NVIDIA GPU pricing tests AI infrastructure budgets

NVIDIA’s premium data-center GPUs remain among the largest line items in AI infrastructure planning. The H100 is presented at $25K-$40K, while the H200 is cited at ~$31K ($315K/8-GPU), the B200 at $30K-$50K, and the B300 at $300K-$350K. Pricing remains opaque because NVIDIA does not publish official data-center GPU price lists, leaving buyers dependent on OEM quotes, reseller listings, cloud rates and market analysis.

Hardware costs are being shaped by memory constraints as much as silicon performance. HBM shortages pushed prices higher, with Samsung and SK Hynix reportedly raising HBM3E prices by nearly 20% for 2026 deliveries. The report also notes that HBM accounts for roughly half of the B200’s $6,400 manufacturing cost, highlighting why newer accelerators carry substantial premiums.

Cloud rentals have become more competitive, with H100 on-demand rates falling to roughly $3-4 per GPU-hour by late 2025 and B200 cloud rates starting around $4-6/GPU-hr. Heavy, steady workloads can still favor ownership, while startups and variable projects often benefit from rental flexibility. Competition from AWS Trainium, Google TPU, AMD and Intel, along with changing export rules for China, could pressure future pricing even as NVIDIA retains a performance premium.