This is GPU Cloud News written for people who are about to paste a credit card, not collect a quote. If you need an NVIDIA H100 this week, the catalog is large enough to waste a day. I rent these cards to train and serve models. The method below is the one I use on GPUBeacon before I open a provider.
An H100 hour is not one product. It is a model, a VRAM SKU, an interconnect, a region, and a billing flag. Mix any two of those and you are not comparing prices. You are comparing press releases. Start on the H100 listings, keep the filter on on-demand, then read the rest of this page.
What “rent an H100” actually means
NVIDIA shipped Hopper in more than one envelope. The H100 you see in GPU cloud news is usually 80 GB. That is the SKU most specialist clouds list. There is still a 40 GB variant in older hyperscaler menus. If a page says “A100/H100” without VRAM, close it.
SXM and PCIe are not the same hour. SXM is the NVLink node you want for multi-GPU training. PCIe is fine for a single-GPU fine-tune or a serving job that never leaves one card. A cheap PCIe H100 next to an SXM H100 with NVLink is not a discount. It is a different machine.
- 80 GB HBM2e is the default. Confirm it on the row, not in the hero.
- SXM when you need NVLink and a multi-GPU job.
- PCIe when one GPU is the whole product.
- MIG slices exist. They are not a full H100. Do not pay a full-H100 story for a slice.
On-demand first. Always.
GPU Cloud News in 2026 is still polluted by three numbers on one table: on-demand, spot, and “from” quotes. On-demand is the product you can start. Spot is a bid against leftover capacity. A quote is a sales cycle. I rank published in-stock on-demand hours. That is the only number I treat as a price.
On GPUBeacon, the H100 from-price in the catalog sample is $1.73 per GPU per hour. That is a marketplace on-demand row, not a hyperscaler list price, and not a promise that your region has that host. Open the live table. If the cheapest row is spot or waitlist, it is not the cheapest hour you can actually run.
I wrote a longer note on on-demand vs spot GPU cloud pricing. Read it before you take a bid that looks like a typo.
The four fields I check before the dollar
Price last. If two of the four fields differ, stop. You are not looking at the same rental.
1. GPU model and VRAM
H100 80 GB. Not “Hopper.” Not “H100 equivalent.” Not an H200 with a footnote. If you actually need 141 GB, you want an H200, and that is a different article.
2. GPU count and interconnect
One GPU, eight GPUs, NVLink, InfiniBand, Ethernet. An 8× H100 SXM node is a training box. A 1× H100 PCIe is a workstation in a rack. Multiply the per-GPU hour by the count, then ask whether the fabric is in the listing. If interconnect is missing, assume the worst fabric the provider sells.
3. Region
Latency, data residency, and egress are part of the bill. US-East cheap and EU-West legal are not interchangeable. If your weights cannot leave a country, filter the region before you fall in love with a number.
4. Stock
In stock, limited, waitlist, talk to sales. GPU Cloud News loves “H100 available.” The row either provisions or it does not. A waitlist is not a price. A quota form is not a price.
Marketplace, managed cloud, hyperscaler
I name the business model before I name the hour. Vast.ai is a marketplace. You underwrite a host. RunPod sells community and secure cloud with a control plane. Lambda sells on-demand that is actually on-demand, plus reserved clusters. AWS, GCP and Azure sell a quote theatre with a VM somewhere behind it.
- Marketplace: cheapest published H100s, host risk, checkpoint or do not sleep.
- Managed GPU cloud: less host roulette, still read the SKU, still confirm stock.
- Hyperscaler: use it when procurement already lives there. Do not use it to “win” a spreadsheet against Vast.ai.
If you need the full list, open GPU cloud providers. If you need two models side by side, use Compare. Do not paste eight vendor tabs into a sheet and call it research.
What an H100 is actually good for
Hopper is still the workhorse. 80 GB fits a lot of 70B work if you are not greedy with context and optimizer states. I rent H100 when the job is compute-bound inside 80 GB, when I need NVLink for a known stack, and when Blackwell is either a waitlist or a software tax I will not use.
- Full fine-tunes and continued pretrain on 7B–70B when the states fit.
- LoRA and QLoRA when you want headroom, not a 24 GB card at 99% VRAM.
- vLLM / TGI serving with a context window that still fits the KV cache.
- Multi-GPU jobs on SXM when you already have NCCL working.
I do not rent H100 as a default for a 7B LoRA that fits a 4090 or a 5090. That is vanity silicon. I also do not rent one lonely H100 and expect it to beat a memory-bound H200 job. Memory is a product. FLOPs are a slide.
H100 vs the cards people upgrade into
GPU Cloud News in 2026 is full of “just get B200.” Ignore that until you can name the bytes that do not fit. I compare H100 to H200 and B200 in a dedicated piece: H100 vs H200 vs B200 cloud GPU.
- Stay on H100 if 80 GB holds weights, optimizer, activations, and the KV cache you actually serve.
- Move to H200 when the extra 61 GB removes a GPU from the job or stops the swap.
- Look at B200 when FP4 and the fabric are the reason, not because the landing page is newer.
How I actually click through a rental
Open all GPUs, search H100, set on-demand. Open the H100 page. Sort by published hour. Discard rows that hide VRAM, interconnect, or stock. Open two providers, not ten. Confirm the live rate on their page. Then provision a five-minute smoke test: nvidia-smi, a tiny forward pass, disk, and the fabric if you paid for it.
Do not start a 48-hour run on a host you have never seen. Marketplace especially. Checkpoint. Measure tokens per second against the hour, not against a blog that used a different batch size.
- Filter on-demand on GPUBeacon.
- Match 80 GB, GPU count, interconnect, region.
- Prefer in-stock over a prettier number.
- Confirm the provider’s live SKU.
- Smoke-test, then train.
The traps I still see every week
- “From $X/hr” that is spot, interruptible, or a 40 GB card.
- Reserved cluster pricing sitting in an on-demand table.
- An 8× node quoted per GPU next to a 1× PCIe as if they were the same SKU.
- A waitlist with a Buy button.
- Affiliate roundups that rank the provider who pays, not the hour that is in stock.
GPUBeacon marks affiliate links. Commissions never move a row. If a GPU Cloud News post cannot say whether the hour is on-demand and in stock, it is an ad.
FAQ: renting an H100 in the GPU cloud
How much does an H100 cost per hour?
On GPUBeacon’s on-demand sample, H100 starts around $1.73 per GPU per hour on a marketplace listing. Managed clouds and hyperscalers sit higher. The live row is the answer. Open H100 prices and confirm with the provider.
Is spot H100 worth it?
Only if you can checkpoint and you can wait. If the run is the product, pay on-demand. Interrupted hours are not a discount. They are a destroyed job.
H100 or A100?
A100 80 GB still exists and can be cheaper. I take it when Hopper is waitlisted and the stack does not care. I take H100 when Transformer Engine, Hopper throughput, or the catalog’s in-stock hour wins. Compare them as 80 GB cards, not as brand names. A100 listings sit next to H100 for a reason.
Can I rent an H100 without talking to sales?
Yes, on specialist GPU clouds and marketplaces, when the row is in stock. Hyperscalers often hide the interesting SKUs behind quota. If the page asks you to book a call, you are not in on-demand anymore.
GPU Cloud News is only useful if it ends on a next click. Open the H100 catalog, keep on-demand on, and rent the hour you can actually start.
Some outbound links in these articles may be affiliate links. Rankings on GPUBeacon stay independent.
