A price comparison between the NVIDIA A100 and the H100 is not a benchmark question, and no benchmark can answer it for you. What a published result can do is fix the denominator — the “per unit of work” half of cost-per-work — so the ratio you build is traceable to something measured rather than asserted. Prices move with vendor, region, contract terms, and date; none of that is measured by a benchmark, and we do not publish prices or quote deltas here.
What does “A100 vs H100 price” mean in practice?
Practically, this means one of two very different questions. The first is “what will each card cost me?” — a procurement question, answered by your own quotes. The second is “how much work do I get per unit of that cost?” — and that half is only well-formed once you name which work you mean.
Each device’s published 26Q3 LynxBenchAI result separates into three category scores: Training, Inference, and Compute. A headline figure that collapses those three into one number averages across categories your workload may never touch. Divide a price by that blended figure and you get an answer to a question nobody asked.
The single defensible move is to pick the category your workload resembles and use that category’s score as the denominator. An inference-serving fleet compares against the Inference score. A fine-tuning cluster compares against Training. A simulation or preprocessing pipeline compares against Compute.e.
The rules a cost-per-work ratio has to satisfy
| Requirement | Why it holds |
|---|---|
| Denominator is a published category score, not a headline average | A blended figure includes categories the workload never runs |
| Same release on both sides (26Q3 vs 26Q3) | Scores are produced by a named backend, driver, and framework; cross-release figures are not comparable |
| Category matches the workload, and is stated | An unstated denominator makes the ratio unauditable |
| Price side is your own quote, with basis declared | Hourly rental and purchase price are different denominators and cannot be mixed silently |
| Spec-sheet FLOPS never substitutes for the category score | Peak arithmetic throughput is not a measurement of delivered work |
If any row fails, the number you produce is a spec-sheet assertion wearing a decimal point.
Two consequences follow directly. A used or secondary-market A100 quote changes the numerator only — the published category score is unchanged by who sold the card, so the same-release constraint still governs the comparison. And when a newer generation enters the set, an H200 or B100 figure belongs in the same ratio only once it has a result published in the same release on a stated stack; otherwise you are comparing a measurement to a projection.
Where the price gap actually decides something
The gap decides something when the matching category score, not the generational uplift claim, sits next to it. Teams justify H100 migrations on uplift figures measured in a category they do not run — the most common version of this error we see is a serving workload argued from training-side numbers. If the Inference-category difference between the two devices is smaller than the price difference for your workload, existing A100 capacity may well be the cheaper source of the work you actually do. That is a conclusion your quotes and the published category scores can support together; neither can reach it alone.
If your workload resembles none of the catalogue’s categories, do not pick the nearest one and proceed. Say so, and treat the published scores as bounding context rather than as your denominator — the reasoning behind that limit is developed in our work on why hardware specifications fail to predict real AI performance, and the per-category 26Q3 results themselves are published on the LynxBenchAI leaderboard. For engineering work on the serving stack that sits behind those numbers, our GPU acceleration practice is the relevant surface.
So the open question is not which card is better value. It is whether your team can name, out loud, which of the three published categories its production workload maps to — because until it can, both price ratios are guesses.
