# AI Hardware Directory

> A curated directory of AI hardware: data-center accelerators, edge and embedded inference chips, GPU clouds, and the server systems built around them. Editorial, unpaid, and checked.

Site: https://aihardwaredirectory.com/
Listings: 26
Updated: 2026-09-11

## What this site is
Training and inference hardware is the most expensive decision in an AI stack and the hardest to reverse. This is a plain directory of who makes the silicon, who rents it by the hour, and who packages it into machines — grouped so you can compare like with like.

## Machine-readable data
- https://aihardwaredirectory.com/entries.json  — the full listing set as JSON (name, url, category, blurb, tags).
  Stable shape. Attribution appreciated; a link back to the listing page is the only ask.

## Editorial policy
Listings are editorial and unpaid. Inclusion is not an endorsement. Outbound links are
checked to resolve before publication. Advertising is labelled and never determines
editorial placement.

## Categories
- Data-center accelerators
- Edge & embedded
- GPU cloud
- Systems & servers
- Photonic & novel

## Questions this site answers directly

### What is the difference between a GPU and an AI accelerator?
A GPU is a general parallel processor that happens to be excellent at the matrix maths neural networks need. A dedicated AI accelerator is purpose-built for that maths and usually trades flexibility for throughput or efficiency. GPUs win on ecosystem and software maturity; accelerators often win on performance per watt for a narrower set of models.

### Should I buy hardware or rent GPU cloud?
Rent while your workload is still changing shape, because the wrong purchase locks up capital in silicon that depreciates fast. Buy when utilisation is high and steady — sustained usage above roughly half of capacity is the point where owning usually beats renting, though the exact crossover depends on your power and hosting costs.

### What does 'inference at the edge' mean?
Running a trained model on a device near where the data is created — a camera, a vehicle, a factory sensor — instead of sending the data to a data centre. It lowers latency, cuts bandwidth cost, and keeps data local, which is often the deciding factor for privacy or regulation.

### Why is memory bandwidth talked about more than raw compute?
Large models spend much of their time moving weights rather than multiplying them. When a chip cannot feed its own compute units fast enough, headline throughput numbers are unreachable in practice. For large-model inference, memory capacity and bandwidth usually constrain real performance before the arithmetic units do.

### Does this directory take payment for listings?
No. Listings are editorial and unpaid. Advertising on the site is labelled as advertising and has no bearing on what gets listed or where it appears.

## Contact
privacy@aihardwaredirectory.com · legal@aihardwaredirectory.com