The directory

Where to buy a GPU for local AI

Vendors listing graphics cards suitable for running language models on your own hardware. Every listing names the exact board you would receive and links to what that board can actually run.

0

vendors listed

0

cards for sale

0

in stock now

0

shipping countries

0 vendors

Updating
Vendor Ships from Offers Website View vendor

No vendors found.

Try a different country, or clear the search.

Buying for inference

What this directory is

Every vendor listed here sells graphics cards suitable for running language models on your own hardware. That is a narrower claim than it sounds: a card is on this site because we hold its full specification and can work out which models it can hold and how fast it will generate text on them. If you are shopping for hardware to run an LLM locally rather than to play games, this is the part of the market that matters.

A listing is attached to a specific board variant rather than to a card name, which is a distinction that matters more here than almost anywhere else. The same name is often sold at several memory capacities, and capacity is what decides whether a model fits at all — so the listing you are looking at names exactly the board you would receive, and links to the page showing what that board runs.

Buying a GPU for AI is not buying a GPU for games

The ordering that matters for inference is almost never the ordering a gaming benchmark produces. Generating a token means reading the model's weights out of memory once, so the two specifications that decide the outcome are how much memory the card has — which sets whether a model fits at all — and how fast it reads that memory, which sets how quickly the tokens come out. Core counts and clock speeds barely enter into it.

The practical consequence is that a slower card with more memory frequently beats a faster card with less, because the faster card simply cannot hold the model. It is the one place where the cheap-looking option is often the right one, and the reason a card page on this site leads with capacity rather than with a benchmark score.

Work out which card you want before you start comparing prices. Open any listing here and you can see what that exact board runs and roughly how fast, which is a better basis for a purchase than a headline number on a box.

How to compare vendors here

Price is the obvious axis and the least complete one. A listing shows the price and the shipping cost separately because both are money spent to own the card, and a low price with expensive delivery from another continent is not the bargain it looks like. Where a card page shows several listings, they are ordered by that total rather than by the headline figure.

Beyond price: whether the card is in stock right now, what condition it is in, how long the warranty runs, which country it ships from, and whether a power supply is included — which is worth more than it sounds on the largest cards, since they often need more than a typical desktop supply provides. A vendor's own page collects everything they sell in one table, so you can see whether they are a specialist or listing a single card.

What we do, and what we do not

We do not sell hardware, take payment, hold stock, or handle any part of a transaction. Every listing is published by the vendor themselves and every purchase happens on their site, under their terms. Nothing here is a recommendation of a particular seller, and no vendor pays to appear or to rank higher — the order is verified vendors first, then whoever lists the most.

Prices are shown as the vendor supplied them and can go stale between their site and this page, so treat the figure here as a starting point rather than a quote. Outbound links carry no ranking signal, which is a deliberate choice: a directory that passed its authority to every listing on it would be a link farm rather than a reference.

What we do add is the part that is hard to find anywhere else — the answer to what a given card will actually run, calculated rather than quoted, sitting one click from the offer.

Step by step

How to buy a GPU for running AI models

The order that saves the most money is not the obvious one: work out what you need to run before you start comparing prices.

  1. 01

    Decide what you need to run first

    Start from the model you want rather than from a budget. The model catalogue lists every model on file, and each one names the cards that can hold it — which turns "how much should I spend" into a much more answerable question.

  2. 02

    Check the card against that model

    Every card has its own page listing the models it can run, the memory each needs, and an estimated tokens-per-second figure. Set the context length you actually work at before reading it off: a long conversation needs considerably more memory than a short one.

  3. 03

    Find the vendors listing that card

    Listings appear on the card's own page, cheapest total first. From there you can open the vendor to see everything else they sell, which is worth doing if you are buying more than one card.

  4. 04

    Filter the directory by where you are

    Shipping cost and delivery time depend on where the card is coming from. The country filter narrows the directory to vendors shipping from a given country, which is usually a faster route to a sensible shortlist than comparing prices first.

  5. 05

    Compare the total, not the price

    Add shipping to the price, then weigh condition, warranty length and whether the card is in stock. A used card with a year of warranty left from a vendor two days away is often a better purchase than a new one arriving in three weeks.

  6. 06

    Buy from the vendor, not from us

    Every listing links out to the vendor's own product page, where the purchase happens under their terms. Check the price still matches before you pay — listings here are as the vendor supplied them and can go stale.

Answers

Common questions

01

Which GPU should I buy to run AI models locally?

The one with enough memory for the model you want, before anything else. Memory capacity decides whether a model fits at all; memory bandwidth decides how fast it generates once it does. Open the model you have in mind and it will name the cards that can run it, smallest that fits and fastest first — then come back here to find who sells them.

02

Do you sell the graphics cards yourself?

No. We hold no stock, take no payment, and are not part of any transaction. Every listing is published by the vendor and every purchase happens on their own site under their terms.

03

Do vendors pay to be listed or to rank higher?

No. The order is verified vendors first, then whoever lists the most cards, then alphabetically. Nothing about that ordering is for sale, and outbound links to vendors carry no ranking signal.

04

What does the verified badge mean?

That we have confirmed the vendor is who they claim to be. It is a check on identity, not an endorsement of any individual listing or a guarantee about a sale, and it is why verified vendors sort to the top.

05

Are the prices up to date?

They are as the vendor supplied them, and they can go stale between their site and this page. Treat a price here as a starting point and check it on the vendor's own page before paying. Prices are all shown in a single currency, so compare within a vendor rather than across borders without checking.

06

Is a used GPU worth buying for AI inference?

Often, yes. Nothing about generating tokens is especially hard on a card, and memory capacity — the specification that decides what you can run — does not degrade with age. A previous-generation card with more memory frequently runs models a newer, smaller card cannot load at all. Check the warranty and condition on the listing, both of which vendors state here.

07

How do I list my own GPUs here?

Register as a vendor, complete your profile, and add listings from your vendor settings. Each listing is attached to a specific board variant from our catalogue, so it appears on that card's page next to the answer to what the card can actually run.

Before you buy

Work out what you need first

Start from a card you are considering, or from the model you want to run. Both directions end at the same answer.

Selling hardware

Interested in becoming a vendor?

List your GPUs for sale right next to the answer to what they can actually run.