# Choosing the Right GPU Server for AI Training Workloads
When a research lab or enterprise AI team asks us to spec a training rig, the first question is almost never "which GPU?" — it is "what are we training, and for how long?"
## Model size drives everything
A 7B parameter fine-tune fits comfortably on a single A100 80GB. A 70B model does not. Once you cross the single-GPU memory wall, you are paying for NVLink, InfiniBand, and the orchestration complexity that comes with them.
## Cooling is the silent killer
Mumbai data centres run warm. A dense 8-GPU chassis that works in a Bangalore colo might thermal-throttle in a Mumbai rack. We always spec headroom — typically one PCIe slot left empty and active fan curves tuned for 28°C intake.
## Our default AI server recipe
- 2 × Intel Xeon Scalable (Sapphire Rapids or newer) - 8 × NVIDIA H100 SXM5 with NVLink - 2TB DDR5 ECC - Dual 400Gb InfiniBand - 30TB local NVMe for dataset staging
This handles everything from vision transformers to 70B LLM fine-tunes. For anything smaller, we downscale to A100 80GB nodes at roughly half the cost.