Live Rental Prices for H100 and Other GPUs Increasing

Key Takeaways

• Vast.ai B200 median price around $4.5‑$4.7 per hour
• RunPod B300 instances cost roughly $6.9 per hour
• Hyperscalers charge $10‑$16+ per hour for B200 GPUs
• Spot pricing is 40‑70% cheaper than on‑demand rates
• Reserved contracts can shave 20‑50% off on‑demand costs

Pulse Analysis

The downward trajectory of H100 rental pricing reflects a maturing market where supply is finally catching up with the explosive demand for AI compute. Early 2023 saw rates near $3.00 per hour, but a combination of increased production capacity, aggressive pricing by peer‑to‑peer platforms, and the entry of newer Blackwell Ultra models has driven median B200 costs into the $4‑$5 range. This trend is tempered by a modest uptick in early 2026, suggesting that as the most efficient chips become scarce, providers may adjust rates to balance utilization and revenue.

A deeper dive into provider dynamics reveals stark contrasts. Vast.ai and RunPod, which operate on a marketplace or managed‑cloud model, consistently undercut traditional hyperscalers, offering spot rates that are 40‑70% lower than on‑demand equivalents. The trade‑off is pre‑emption risk, which many AI researchers accept for batch training jobs. Reserved contracts, meanwhile, provide a middle ground, delivering 20‑50% discounts while guaranteeing capacity—a crucial factor for enterprises with steady inference workloads. This pricing elasticity enables organizations to fine‑tune cost structures based on workload criticality and predictability.

For businesses planning AI initiatives, the evolving price landscape reshapes budgeting and vendor selection. Companies can now blend spot, on‑demand, and reserved instances to optimize spend, leveraging cheap interruptible capacity for non‑time‑critical experiments while locking in reserved capacity for production services. However, the tighter availability of Blackwell Ultra (B300) units signals that demand for top‑tier performance may outpace supply, potentially driving future price rebounds. Monitoring marketplace trends and negotiating flexible contracts will be essential to maintain competitive compute costs as the AI ecosystem continues to scale.