Tesla V100 modded to PCIe card proves decades-old GPUs still viable for LLM inference at $200

A hacker has retrofit a Tesla V100 data-center GPU from 2017 onto a custom PCIe card with 3D-printed cooling, demonstrating that legacy Nvidia server GPUs can rival modern mid-range inference accelerators when optimized. The mod underscores remaining performance and power headroom in older silicon—relevant for cost-conscious inference deployments and secondary-market GPU economics.

For IT procurement, this points to opportunities for GPU workload tiering: batch inference, fine-tuning, and development environments can leverage used-market V100s and H100s at substantial capex reduction, freeing newer SKUs for production inference and training.