Blackwell GPU rents spike

Renting Nvidia’s newest Blackwell graphics processors now costs about $4.08 an hour, up from roughly $2.75 two months ago, according to the Ornn Compute Price Index. (techmeme.com) The move amounts to a 48% jump in a market that tracks what cloud customers pay to lease a single Blackwell chip by the hour. Techmeme’s roundup, citing The Wall Street Journal, said the increase was driven by demand for “agentic” artificial intelligence systems that keep running after a prompt instead of answering once and stopping. (techmeme.com) A graphics processing unit is the specialized chip that trains and runs large artificial intelligence models, and Blackwell is Nvidia’s latest data-center generation after Hopper. Nvidia introduced the Blackwell platform on March 18, 2024, saying it was built for trillion-parameter model training and inference. (nvidia.com) Blackwell systems are designed to pack more memory and faster links between chips into each server. Nvidia says the DGX B200 server uses eight Blackwell graphics processors connected with fifth-generation NVLink, while the GB200 NVL72 rack ties together 72 Blackwell graphics processors in one system. (nvidia.com, nvidia.com) That hardware matters because renting compute is how many artificial intelligence companies get capacity before they buy their own servers. When hourly prices rise this fast, the shortage shows up first in cloud bills, not just in Nvidia’s chip backlog. (techmeme.com, morningstar.com) The index itself is new enough that Ornn AI said on April 2, 2026, that its Compute Price Index had just been added to the Bloomberg Terminal after six months as a live index. That gives investors and infrastructure buyers a market-style gauge for graphics processor prices that used to be scattered across private cloud contracts. (morningstar.com) Public cloud listings show the market is not moving in lockstep. One live pricing tracker published April 12 showed Nvidia B200 rates starting around $2.25 an hour at some providers, while another March 2026 survey listed prices near $6 an hour on RunPod and Lambda and higher on larger clusters. (datastorage.com, deploybase.ai) Supply is also constrained by the way these systems are built. Blackwell racks rely on advanced packaging, high-bandwidth memory, liquid cooling, and dense interconnects, and industry reports have pointed to packaging capacity at Taiwan Semiconductor Manufacturing Company as a bottleneck for Nvidia’s latest artificial intelligence chips. (nvidia.com, wccftech.com) For customers, the result is simple: the newest artificial intelligence compute is getting more expensive to rent even as providers keep adding Blackwell capacity. The next signal to watch is whether more supply pushes hourly rates back down, or whether sustained demand keeps cloud prices elevated. (techmeme.com, getdeploying.com)