Compute is the new choke point
Customers are trying to reserve huge blocks of cloud AI capacity, turning compute itself into a scarce product rather than a simple input, and AWS is pushing its Trainium chips in response. (networkworld.com) Nvidia still faces data‑centre power limits but reportedly has enough inventory to meet demand through 2027 for Blackwell and Rubin accelerators, keeping supply tight but steady. (aol.com) At the same time AMD is making gains with aggressive launches, suggesting buyers may slowly get more alternatives to Nvidia over time. (ibtimes.com.au)
Artificial intelligence builders are no longer just buying chips or cloud time. They are trying to lock up entire blocks of computing capacity years in advance, and Amazon said two customers asked to reserve all available Graviton capacity for 2026. (aboutamazon.com) A cloud region’s real limit is now power, buildings, and installed servers, not a line item called “compute.” Amazon said Amazon Web Services added 3.9 gigawatts of new power capacity in 2025 and expects to double total power capacity by the end of 2027. (aboutamazon.com) That helps explain why Amazon is pushing its own chips harder. Its new Trainium3 UltraServers, announced in December 2025, promise up to 4.4 times higher performance and 4 times better performance per watt than Trainium2 systems. (aws.amazon.com) In plain terms, large language models need vast racks of accelerators, electricity, cooling, and networking to train and answer prompts. When customers reserve those racks months or years ahead, cloud capacity starts to look more like leased industrial infrastructure than on-demand software. (aboutamazon.com) Nvidia still sits at the center of that buildout. At its March 16, 2026 GTC event, the company said it had visibility into at least $1 trillion of Blackwell and Rubin demand through 2027. (nvidia.com) Nvidia is also framing the next bottleneck as factory scale, not just chip shipments. On March 16, 2026, it said the Vera Rubin platform was in full production with seven chips designed to build larger “artificial intelligence factories” for training and inference. (investor.nvidia.com) Amazon argues the market will not stay single-sourced forever. In his April 9, 2026 shareholder letter, Andy Jassy said cheaper custom silicon, including Trainium, and a more competitive chip market should reduce artificial intelligence costs over time. (aboutamazon.com) Advanced Micro Devices is trying to speed up that shift. At its June 12, 2025 Advancing AI event, the company put the Instinct MI350 family into production and said the MI400 platform would follow in 2026 as part of a full rack-scale artificial intelligence system. (ir.amd.com) Advanced Micro Devices had already said in October 2024 that its MI400 series was expected in 2026, extending a yearly data-center accelerator cadence aimed directly at Nvidia’s release cycle. (ir.amd.com) The immediate shortage, though, is not demand for artificial intelligence itself. It is the slower work of adding substations, data centers, and enough installed machines that no single customer can ask to buy the whole cloud. (aboutamazon.com)
