Compute demand and vendor partnerships rise
Reports say Grace Blackwell is shipping at volume and the industry is buying Nvidia compute faster than supply, while Cadence and Nvidia are extending their partnership to optimise large training and inference estates. These vendor narratives point to ongoing tightness in premium AI infrastructure and growing co‑engineering for large deployments. (techi.com) (edge-ai-vision.com)
Artificial intelligence training is the stage where companies feed huge datasets into chips to build a model, and inference is the stage where that trained model answers prompts for users. Nvidia says its Grace Blackwell systems are now in full production as customers keep adding more of both. (nvidia.com) Nvidia said in a blog post that “thousands” of Grace Blackwell graphics processing units are already live at CoreWeave, and described GB200 NVL72 systems as being in full production. CoreWeave said the rollout supports large model development, agentic artificial intelligence workloads and real-time inference. (nvidia.com) The company has been signaling tight supply for months. In its third-quarter fiscal 2026 results released on November 19, 2025, Chief Executive Jensen Huang said Blackwell sales were “off the charts” and cloud graphics processing units were sold out, and in fiscal 2026 fourth-quarter results Nvidia said customers were “racing to invest in AI factories.” (investor.nvidia.com 1) (investor.nvidia.com 2) That demand is showing up in the broader server market too. TrendForce said on April 8, 2026, that Blackwell’s share of Nvidia’s high-end graphics processing unit shipments is expected to rise to 71% in 2026 from 61%, helped by strong artificial intelligence demand and Nvidia’s push for rack-scale systems with more chips per deployment. (trendforce.com) Cadence sits on a different part of the stack: it sells the software engineers use to design chips, simulate systems and model factories before they are built. On April 15, 2026, Cadence and Nvidia said they were expanding their partnership to connect Cadence AgentStack, its Physical AI Stack and artificial intelligence factory digital twins with Nvidia accelerated computing. (cadence.com) A digital twin is a software copy of a real machine or facility, used to test layout, cooling, networking and operations before spending money on the physical version. Cadence said the expanded tie-up is aimed at moving customers from early design and model training to deployment faster, with tools for data centers, robotics, life sciences and semiconductor design. (cadence.com) The companies had already been deepening that work. In March 2026, Cadence said it expanded design software accelerated by Nvidia Grace central processing units and Blackwell graphics processing units, and said its Millennium M2000 supercomputer could deliver up to 80 times greater throughput and up to 20 times lower power consumption on some workloads. (cadence.com) Nvidia has also been using Cadence tools in its own chip-development pipeline. Nvidia said in May 2025 that its engineering teams used Cadence Palladium emulation and Protium prototyping systems during Blackwell design verification and chip bring-up, while Cadence used Grace Blackwell systems for fluid-dynamics simulations in aerospace. (nvidia.com) The immediate picture is not just more chip shipments, but more software and system vendors building around the same hardware base. As more customers buy rack-scale Nvidia infrastructure, companies like Cadence are positioning their tools to plan, simulate and run those larger estates before and after the hardware arrives. (nvidia.com) (cadence.com)
