Is on-demand capacity reserved for me?
No. On-demand GPUs have no capacity set aside for you, and availability differs by region. In March 2026 roughly half of GPU providers had no on-demand H100 stock at all; reserved customers were not affected. Reserved clusters allocate capacity for the full term at a fixed monthly price, 3% to 14% below on-demand rates for 3–36 months. Dedicated servers are not shared with other customers, and you choose the region when you reserve.
See: On demand vs reserved GPUs and Reserved GPU cluster cost.
What should I do right now if my preferred GPU is out of stock?
- Try a nearby region or an adjacent model. For example, H200 holds bigger models than H100, while the RTX PRO 6000 is cost‑effective for single‑GPU inference and image/video generation. See H100, H200, RTX PRO 6000, and H100 vs H200.
- Match the GPU to the workload. For LLM serving, pick by model size and context: LLM inference guide. For image/video, run as a job queue across several GPUs: Image & video generation.
- If you need a firm start date, request a reserved node and we quote available capacity at a fixed monthly price. See Dedicated GPU cluster (3–12 months).
Plan for bursts: queue and overflow
Production generation and inference run best behind a queue spread over several GPUs. Keep a small reserved cluster for steady traffic and add on‑demand GPUs for peaks or campaigns.
See: Measure before you scale and More inference per GPU.
Reserve capacity to avoid volatility
Reserved clusters allocate capacity for the full term with a fixed monthly price. Terms: 3, 6, 12, and 24–36 months (3% to 14% below on‑demand). You choose the region; dedicated servers are not shared with other customers.
Details: Reserved GPU cluster cost and Dedicated GPU cluster (3–12 months).
Third-party fallbacks
If you need a GPU in the next hour, other providers are a reasonable fallback: RunPod (Community and Secure Cloud), Vast.ai, Lambda and CoreWeave. Marketplaces and community clouds are usually the lowest hourly price but rent other people's servers and vary in reliability; dedicated servers in Tier III datacenters cost more and come with a contract and an uptime commitment. Current list prices per H100 GPU-hour are on H100 rental price.
Comparisons: RunPod alternatives, Strata vs RunPod, Strata vs Lambda.


