A THOUGHT EXPERIMENT, MADE TANGIBLE

What if we measured the world’s AI compute in brains?

A simple comparison with one enormous caveat: nobody agrees how much compute a brain requires.

— theoretical brain-equivalents

Loading the latest Epoch snapshot…

01 / THE IDEA

The entire argument fits on one line.

Take the world’s estimated AI chip capacity. Divide it by an estimate for one brain. The arithmetic is easy. The assumptions are not.

WORLD CAPACITY — H100e

EACH H100e 1.979 PFLOP/s

ONE BRAIN 10¹⁸ FLOP/s

02 / TRY IT YOURSELF

Two choices.
Very different worlds.

Start with the data. Then decide how much biological detail your simulated brain needs.

1 World AI compute

Epoch’s estimate of cumulative accelerator sales, normalized to H100-equivalents.

Source snapshot loading…

2 Choose a brain-compute assumption

These are scenario anchors, not a measured conversion between brains and FLOP/s.

10¹⁸ FLOP/s

Spiking-neuron emulation scenario

Where do these values come from?

10¹⁵ FLOP/s is Joe Carlsmith’s 2020 best-guess median for compute sufficient to match human task performance. His plausible mechanistic range is 10¹³–10¹⁷. This is not a whole-brain emulation estimate.

10¹⁸, 10²², and 10²⁵ FLOP/s come from the Oxford technical report Whole Brain Emulation: A Roadmap by Anders Sandberg and Nick Bostrom (2008). It estimated real-time processing demand at spiking-neuron, electrophysiological, and metabolome simulation resolutions.

The slider connects these unlike estimates to expose how strongly the result depends on what “brain-equivalent” means.

Carlsmith 2020 ↗ Whole Brain Emulation: A Roadmap (2008) ↗

YOUR ESTIMATE

— brain-equivalents

—

03 / ANOTHER WAY TO SEE IT

One brain would need roughly — H100-equivalents.

At the selected brain estimate.

Each tile represents H100-equivalents.

OPTIONAL REALITY CHECK Peak hardware is not usable compute Add efficiency estimate

Epoch reports theoretical peak capacity from chips sold. If you want a practical scenario, apply one combined factor for deployment, downtime, and sustained utilization.

Usable share 20%

Effective estimate — brain-equivalents

04 / KEEP IN MIND

This compares operations.
Not minds.

01 H100e compares peak 8-bit operations, not memory, networking, or software.

02 Chip sales are not the same thing as deployed, available hardware.

03 FLOP/s may be a poor language for biological cognition in the first place.

What if we measured the world’s AI compute in brains?

The entire argument fits on one line.

Two choices.Very different worlds.

One brain would need roughly — H100-equivalents.

This compares operations.Not minds.

Two choices.
Very different worlds.

This compares operations.
Not minds.