The same GPU, at very different prices

The finding

29 GPU models are rented on-demand by 3 or more of the providers listed here. On every one of them, the providers disagree about the price — and on most, they disagree by more than a rounding error.

The median gap between the cheapest and dearest seller of an identical card is 3.7×. The widest is the RTX-A4000-16GB: $0.10 per GPU-hour at TensorDock against $1.04 at Runcrate, a gap of 10.4×. An H100 SXM 80GB — the card most of this market is actually built on — runs from $1.68 to $6.16 across 13 sellers, a gap of 3.7×.

These are headline on-demand rates, before any commitment discount, for the same part. Not a similar part, not a different memory configuration, and not one region against another.

What we compared, and what we refused to

Four constraints decide whether a number like this means anything, and each one removes rows we would rather have kept.

On-demand only. A spot rate is a different product and a committed rate is a different contract; putting them in one range would compare things nobody chooses between.

One rate per seller per card, their cheapest. A provider listing an H100 at four instance sizes is one seller with one headline rate. Counting each instance size separately would let a single price list manufacture a spread on its own.

Headline rates rather than regional ones. Comparing Virginia against Frankfurt and calling the difference a market would be a mistake about geography wearing the costume of a finding.

A minimum of 3 sellers. A couple of providers disagreeing is an anecdote. This bar is what reduces the set to the 29 cards compared, and it is the constraint we would relax last.

Scarce silicon is priced more consistently than old silicon

The newest parts — B200, B300 and H200 — vary by an average of 2.6× across their sellers. Everything older varies by 4.6×.

That is the opposite of the intuition that new things are priced erratically, and it has a straightforward explanation: when supply is tight and every seller is paying close to the same amount for the same scarce card, there is very little room to differ. As supply loosens, price stops tracking cost and starts tracking whatever each provider thinks it can charge — support, contract terms, inertia, or simply not having looked at a competitor lately.

Worth stating the weakness plainly: that first figure rests on only a handful of cards. It is a pattern worth watching rather than a law, and it is exactly the sort of claim that a longer price series would either confirm or kill.

Who is cheapest, and why that is not a recommendation

Across the 29 cards compared, the lowest rate tends to sit with marketplaces and community clouds rather than managed providers. The widest gap, on the RTX-A4000-16GB, is TensorDock against Runcrate.

That is a fact about price and nothing else. It says nothing about availability when you need a whole cluster at once, about what happens when an instance is reclaimed, about support, or about whether the network between your nodes is any good. Those differences are most of why the dearer providers exist, and none of them are visible in a per-hour rate.

So this is not a buying recommendation, and there is no "best value" column anywhere on this site. The number tells you what to ask about, not what to choose.

This is not an index, yet

An index tracks movement, and movement needs a series long enough that a change means something — twelve weeks, give or take. We hold a handful of capture days. A line drawn through them would be a line through noise, and drawing it would break the only rule this directory has.

Spread needs no history at all. It is a fact about today, computed from the same rows the tables show, and it is the thing an index would eventually be built to explain. When there is enough series to say whether these gaps are closing, that piece gets written.