GPU
Coming soonGPU compute on demand
Accelerators by the hour for training, fine-tuning and self-hosted inference. Hostacker is being built to source the capacity so you get a machine and a price.
$hostacker gpu launch --model l40s --hours 4✓Capacity matched Region Germany VRAM 48 GB Scratch NVMe 512 GBEstimate returned before launchLaunch instance? Y/nAccelerators
Pick the accelerator that fits the job
The accelerator classes Hostacker intends to offer. Rates are published once capacity is contracted.
- Preview
L40S
48 GB VRAM
- Interconnect
- PCIe Gen4
- Inference serving
- Fine-tuning up to 13B
- Image generation
- Coming soon
A100
80 GB VRAM
- Interconnect
- NVLink
- Training
- Long-context inference
- Batch embedding jobs
- Coming soon
H100
80 GB VRAM
- Interconnect
- NVLink / NVSwitch
- Large-scale training
- 70B+ serving
- Distributed jobs
Pricing is set once capacity is contracted. Join the access list to be quoted for your workload.
Use cases
Hourly capacity, not annual commitments
The billing model GPU is being built around: an instance runs for the length of the job, and the meter stops when it ends.
Serve open models
Designed to run 8B to 70B models behind your own endpoint with predictable per-hour cost instead of per-token pricing.
Fine-tune on your data
Attach NVMe scratch storage, mount a dataset and stop the instance the moment the job finishes.
Batch and offline jobs
Embedding backfills, transcription and evaluation runs that only need capacity for a few hours.
Availability will vary by region
Accelerator supply moves constantly, so no GPU class will be offered everywhere. Tell us the workload and the region you need and we will factor it into where we contract capacity first.
Get GPU capacity when you need it
Join the access list and we will reach out when capacity opens in your region.