H200 beats H100 for AI inference.
But not for the reason you think.
We ran the same DeepSeek model on both GPUs with the same traffic for 10 days. The H200 delivered 2.5× more tokens for just 33% more rental cost.
But the biggest lesson wasn't about the GPUs. It was about how they're connected.
NVSwitch vs PCIe changed which workloads and configurations were actually possible.
So, when you're choosing AI infrastructure, don't just read the GPU spec sheet. Check the interconnect.
But not for the reason you think.
We ran the same DeepSeek model on both GPUs with the same traffic for 10 days. The H200 delivered 2.5× more tokens for just 33% more rental cost.
But the biggest lesson wasn't about the GPUs. It was about how they're connected.
NVSwitch vs PCIe changed which workloads and configurations were actually possible.
So, when you're choosing AI infrastructure, don't just read the GPU spec sheet. Check the interconnect.