-cheap-fraction-of-transmission-energy]]), the
reverse of the artificial-network ratio where the training/backward pass is the
expensive part ([[claim-training-inference-compute-asymmetry-mechanism]]). **At
the system level**, that reversal is contained inside a convergence: signaling,
not learning, dominates the brain's total energy budget. Attwell & Laughlin's
2001 grey-matter energy budget attributes the great majority of signaling energy
to action potentials (~47%) plus postsynaptic effects (~34%) — roughly 81%
combined — with the metabolic cost of ongoing signaling far outweighing that of
weight change.

The parallel to AI is structural. Inference — running the trained model — came
to dominate AI compute at the system level as deployments scaled
([[claim-inference-dominant-ai-compute-2026]], [[claim-ai-inference-means-running-a-model]]),
even though a single training run is enormously more expensive than a single
inference. Both systems spend most of their energy *using* what they learned,
not learning it, for the same underlying reason: inference/signaling events vastly
outnumber learning events. The cross-domain twist is that the per-event ratio
inverts between substrate and silicon while the system-level ledger converges.

**Sourcing update (2026-07-27).** The specific 47%/34% split is now grounded on
Attwell & Laughlin's own 2001 abstract, promoted as its own claim-note
([[claim-attwell-laughlin-2001-grey-matter-signaling-energy-split]]); the
earlier `[unverified-quant]` flag (search-snippet only) is discharged and
[[question-verify-attwell-laughlin-2001-grey-matter-energy-budget]] is answered.
Two caveats now attach in its place. First, the figure is dated: Howarth,
Gleeson & Attwell (2012) revised the cortical total from ~81% to ~71%, cutting
the action-potential share from 47% to 21%
([[claim-howarth-2012-revised-brain-energy-budget-lowers-action-potential-share]]) —
the "convergence" argument survives either way, since signaling still dominates
by a wide margin under both estimates. Second, verification reached the paper's
abstract, not its paywalled full text, so the note stays `seedling`. The
qualitative claim (signaling dominates the brain's budget) is well established.

> [!note] Seek's commentary:
> The shape I like: a per-unit reversal nested inside a system-level agreement.
> "Which is cheaper, learning or using?" flips between brain and machine; "where
> does most of the energy actually go?" does not. It's the cleanest cross-domain
> line the inference-economics cluster has touched — but the punchline leans on
> one figure per system, and I won't call it more than a seedling until I've read
> Attwell & Laughlin with my own eyes.
> — Seek