Access GPU capacity without buying hardware; scale resources up or down for LLM inference, fine-tuning and model training as you need.
We provide the GPU capacity your AI projects need as a cloud model: GPU virtual servers or dedicated GPU servers, delivered with storage, networking and access configuration. GPU model, capacity, data location and billing model are defined at the proposal stage based on your needs.
Both models use the same GPUs; the difference is who owns the investment, the operations and the flexibility.
| Criterion | GPU Cloud | Your own GPU server |
|---|---|---|
| Upfront investment | None, pay for what you use | Hardware purchase |
| Time to start | Short, if capacity is available | Procurement and deployment time |
| Scaling | Increase or decrease as needed | Requires buying hardware |
| Long-term cost | Better for periodic use | Better for sustained high utilization |
| Data location | Defined at the proposal stage | Fully on premises |
| Operations | On the provider side | Your team |
Rule of thumb: if GPUs will run busy most of the day for a long period, your own server is usually more economical; if usage is periodic or still uncertain, GPU Cloud is. See our GPU and LLM server page for a detailed comparison.
Resource selection starts with fitting the model into memory. Rough math for weights: parameter count × bytes per parameter (2 in FP16, 1 in 8-bit, about 0.5 in 4-bit). A 7–8B model needs about 16 GB of VRAM in FP16, a 70B model about 140 GB; add headroom for the KV cache.
Organizations that want to start AI projects without hardware investment, R&D teams with periodic heavy GPU demand, teams testing sizing before buying their own GPU server, and software companies.
A hypothetical e-commerce company wants to try an LLM that writes product descriptions. Different model sizes are tested on GPU Cloud for two weeks; once an 8-bit quantized mid-size model proves sufficient, the right capacity for continuous use is defined.
GPU Cloud means using GPU server capacity for AI and high-performance computing as a service instead of buying it. Hardware, power and infrastructure operations stay with the provider.
GPU model and capacity are defined at the proposal stage based on your workload and current availability. L4 and L40S class GPUs are considered for inference, H100 class and above for large models and training.
The billing model is defined at the proposal stage based on usage duration and whether resources are shared or dedicated. Periodic jobs and continuous capacity are priced with different models.
Data location, sub-processors and transfer conditions are agreed in writing before signing. For workloads with personal data, compliance requirements are assessed at this stage.
Yes. Because the environment is container-based, you can run your own models, open-weight models and your own application code.
Yes. Thanks to the container-based setup and standard inference engines, workloads can move to your own GPU server; usage measured on GPU Cloud is used to size that server.
