GPU class · 141 GB
H200
What does H200 cost and what fits on it?
H200 has 141 GB and costs $4.55 per hour to train on, $5.25 to serve. 41 models in the catalogue fit half-precision LoRA on it and 2 more fit in four-bit.
The frontier class. A hundred and forty-one gigabytes is what large mixture-of-experts models want during training, when total parameters have to be resident even though only a fraction are active per token. Also the class to use when context length rather than parameter count is the binding constraint.
When to pick this class
The frontier class, and the specific thing it unlocks is training large mixture-of-experts models, where total parameters have to be resident even though only a fraction are active per token. A hundred and forty-one gigabytes is what that shape of model wants.
It is also the class to use when context length rather than parameter count is the binding constraint, because the attention cache grows with sequence length and concurrency. If neither of those describes your run, the 80GB classes will do the same work for considerably less.
Rates
| VRAM | 141 GB |
|---|---|
| Training | $4.55 / GPU-hourMetered per GPU-second, one-minute floor |
| Serving | $5.25 / GPU-hourWhile a replica is resident |
| Warm for a day | $12624 hours resident, regardless of traffic |
| Warm for 30 days | $3,780Set the endpoint to scale to zero if nobody is waiting |
| Models — LoRA | 41 |
| Models — QLoRA only | 2 |
Largest models that fit for LoRA
Half precision, frozen base, within 141 GB.
Llama 3.3 70B
70BNeeds 140 GB · Llama 3.3 Community
Qwen3 32B
32BNeeds 64 GB · Apache 2.0
OLMo 3 32B
32BNeeds 64 GB · Apache 2.0
Qwen3 30B-A3B
30BNeeds 64 GB · Apache 2.0
Gemma 3 27B
27BNeeds 56 GB · Gemma Terms of Use
Mistral Small 3 24B
24BNeeds 48 GB · Apache 2.0
LFM2 24B-A2B
24BNeeds 48 GB · LFM Open
gpt-oss 20B
20.9BNeeds 40 GB · Apache 2.0
Models that need four-bit training here
These exceed 141 GB in half precision but fit quantised. Quantised training is slower per step, so compare the total run cost against the next class up rather than the hourly rate alone.
Compared with its neighbours
The classes immediately either side of H200 on capacity and rate.
| Class | VRAM | Training | Serving | Versus this one |
|---|---|---|---|---|
| A100 80 GB | 80 GB | $1.68 | $1.94 | 61 GB less, $2.87/hr cheaper |
| H100 80 GB | 80 GB | $2.59 | $2.99 | 61 GB less, $1.96/hr cheaper |
Frequently asked questions
How much does H200 cost per hour?
$4.55 per GPU-hour for training and $5.25 for serving, metered per GPU-second above a one-minute floor. A continuously warm endpoint on this class is about $126 a day, or $3,780 over thirty days.
What models fit on H200?
41 models in the catalogue fit half-precision LoRA training within 141 GB, and 2 more fit in four-bit. 43 can be served from this class before accounting for the attention cache.
Is H200 the cheapest option?
Cheapest per hour is not the same as cheapest per run. Total cost is the rate multiplied by wall time, so a faster class that finishes sooner often costs less — particularly when the cheaper alternative would require quantised training, which is slower per step.
Last verified 6 August 2026.
Start with the free tier
A magic link creates your account, your tenant and your first API key. No card until you ask for compute.