NVIDIA's H100 costs up to $40,000. The RTX 5090, a gaming card, costs about $2,000. For actually running AI, how much of that 15x gap is real? Less than you'd think, for most people. For any workload that fits in 32GB (models up to ~13B at full precision, ~32B quantized), the $2,000 5090 is the best performance per dollar you can buy. It is not close. Fine-tuning a small model? A run that costs ~$2 of electricity on a 5090 you own costs ~$50 to rent the equivalent cloud H100 time. The card pays for itself in a year. But the H100 earns its price the moment you cross three lines: - Memory: a 70B model needs ~40GB even at 4-bit. The 5090's 32GB cannot hold it. - Reliability: ECC memory catches the bit-flips that silently corrupt a 48-hour training run. The 5090 has none. - Scale: NVLink connects H100s at 7x PCIe speed. The 5090 has no NVLink at all. The expensive mistake is not picking the "worse" card. It is picking the wrong one for your job: overpaying for an H100 you will never fully use, or under-buying a 5090 and hitting a wall you cannot quantize around.