Abhishek S.
Shipping in public. Listening in private.

Abhishek

I lead women’s Indo-Western & Premium at Max Fashion. I also wrote the AI that runs the buying floor.

Rare profile. Category operator who ships production code.

Senior Buying Leader · Max Fashion Women’s Indo-Western & Premium · 530+ India stores NIFT ’12 · Twelve years on the floor

abhishek@bengaluru ~ %
>role: senior buying lead
>dept: women’s indo-western + premium
>floor: 530+ stores india

The Brain's Energy Budget

The adult brain is about 2% of body mass but consumes roughly 20% of the body's resting oxygen and calories. Raichle and Gusnard quantified that mismatch in 2002, then pointed out the stranger part: deliberate thought changes the total surprisingly little. The brain is not switched on by a task; it redirects an energy budget already near 20 watts.

How neurons spend ATP

Neural computation is largely ion bookkeeping. Each spike and synaptic current lets sodium, potassium, or calcium cross a membrane. The sodium-potassium pump then spends one ATP molecule to move 3 sodium ions out and 2 potassium ions in.

Howarth, Gleeson, and Attwell's 2012 model assigned the signaling budget of rodent neocortex as follows:

Process Share of signaling energy
Postsynaptic glutamate receptors 50%
Action potentials 21%
Resting membrane potentials 20%
Presynaptic transmitter release 5%
Transmitter recycling 4%

The table is a model, not a meter reading. Its central claim survives later revisions: receiving and resetting signals costs more than the abstract operation performed on them.

Why doing nothing costs so much

Raichle's concept default mode network explains part of the baseline. When a scanner subject receives no task, the brain continues constructing memories, possible futures, social models, and a continuous sense of self. Rest removes an experimenter's instruction, not neural activity.

The budget also constrains coding. Peter Lennie estimated in 2003 that fewer than 1% of cortical neurons could be strongly active at once under the available energy supply. That estimate depends on firing rates and synaptic costs, but the direction is clear: sparse activity is not merely elegant representation. It is rationing.

There is almost no reserve. Complete interruption of cerebral blood flow can cause unconsciousness within 10 seconds because stored oxygen is exhausted. A liver can warehouse fuel; a cortex depends on delivery.

What's contested

The percentages have already moved. Attwell and Laughlin's 2001 model assigned 47% of signaling energy to action potentials; the 2012 revision reduced that to 21% after measurements showed mammalian spikes waste less sodium current than squid-axon estimates implied.

Fuel routing is also disputed. Pellerin and Magistretti proposed in 1994 that astrocytes consume glucose and pass lactate to neurons. Researchers agree that neurons can oxidize lactate and that astrocytes participate in metabolic support. They still contest when lactate is the main delivered fuel and when neurons take up glucose directly.

Why this has to do with other realms

Concept information theory asks how many bits a channel can carry. Neuroenergetics adds the invoice: how many useful distinctions can a circuit preserve per joule? Attention, sparse firing, and local blood-flow control look different when treated as energy-allocation mechanisms.

This also changes the comparison with concept transformer architecture. A biological brain colocates memory, communication, and computation across synapses; a digital model repeatedly moves values between separate memory and arithmetic hardware. Comparing them by operation count alone hides a major cost: moving the signal.

An open question

If energy scarcity shaped neural sparsity, attention, and wiring, what machine-learning architecture would appear if every communication edge carried a joule price from its first training run?

Key Sources

Further Reading

Abhishek's take

The number that stays with me is Lennie's estimate that fewer than 1% of cortical neurons can be strongly active at once. Intelligence may depend less on activating more machinery than on choosing which tiny fraction earns the ATP.

Tags: #brain-energy #neuroenergetics #neural-coding #metabolism #information-theory