Dharma Insights — Operational№ 251 · AI Systems
← The Signal№ 251 · AI Systems · March 13, 2026 · 1 min read

Most conversations about AI hardware start with one company

Most conversations about AI hardware start with one company: NVIDIA. Its GPUs power the majority of modern AI training and large-scale inference. But as AI moves beyond data centers into…

Most conversations about AI hardware start with one company: NVIDIA.

Its GPUs power the majority of modern AI training and large-scale inference.

But as AI moves beyond data centers into the physical world, a different type of processor is becoming important: Neural Processing Units (NPUs).

Unlike GPUs, which were originally designed for graphics and parallel computing, NPUs are built specifically to run neural networks efficiently.

Companies like DEEPX are designing these chips for Edge AI — where models run directly on devices rather than in the cloud.

The difference becomes clear when comparing use cases.

GPUs

Best for:
• training large models
• data-center scale AI
• massive compute workloads

NPUs

Best for:
• running AI on devices
• real-time inference
• low-power environments

A typical data-center GPU can consume 300–400W of power.

Many edge devices operate under 5–15W.

That makes GPUs impractical for systems like:

• drones
• robots
• smart cameras
• industrial machines

This is where specialized AI processors come in.

Instead of centralized AI running in a few data centers, the future may involve millions of devices running AI locally.

And that shift could reshape how AI infrastructure is built.

In the article below, I explore a bigger question:

Why energy-efficient AI hardware may become one of the most important layers of the AI stack.

View all signals →