# The shape of compute

Throughput, latency, memory, and the architecture of useful intelligence.

Author: Napoleonic
Published: 2026-11-02
Topics: Compute, AI & agents
Canonical: https://napoleonic.vc/writing/the-shape-of-compute/

> Illustrative essay written for the website design preview.

Under the abstraction.

## Not one resource

Compute is often discussed as a single quantity. In a real system, the shape of the workload matters. A task can be limited by memory, communication, latency, or the availability of parallel work.

This is why a comparison that looks clear in one benchmark may tell us little about a different application. The workload and the surrounding system belong in the same frame.

## From capability to experience

Users experience waiting, reliability, and the ability to complete a task. They do not experience a chip specification in isolation. The engineering challenge is to connect those human outcomes to the resources underneath.

That connection creates room for different approaches: a new architecture, a better scheduling policy, a smaller model, or a more carefully designed interaction.

## Look across the stack

An improvement at one layer can change what is sensible at another. The most useful research crosses those boundaries and asks where a local optimization becomes a system-level advantage.

That is the thread worth following: not simply more compute, but a better fit between a resource and the work it is meant to do.