MI300X rental prices and who sells them

Who lists the MI300X, and at what rate

The MI300X 192GB has a current on-demand rate at 4 providers we hold. The lowest listed rate is $2.39 per GPU-hour, at RunPod, and the highest listed rate is $3.59. Each is the seller's cheapest on-demand list rate, read from its own pricing page and hidden once it is more than 21 days old. Which seller sits at which end says nothing about which one to choose.

Where more than one seller lists the card, the highest rate belongs to Spheron. Where only one does, that name reads as no figure on record and the low and high rates are the same quote. At the lowest rate, 730 hours comes to $1,744.70. Every seller and source page is on the MI300X page.

The title asks who sells the card, so this post is careful with names. A seller appears here only through those fields, which fill from the data each time the page loads, and the card page carries the full list in rate order.

An AMD part among NVIDIA cards

The MI300X is an AMD Instinct part, and the table can set it against NVIDIA parts on two things only: the hourly rate and the memory on the card.

Its 192 GB matches the B200 and exceeds the H200 at 141 GB. The lowest listed B200 rate is $3.74 and the lowest listed H200 rate is $2.09, both covered in the B200 post and the H200 post. Where memory per card is the deciding question, those rates belong beside the MI300X range, whatever they are on the day.

Anyone moving from an NVIDIA card also has to settle whether their own code runs on an AMD part. That question sits outside every rate, and the seller is better placed to answer it on its own page than anyone reading a price table.

Memory per card is the argument to make

Weights need about the parameter count times the bytes per parameter. A 70 billion parameter model in 16-bit needs about 140 GB, which fits on one 192 GB card with room to spare. A model of about 90 billion parameters in 16-bit needs about 180 GB and still fits, just. The key-value cache and activations take memory on top, grow with context length and batch size, and cannot be given as one figure.

For a seller, that is the line to write. Someone who would otherwise split a model across several 80 GB cards may prefer one card with 192 GB, and a page that says which model sizes fit, with the cache caveat stated, does the arithmetic for the buyer before they leave to do it elsewhere.

The same arithmetic tells a seller where the argument stops. A model that fits comfortably in 80 GB gains nothing from the extra memory, and for its owner the MI300X is compared on rate alone against the lowest listed H100 SXM rate, $1.68. Knowing which buyer you are writing for decides which comparison the page should invite.

How the MI300X rows are read

We read each provider's own pricing page and keep one rate per seller for the card: its cheapest on-demand list rate, per GPU per hour. Reserved prices, interruptible capacity and quotes on request stay out. A rate whose last good read is more than 21 days old is hidden rather than shown at a stale value.

On a card listed by few sellers this rule has visible effects. If one seller drops out because its page could not be read, the count, both ends of the range and the names attached to them can all change on the next read. How the data is gathered describes the read and what is refused.

Checks for an MI300X seller

Name the part exactly. The MI325X, with 256 GB, is a different card, and a page that lists both under one heading risks a rate landing in the wrong row. Put MI300X and 192 GB together in the product name buyers will search for, and keep the same name in the page title, the table and any instance names, so every place a reader looks says the same thing.

Quote the rate per GPU per hour. A price per node is easy to publish and hard to compare, and a node price with no plainly stated card count cannot honestly be turned into a per-GPU rate. Say whether the rate is on-demand or tied to a term, beside the rate rather than in a footnote.

A seller missing from the card page, or holding a figure it believes is wrong, a claimed listing can correct it and a dispute can be filed against any value with evidence.

What the MI300X data cannot answer

It holds no benchmark data, so nothing here compares the MI300X with the B200 or the H200 on work done per hour. It holds nothing on software compatibility, and nothing on which seller has cards free today. Those are the questions a buyer asks next, and the seller's own page is where any answer has to live.

Read the seller count before quoting the range. When 4 is a small number, the market for the card on this site is thin, and each end of the range is one seller's page on one date rather than a price the market has settled on.