GPU rents spike, memory crunch bites

Renting Nvidia’s new Blackwell graphics processors in the cloud got more expensive in April, with one live pricing tracker showing the index at $4.56 an hour on April 13. (computepulse.net) Blackwell is Nvidia’s latest data-center chip family, built for training and serving artificial intelligence models, and Nvidia says a DGX B200 server packs eight Blackwell graphics processors. Nvidia says that system delivers 15 times the inference performance of its previous generation. (nvidia.com) Inference is the step where a trained model answers prompts, and Blackwell is tuned for that job at large scale. Nvidia says its GB200 NVL72 rack links 72 Blackwell graphics processors and is designed for trillion-parameter models, with up to 30 times faster real-time large language model inference than comparable Hopper setups. (nvidia.com) Spot rental prices rise when buyers need capacity now instead of waiting for long contracts or new hardware deliveries. Compute Pulse says it updates prices every 60 seconds from provider application programming interfaces, and independent trackers show Blackwell rentals spanning several dollars an hour depending on provider and term length. (computepulse.net) (tech-insider.org) The squeeze is showing up in memory too. TrendForce said on March 31 that suppliers are prioritizing higher-margin server memory, near-term supply remains tight, and limited capacity allocated to graphics double data rate memory continues to constrain graphics memory supply in the second quarter of 2026. (trendforce.com) Graphics double data rate memory is the memory sitting next to a graphics processor, while high-bandwidth memory is the stacked, ultra-fast memory used in top artificial intelligence accelerators. TrendForce said in December that, measured by wafer use, artificial intelligence demand for high-speed memory such as high-bandwidth memory and graphics double data rate 7 could absorb nearly 20% of global dynamic random-access memory supply in 2026. (trendforce.com) That pressure has spilled into consumer graphics card rumors. PCMag reported in January that Nvidia did not unveil the expected GeForce RTX 50 SUPER refresh at the 2026 Consumer Electronics Show, and trade reports have tied the delay to shortages of denser 3 gigabyte graphics double data rate 7 chips planned for higher-memory card variants. (pcmag.com) (techpowerup.com) Some of those reports say Nvidia may be steering scarce 3 gigabyte memory chips toward more profitable workstation and data-center products instead of gaming cards. TweakTown and VideoCardz both reported that the rumored SUPER cards depended on those denser modules for 18 gigabyte and 24 gigabyte configurations. (tweaktown.com) (videocardz.com) Micron, one of the biggest memory suppliers, has also signaled how tight the artificial intelligence side of the market has become. Reports on Micron’s March results said its entire 2026 high-bandwidth memory output was already sold out, locking more supply into data-center builds. (finance.yahoo.com) Put together, the April jump in Blackwell rental prices and the widening memory shortage point to the same bottleneck: companies are still bidding aggressively for the chips and memory needed to run large artificial intelligence systems right now. (computepulse.net) (trendforce.com)