GPU compute remains tight
GPU compute is still in short supply — Blackwell instance rentals are up about 50% since January as demand outpaces available hardware (x.com). Social posts tracking cloud and rental markets say compute shortages persist even as organizations scramble for inference capacity (x.com).
Companies still cannot get enough rented graphics-processing-unit capacity, even after NVIDIA’s Blackwell chips began reaching cloud providers in volume. (semianalysis.com) A graphics processing unit, or GPU, is the specialized chip that runs many artificial-intelligence calculations at once, like adding more checkout lanes to handle a rush. NVIDIA said on March 18, 2024 that Blackwell was built to run large language models at lower cost and energy than the prior Hopper generation. (nvidia.com) Blackwell hardware is now broadly landing in the cloud, but mostly through limited new instance families rather than a glut of spare machines. AWS made its Amazon Elastic Compute Cloud G7e instances with NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs generally available on January 20, 2026. (aws.amazon.com) AWS said in March 2026 that it plans to add more than 1 million NVIDIA GPUs across its regions starting in 2026, a sign that the largest cloud operators are still racing to build supply rather than declaring the shortage over. The same announcement tied that buildout directly to growing demand for model fine-tuning and inference, the step where a trained model answers live user requests. (aws.amazon.com) That demand has shifted from training giant models to serving them continuously. AWS said G7e is aimed at generative-artificial-intelligence inference, and NVIDIA has pitched Blackwell as a way to speed token generation, the word-by-word output users see in chatbots and coding tools. (aws.amazon.com, nvidia.com) Cloud providers are also advertising bigger connected clusters, not just faster single servers. NVIDIA said CoreWeave’s GB200 NVL72 setup links 72 Blackwell GPUs and 36 Grace central-processing-unit chips in one rack-scale system, and can scale to as many as 110,000 GPUs with InfiniBand networking. (nvidia.com) CoreWeave had already made RTX PRO 6000 Blackwell Server instances generally available on July 9, 2025, saying the systems could be configured with up to eight GPUs per instance. The company said those chips were designed for large-language-model inference, fine-tuning and graphics workloads. (coreweave.com) New supply is arriving from smaller providers too, but their own announcements still describe the market as constrained. IREN said last month that it was expanding its artificial-intelligence cloud with 2,400 NVIDIA Blackwell GPUs and called the market “supply-constrained” as demand from enterprises, cloud operators and hyperscalers kept rising. (markets.ft.com) Independent market trackers are still framing 2026 as a shortage story. SemiAnalysis published “The Great GPU Shortage” on April 1, 2026 and launched a rental-pricing dashboard the same day, underscoring that the market is now being watched machine by machine, contract by contract. (newsletter.semianalysis.com) The result is a market where better chips have not yet produced cheaper, easier access. Blackwell is in the cloud, but providers are still adding capacity as fast as they can and customers are still competing for it. (aws.amazon.com, semianalysis.com)
