Adapt open-weight models to your data in an afternoon
Most fine-tuning fits on one machine. Launch an H100 or H200 in seconds, run LoRA or full fine-tunes with the frameworks you already use, and stop paying the moment the job exits.
Recommended start
H100 SXM $2.69/hr
141 GB
HBM3e on H200 for larger batches
1 second
billing granularity
60 s
minimum charge per run
What makes this workload expensive
The bill is rarely about the GPU price alone. These are the problems we design around.
Memory, not compute, is the wall
Optimizer states and activations decide whether a run fits. Picking the wrong card means OOMs or wasted money.
Short, bursty jobs
Fine-tunes run for hours, not months. Hourly rounding and idle nodes quietly double the bill.
Reproducibility
Drivers, CUDA, and framework versions drift between machines and break runs that worked yesterday.
Built for the shape of the job
Right-sized memory
80 GB H100 and A100, 141 GB H200, 180 GB B200 — choose by what your optimizer states need, not by what is in stock.
Per-second billing
Jobs that end at 2h 07m are billed for 2h 07m. Idle instances can be stopped without losing their volumes.
Prebuilt images
PyTorch and CUDA images keep the stack identical between your prototype and your production run.
Datasets on NVMe
Keep datasets and adapters on block volumes that move between instances as you change GPU type.
Where to start, where to scale
H100 SXM
Start here80 GB VRAM · 1× NVIDIA H100
$2.69/hr
H200 SXM
141 GB VRAM · 1× NVIDIA H200
$4.39/hr
A100 80GB SXM
80 GB VRAM · 1× NVIDIA A100
$1.69/hr
From first run to production
-
1
Pick a card by memory: estimate weights + optimizer states + activations.
-
2
Launch from the console with a PyTorch image and attach your dataset volume.
-
3
Run LoRA or a full fine-tune; checkpoint adapters to the volume.
-
4
Stop or terminate — billing ends with the second.
Guides for this workload
August 28, 2026
FP8 training on Hopper and Blackwell: a practical guide
June 9, 2026
Why VRAM, not FLOPS, limits your batch size
May 12, 2026
Fine-tuning Llama 3 70B on a single 8×H100 node
Yes — LoRA on a single 8× H100 node is a common setup. Our guide walks through the memory budget.
Yes. Block volumes persist independently and can be re-attached to a different instance type.
Anything that runs on Linux with CUDA: PyTorch, JAX, Hugging Face libraries, DeepSpeed and more.
Other workloads
Start training in minutes
Create an account, add a card, and launch your first GPU instance. Per-second billing means you only pay for what you use.
No minimum commitment · Cancel anytime · $10 free credit for new accounts