GPU Rental Costs Soar to Record High Amid AI Compute Demand
Key Summary
Blackwell Ultra GPU rental prices have reached an all-time high for the 17th consecutive day, driven by strong demand for AI compute and supply constraints at Supermicro, the manufacturer of B300 systems. On-demand rental prices have increased by approximately 64% since November 2025, with the current price range being $7 to $10 per GPU-hour.
What the Numbers Show
On-demand B300 rentals currently run around $7 to $10 per GPU-hour. Back in November 2025, the same compute cost approximately $4.55 per hour.
That works out to a rise of approximately 64%. Anyone who budgeted for AI training runs last winter is probably revisiting that spreadsheet.
Why the B300 Commands a Premium
The B300 is not just a refresh of the prior Blackwell B200. It packs 288 GB of HBM3e memory and 8 TB/s of memory bandwidth.
The Tradeoff is Power
The B300 can draw up to 1,400 W, a meaningful step up that affects how data centers plan cooling and electrical capacity.
Rental Markets Have Shown Sensitivity to Supply Constraints
Rental markets have shown sensitivity to both Supermicro server availability and waves of AI model deployments. When a new batch of models lands and everyone wants to run them at once, a constrained supply pipeline has nowhere to give except price.
A Disconnect in the Public Markets
Despite record rental rates, public companies tied to compute leasing and mining are seeing mixed market reactions. Some are trading near their 52-week lows.
What This Means
For companies that already own B300 capacity, the current environment looks favorable. Higher hourly rates and a nearly 40% gain in hardware value over three months strengthen both their income and their balance sheets.
For AI developers and startups renting compute, an approximately 64% jump since November 2025 raises the cost of training and serving models, which may push some teams toward smaller models, older chips, or longer-term contracts to lock in pricing.