A100 vs H100 vs H200: Real Cloud Prices, Side by Side (October 2026)

You will not find these numbers in a spec sheet, and a lot of them will not last the week. I pulled them this morning from the two GPU marketplaces I actually keep a tab open on, and I am putting them in front of you while they are still real. If you are deciding between the A100, H100 and H200 for a workload you want to run next month, the chip on paper is the easy part. The price you actually pay, right now, is the part that keeps tripping people up.

What you are actually buying

Same brand, three different eras of silicon. The A100 is the Ampere generation, 80 GB of HBM2e, roughly 2 TB/s of memory bandwidth, and a battle-tested workhorse that every fine-tuning pipeline on Earth has run on at some point. The H100 is Hopper, 80 GB of HBM3, about 3.35 TB/s, and the first chip that made FP8 a real thing for inference. The H200 keeps the H100 compute die but swaps in 141 GB of HBM3e at about 4.8 TB/s. So the H200 is not a faster H100 in the raw-FLOPs sense; it is an H100 with 76% more memory and a wider bus, which is exactly what long-context LLM inference and big-batch serving want.

That distinction matters for the price math below, because it explains why the H200 is the weird one in the table.

The numbers, as of this morning

Here is what the two providers were quoting at the time I checked (on-demand per GPU per hour, with the lowest spot quote where one exists):

| Chip | Vast.ai | RunPod |

|---|---|---|

| A100 (PCIe) | $0.47 | $1.19 |

| A100 (SXM) | $1.04 | $1.00–$1.64 (spot from $1.19) |

| H100 (SXM) | $1.59–$1.98 (spot from $1.93) | $2.59–$3.79 (spot from $2.79) |

| H200 (SXM) | $6.62 (spot $2.67) | $0.50–$4.39 (spot from $3.59) |

Two things jump out before you even pick a card. First, the A100 is still the value play, full stop. A single A100 PCIe at $0.47 an hour on the marketplace side is the cheapest serious GPU I could find this morning, and an A100 SXM will run a 70B model in 4-bit with room to spare. If your workload fits in 80 GB, buying the "newer" chip is a decision you are making for the spec sheet, not for the invoice.

Second, look at the H100 row and you can see the spread between providers is doing more than the chip is. Vast.ai was quoting the H100 SXM from about $1.59 while RunPod's floor was $2.59. Same chip, same memory, same die. That is a 40% gap on the card you do not even have to configure differently. The marketplace model is why: Vast is aggregating cards from a large pool of individual hosts, so the price tracks supply almost like a spot market, and RunPod is a managed cluster that prices for the reliability you get back. Neither is wrong; they are selling different things and charging for it.

Where I would be careful with that H200 row

I am flagging this straight, because someone is going to try to book it. RunPod showed an H200 listing starting at $0.50 an hour. That is 92% below the next cheapest H200 I could see, and it reads like a single-GPU spot grab or a quote with one card in it, not a number you can plan a training run around. Vast's $6.62 is the opposite failure mode: one listing, priced well above where the H200 has been trading, and it has not refreshed in a couple of days. My read is that H200 supply on both platforms is thin right now, so both of those numbers are volatile and you should treat either as a starting point for a conversation, not a locked-in rate. When a quote is that far from the cluster of others, the honest move is to check the listing details before you commit budget to it.

So which one is the right answer

There is no single right answer, which is the whole point of putting the spread on one line. If the model and context fit in 80 GB, the A100 is the cheapest way to get there and the H100 is what you are paying a premium for that you may not use. If you are doing long-context serving, big-batch inference, or a model that wants 141 GB, the H200 is the one that removes the tensor-parallel hop, but budget for the thin supply and verify the quote per listing. And whichever card you pick, the provider you pick is going to move your hourly cost more than you expect, because the same H100 cost almost 2.5x as much on one platform as on the other this morning.

The practical move: pull the live quote for the exact configuration and billing type you want, check how many hosts are actually offering it, and re-check before you start the job. Prices here are a snapshot, not a price list, and that snapshot was taken today.

Two quick questions, answered honestly

Is the H200 worth it over the H100? Only if your workload needs the extra memory or bandwidth. For most training and serving that fits in 80 GB, you are paying 60–100% more for headroom you are not using. For long-context or 141 GB-class models, it is the difference between one node and two.

Which provider is cheaper? Right now, for the H100 and A100, Vast.ai's marketplace was consistently lower than RunPod's managed floor. The trade-off is that the lower marketplace price comes from a larger pool of individual hosts, so availability and single-node stability are more variable than a managed cluster. If uptime and a flat, predictable rate matter more than the floor price, the managed platform is the premium you are buying.

This page compares prices collected from public provider listings on Vast.ai and RunPod, checked on October 19, 2026. Rates change frequently and the figures above are a point-in-time snapshot, not a guarantee. Some provider links on this site are affiliate links, meaning we may earn a commission if you sign up through them; that does not change the prices listed here, and we have not altered any figure to favor a partner. Always confirm the live rate on the provider's own site before starting a job.