AI HARDWARE, EXPLAINED

Start with what you want to do.

Choose your work. Set a limit. See what fits, what doesn’t, and why.

See my recommendation ↓

01

What do you want to do?

Pick as many as you need. These are common workloads, not a ranked popularity chart.

02

Your new-hardware budget

$USD

Use what I own$5 million

A ceiling, not a target. Hardware estimates exclude tax, shipping and facility work.

03

How you want it to work

What matters most?

How demanding are your projects?

This sets planning assumptions. You can inspect every memory allowance in the results.

What needs to stay running?

The result tells you the facility required and estimated parallel capacity. Start with a setup that fits one instance of each selected model family under your chosen residency mode.

Model sizes & advanced options

The defaults suit smaller local models. Open this when you need larger models, original-weight precision, a team service or specialized infrastructure.

What kind of operation?

Managed service restricts new language-model hardware to professional multi-GPU workstations or servers. Facilities are an output, never a hidden filter.

Inference language-model size

How much text at once?

How should inference weights be stored?

Quantization reduces weight precision, not video resolution. Published original weights are not always 16-bit. More precision does not guarantee a visibly better result.

Video delivery target

Add machines or worker fleets when useful · up to 128 nodes Split one model across GPUs / include experimental ports Consider slower CPU/RAM offload where supported Consider preorders and out-of-stock hardware

Leave experimental ports and distributed setups off for the more straightforward software path. Exact model support still needs checking.

04

What you already own

Existing devices cost $0 in the plan. Their memory stays separate.

THE OFFLINE CATALOG

From the desk to the data center.

Representative hardware families, with sources and explicit limits. Not every retail configuration.

Hardware class

NO CHATBOT REQUIRED

You should be able to question the answer.

Every result comes from bundled data and visible rules. No account, API key or model subscription.