H200 rental prices across providers

The H200 range as listed today

For the H200 SXM 141GB, the lowest listed on-demand rate is $2.09 per GPU-hour, at Beam, and the highest listed rate is $6.31. The count of providers with a current rate for the card is 26. These are list rates taken from each seller's own page, one per seller, and a rate is not a recommendation.

At the lowest rate, a card kept busy for a 730-hour month costs $1,525.70. The full list of sellers, with the page each rate came from and the date it was read, is on the H200 page.

Memory is the case a seller can make

The H200 and the H100 are both Hopper parts. The H200 carries 141 GB of memory against the H100's 80 GB, and memory is the difference a seller can point to when it sets an H200 rate. The argument is about the size of the buyer's model.

Arithmetic shows why the argument lands. Weights need roughly the parameter count times the bytes per parameter, so a 70 billion parameter model in 16-bit needs about 140 GB for its weights. That does not fit on one 80 GB card and nearly fills a 141 GB one. The key-value cache and activations need memory on top, grow with context length and batch size, and cannot be given as one figure.

In 8-bit the same weights come to about 70 GB, which an 80 GB card holds with little room left for the cache. For some models the choice between an H200 and an H100 is a choice between one card and splitting the model across several, and the rate table shows only the price half of that choice.

Setting the H200 beside the H100 SXM

For the H100 SXM 80GB the lowest listed rate is $1.68 and the highest is $6.16, shown on its own page. Next to the H200 figures, they give a seller the comparison a buyer is making: what extra memory costs per hour at each end of the list.

We do not compute a premium and this post does not state one. A premium built from the lowest rate of each card can be moved by a single seller at either end, and the sellers at the low end of each card need not be the same company. Read the figures side by side and draw the line yourself.

Sellers of both cards should check that the two pages tell one story. A page that argues memory for the H200 should say which model sizes the extra 61 GB brings onto one card, so the reader can test the argument against their own model.

How an H200 rate is read

One rate per seller per card goes in: the cheapest on-demand list rate on the seller's own pricing page, per GPU per hour. Rates not read in the last 21 days are hidden. The H200 range is therefore a current list-rate range, and a seller whose page has not been read lately is out of it rather than in it at an old price.

A seller that changes its H200 price should expect the table to follow on the next read, not on the day of the change. How the data is gathered describes that read and what is refused publication.

The rule cuts both ways for a seller. A rate that was cut last week and has not yet been read again is still shown at its old value until the next read, provided that read is under 21 days old. A rate we cannot read at all for longer than 21 days disappears, which is the outcome a seller is most likely to notice and least likely to connect with its own page.

What a seller outside the range should check

A seller whose H200 rate undercuts the low end shown ($2.09) without appearing there should check first that the page names the card as an H200 and states 141 GB. A page that says only H200 next to a node price leaves the per-GPU rate to be worked out, and we do not work it out where the card count is not plain.

A seller at or past the high end ($6.31) may be pricing more than the card. Say what the hour includes: CPU, system memory, local disk, the network between nodes. A buyer comparing a bundled rate with a bare one needs the parts listed to see the reason for the difference.

If a figure held for you is wrong, a claimed listing can correct it, and any value can be challenged with evidence on the dispute page.

What the H200 figures cannot show

Whether buyers pay anything extra for memory is not in the data. We hold list rates, not orders, and nothing about how many H200 hours sell at any price.

Coverage may also be thin on the day you read this. Where the seller count above is small, one seller changing its page moves the low or the high end, and the range is better read as a handful of quotes than as a market price.

Nor does the table see what happens after a buyer leaves it. A seller near the high end may close sales on terms, support or delivery dates that no rate captures, and a seller near the low end may lose them for the same reasons. The list is where the comparison starts, which is the reason to keep your entry in it correct.