L40S rental prices across providers

The L40S rate range

An L40S 48GB rents on demand from $0.38 per GPU-hour, the lowest listed rate, which belongs to Lium. The highest listed rate is $3.51, and 20 providers have a current rate for the card. Every figure is a seller's cheapest on-demand list rate, read from its own page; a rate older than 21 days does not appear.

A month of 730 hours at the lowest rate comes to $277.40. Where more than one seller lists the card, the top of the range belongs to Replicate. Rates, sellers and source pages are on the L40S page.

With a single seller there is no top-of-range name, and that slot reads as no figure on record. The seller count above shows which case applies on the day you read this.

One letter between the L40S and the L40

The L40S and the L40 share an architecture, Ada Lovelace, and a memory size, 48 GB. They are listed here as separate cards, and a rate for one is never counted as a rate for the other. For a seller the risk is plain: a page that writes L40 where it means L40S, or the reverse, puts its rate in the wrong row.

On the L40 page the lowest listed rate is $0.45, at Vast.ai, across 9 sellers. A buyer searching for 48 GB on an Ada card will see both cards and both ranges, and the product name on your page decides which one you are compared in.

The A40, an Ampere card, also carries 48 GB and turns up in the same searches. Its lowest listed rate is $0.52. A seller of any of these does well to put the card name and the memory together in the title of the offer.

Where the L40S sits for a buyer

With 48 GB, an L40S holds the 16-bit weights of a model up to about 24 billion parameters, since weights need about 2 bytes per parameter at that precision, before any room for the key-value cache. In 8-bit the limit roughly doubles. The cache grows with context and batch, so the real limit is lower and varies by workload.

For a seller, that arithmetic is the pitch. A buyer whose model fits in 48 GB does not need an 80 GB Hopper card, and the lowest listed H100 SXM rate ($1.68) is the obvious figure to read beside the L40S range. Say on the page which model sizes the card is meant for.

How the range is put together

Each seller enters once, at the cheapest on-demand list rate on its own pricing page, per GPU per hour. Reserved and interruptible prices stay out. A page that has not been read in the last 21 days drops out of the range, so a seller who changed its page recently may be missing until the next read.

The full method, including what we will not publish, is on the data page. Nothing on it weights a seller by size or ranks one seller above another.

For this card the 21-day rule has a practical edge. Sellers who list the L40S often list the L40 and the A40 on the same page, and if that page changes layout and cannot be read, all of those rates go hidden together. A seller missing from several 48 GB rows at once should look at the page before looking at the prices.

Checks for a seller outside the range

For a rate under the low end shown ($0.38) that is missing from it, check the card name first, since an L40S filed as L40 lands in the other row. Then check that the rate is per GPU, not per instance, and readable without a sign-in. A rate that fails any of those is not wrong, only unreadable as a per-card price.

For a rate at or past the high end ($3.51), remember that an instance with generous CPU, memory or storage will sit high against a bare rate. List the parts beside the price so a buyer can see what the difference pays for, and state whether the rate is on-demand or tied to a term.

A figure held for you that is wrong can be corrected through a claimed listing, or challenged with evidence on the dispute page.

Check the date too. If the rate on your page changed recently and the table still shows the old one, the next read will replace it; if the table shows nothing at all for you, the last good read is probably older than 21 days and the page is the place to look.

What the L40S figures do not say

They do not say which card a buyer should choose. We hold no benchmark data, so nothing here compares the L40S, the L40 and the A40 on work done per hour, only on listed rate and memory.

They also hold no orders, no availability and no private discounts. A seller at the low end may be fully booked, and the table cannot tell.

They cannot say why a seller set its rate where it did. A rate near the low end may be a choice to fill idle cards, and one near the high end may carry support or a fixed location. The source page beside each figure is the only place that explains it, and only if the seller wrote it down.