Blackwell GPU rents spike
Hourly rental prices for Nvidia’s Blackwell GPUs reportedly jumped to about $4.08 from $2.75 two months earlier, a near 48% increase driven by rising demand from AI workloads. The spike — flagged by compute‑pricing indexes and market reporting — tightens the operating economics of training and inference for firms that rent rather than own hardware. That price pressure is translating hype into a concrete cost line for many AI teams. (intellectia.ai)
Renting Nvidia’s Blackwell chips just got markedly more expensive, with tracked hourly prices rising to $4.08 from $2.75 in roughly two months. (techmeme.com) The move was flagged Monday, April 13, by Techmeme’s roundup of a Wall Street Journal report citing the Ornn Compute Price Index, which tracks market prices for graphics processing unit compute. Ornn said on April 2 that its index had been live for six months and was added to the Bloomberg Terminal. (techmeme.com) (morningstar.com) A graphics processing unit is the specialized processor that trains and runs large artificial intelligence models, and Blackwell is Nvidia’s current high-end family for that work. Nvidia says the GB200 system links Blackwell processors so they can act like one large machine for large language model inference. (nvidia.com) Blackwell hardware has been moving into cloud fleets over the past year, giving renters more access to the newest Nvidia systems without buying servers outright. Nvidia said CoreWeave made Blackwell generally available in February 2025, and Microsoft published Azure documentation for its ND GB200 v6 virtual machines last week. (coreweave.com) (learn.microsoft.com) Those rentals are not small line items. Azure lists one ND GB200 v6 configuration at $108.16 an hour in West US 3, and third-party trackers show single Blackwell B200 rentals spanning from the low-$2 range on spot markets to about $6 an hour on on-demand listings. (cloudpricer.io) (datastorage.com) (deploybase.ai) The price jump lands as cloud providers pitch Blackwell for “reasoning,” agent systems, and other workloads that generate more tokens and spend longer on each request. Nvidia said that kind of processing needs more memory, compute, and high-speed links to keep responses fast. (blogs.nvidia.com) For companies that rent instead of own, that means the cost of training a model or serving answers can rise before they change a single line of code. Ornn markets futures and swaps tied to graphics processing unit capacity, a sign that compute pricing is starting to look more like an input cost that firms may want to hedge. (ornn.com) Blackwell’s appeal is straightforward: more memory, faster links, and systems designed for very large models. Microsoft says each ND GB200 v6 virtual machine uses four Blackwell graphics processing units tied together with fifth-generation NVLink, while Nvidia says its larger GB200 NVL72 rack connects 72 graphics processing units in one NVLink domain. (learn.microsoft.com) (nvidia.com) If rental markets stay tight, the practical effect is simple: the newest artificial intelligence infrastructure remains available, but it is getting harder to rent cheaply. (techmeme.com)
