Head to head · verified Sep 5, 2026
H100 vs A100: FP8 vs a cheaper 80 GB
Same 80 GB envelope on the common SKUs. H100 is the FP8 hour. A100 is the leftover that is still in every catalog. I rent H100 when I will use Transformer Engine. I rent A100 when the job is already Ampere-shaped and the H100 delta does not buy a shorter epoch.
H100$1.73
VRAM80 / 80 GB
A100$1.19
The resultDefault: H100
The result
Default: H100
Pick: H100
H100 for new training and FP8 serving. A100 when you already have Ampere software, a lower hour, and no need for Hopper’s engine. Do not buy H100 as a status A100.
In brief
What to know before you pick
- Common SKUs share an 80 GB envelope. H100 is Transformer Engine. A100 is the leftover that still trains.
- Do not buy Hopper as a status Ampere.
- A100 40 GB and A100 80 GB are different products wearing one name.
Same fields
The comparison
Catalog facts first. The pick above is editorial. Confirm the live on-demand row before you provision.
| Field | H100 | A100 |
|---|---|---|
| VRAM | 80 GBTie | 80 GBTie |
| Memory | HBM3Tie | HBM2eTie |
| Bandwidth | 3,350 GB/sWins this row | 2,039 GB/s |
| Catalog from-price | $1.73 | $1.19Wins this row |
| Cheapest indexed on | Vast.ai | Vast.ai |
| On-demand rows in sample | 15 | 7 |
| In-stock rows | 10 | 4 |
| Architecture | hopper | ampere |
Rent H100 if
New runs, FP8/Transformer Engine, NVLink 900 GB/s on SXM, anything you will keep for months. The catalog actually stocks it.
Skip it if Context windows that blow the KV cache, multi-node jobs that need a reserved fabric, or anyone shopping a 4090 workload with Hopper money.
Rent A100 if
Inference or fine-tunes that already fit A100, a shop that will not retune kernels, or a quoted Hopper hour that is a waitlist in costume.
Skip it if Greenfield FP8 training, people who needed H200 memory, anyone offered a 40 GB A100 as if it were 80.
H100
- The SKU every catalog actually stocks. You can compare an hour, not a waitlist
- FP8 Transformer Engine is the reason this generation still trains at a sane $/step
- MIG exists when you want to split an 80 GB card instead of overpaying for a second one
A100
- 80 GB with NVLink and MIG, a real datacenter product, not a GeForce
- Deep catalog, mature images, fewer Blackwell surprises
- Often the cheapest *serious* 80 GB hour
New work belongs on H100
If you are starting a 70B fine-tune, serving with FP8, or keeping a stack for months, rent H100 SXM. Transformer Engine and memory bandwidth are the reason the hour costs more. Ampere still works. Ampere is not the default for new FP8 work.
Measure an epoch. If the H100 hour does not buy a shorter run, and A100 is honestly cheaper on-demand, stay. Vanity Hopper is how teams light money on fire.
When A100 still wins the bill
The job already fits Ampere. The software will not be retuned. The quoted Hopper hour is a waitlist. Then A100 is the adult choice, not nostalgia.
Confirm 80 GB vs 40 GB. Confirm SXM vs PCIe. A marketplace H100 PCIe next to a Lambda A100 SXM is not the same comparison twice.
Stock still decides the week
H100 is usually the deeper catalog. That is part of the product. If Hopper is waitlist this week and A100 is live, you do not wait for a badge. You train.
The trap
How this comparison usually goes wrong
Paying Hopper money for an Ampere job. Mixing A100 40 GB with A100 80 GB. Comparing a marketplace H100 PCIe to a Lambda A100 SXM as if they were the same product.
What I would do
The move
If you are starting a 70B fine-tune this week, rent H100 SXM. If the model already runs on A100 and the hour is honestly cheaper on-demand, stay. Confirm 80 GB vs 40 GB before you paste a card.
Catalog listings
H100 and A100 in this catalog
On-demand hours first. Spot and waitlists are labeled. From-prices are catalog samples, not a reservation.
ProvidersRegionBilling typeInterconnect/GPU/hr
ProvidersRegionBilling typeInterconnect/GPU/hr
Full review
The two reviews
Same product, full page. Specs, listings, who it is for. Confirm the live hour there.
Full review
NVIDIA H100
The default datacenter hour. 80 GB of HBM3 is enough for most 70B work if you manage the KV cache. It is not a bargain because the FLOPs slide is pretty. It is a bargain when the job already fits.
Read the H100 reviewFull review
NVIDIA A100
Ampere’s 80 GB workhorse. Slower than Hopper, cheaper than denial. Still the right hour when the software is stuck on A100 images or the H100 premium is a tax you will not earn back.
Read the A100 reviewFAQ
H100 vs A100 questions
Is A100 obsolete?+
No. It is older. It still trains. It is the wrong default for new FP8 work. It is the right default when the H100 hour is a waitlist or a quote.
Same VRAM, why pay more for H100?+
Transformer Engine, memory bandwidth, NVLink on SXM. If you never leave FP16 on a single GPU, the delta is thinner. Measure an epoch, not a FLOPs tweet.
H100 vs A100 for inference?+
If the model already serves on A100 and the hour is cheaper on-demand, stay. New FP8 serving: H100. Confirm 80 GB vs 40 GB first.
Can I mix A100 40 GB with H100 80 GB in one comparison?+
No. That is two products. Rank 80 GB against 80 GB, then decide whether 40 GB was ever the job.
Keep comparing