info snack
NVIDIA H100
The Hopper GPU that became the default high-end AI accelerator.
companion card
H100s are 80GB Hopper GPUs, often around $2-8 per GPU-hour, built for high-throughput training and large-model inference.
What it means
H100s have 80GB VRAM and strong tensor performance. They are widely used for training, fine-tuning, and serving larger language models.
Why product teams care
H100s can make sense for high-throughput 70B-class serving, fine-tuning, or workloads where speed matters enough to justify the price.
Understudy angle
Understudy can compare H100 routes against cheaper A100, L40S, or open-provider routes on the same workload.
take this with you
H100 is the familiar premium default, but not every workflow needs it.
Do not buy H100 latency until an eval proves the cheaper route is insufficient.