Older GPUs: Rental Prices Spike
Rental prices for older Nvidia GPUs keep climbing despite new silicon, signaling persistent demand for 'last‑gen' hardware with high VRAM and practical value for local ML development. That tight rental market is keeping smaller teams dependent on legacy cards for training and inference. (officechai.com)
a16z’s rental chart showed the H100 RT Index climbed to $2.41 per hour and the A100 RT Index reached $1.38 per hour by late February 2026. (officechai.com) Silicon Data’s SDH100RT index — published to Bloomberg — has been tracking H100 spot rental rates since May 27, 2025. (0e190a550a8c4c8c4b93-fcd009c875a5577fd4fe2f5b7e3bf4eb.ssl.cf2.rackcdn.com) Silicon Data rolled out updates to its A100 and H100 on‑demand rental indices and launched a B200 rental index in December 2025 to broaden market coverage. (einpresswire.com) Vast.ai marketplace listings show H100 PCIe offers around $1.20 per hour while RTX 3090 nodes are appearing at roughly $0.13 per hour on the same peer‑to‑peer marketplaces. (vast.ai) NVIDIA’s A100 is available in 40GB and 80GB HBM2e configurations and the 80GB variant doubles memory capacity and pushes memory bandwidth toward ~2 TB/s for largest training workloads. (nvidia.com) Consumer/prosumer RTX 3090 cards (24GB VRAM) remain a practical option for local fine‑tuning and experimentation, and techniques like QLoRA and LoRA are routinely used to enable large‑model tuning on such cards. (fluence.network) Published cost comparisons show an 8×H100 multi‑GPU project can push compute bills past $100,000 per month, while persistent waitlists and hyperscaler price spreads have pushed teams to allocate older, high‑VRAM cards instead of buying new units. (researchsnipers.com)
