Renting Nvidia’s top-tier AI hardware just got pricier for the 17th day in a row. As of October 15, 2026, on-demand rental prices for Blackwell Ultra (B300) GPUs have pushed to an all-time high.
The figures come from Compute Desk, which runs an index built on verified rental transactions from US neocloud providers. Matthew Sigel, head of digital assets research at VanEck, flagged the trend and pointed to two drivers: strong demand for open-source models and supply constraints at Supermicro, which produces B300 systems.
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.
The hardware itself has gotten more valuable too. Mid-October reports indicate the B300 has gained nearly 40% in value over the past three months.
Compute Desk’s index matters here because GPU rental pricing has historically been opaque. Many deals happen privately, and posted list prices do not always match what customers actually pay. An index anchored to verified transactions gives the market something closer to a real price signal.
The neocloud providers it tracks are a newer class of cloud company. Unlike general-purpose cloud giants, they focus narrowly on renting out GPU capacity for AI workloads, which makes their pricing a fairly direct read on demand for compute.
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.
Because Supermicro builds the B300 systems, any bottleneck on its end flows straight into rental availability.
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.
Sigel’s emphasis on open-source models is worth noting. Open models can be downloaded and run by anyone with access to hardware, which spreads demand across a much wider pool of buyers than a handful of proprietary labs.
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.
The rental boom also does not map neatly onto any specific digital asset. Investors looking for a clean crypto proxy for AI compute demand will not find one in this data.
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.
Supermicro’s production pace is the clearest variable to monitor. If its supply constraints ease, rental prices could find a ceiling. If they persist while new model deployments keep arriving, the Compute Desk index has room to keep printing records.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.