Compute and storage arms race

The race to build artificial intelligence infrastructure widened on April 22, as Google Cloud and Nvidia unveiled larger “AI factory” systems and Vast Data raised $1 billion at a $30 billion valuation. (blogs.nvidia.com) Google Cloud said its new A5X bare-metal instances will use Nvidia’s Vera Rubin NVL72 systems and scale to 80,000 Rubin graphics processing units in one site and 960,000 across multiple sites. Nvidia said the systems are aimed at training and serving “agentic” software and robotics models on Google Cloud. (blogs.nvidia.com) Nvidia and Google announced the new setup at Google Cloud Next in Las Vegas this week. Nvidia said A5X is part of Google Cloud’s AI Hypercomputer, while Google had already previewed broader Rubin support at Nvidia’s GTC conference in March. (blogs.nvidia.com, cloud.google.com) An AI model needs two basic inputs from infrastructure: processors to do the math and storage systems to keep training data, model weights, and outputs moving fast enough to avoid bottlenecks. Vast Data sells that second layer, pitching a system that combines storage, databases, and compute orchestration in one platform. (vastdata.com) Vast said its Series F valued the company at $30 billion, up from $9.1 billion in late 2023. The company said the roughly $1 billion transaction included both new capital and secondary sales, and that Drive Capital led the round with Access Industries as co-lead. (vastdata.com) CNBC reported Nvidia, Fidelity Management and Research Co., and NEA participated in the financing. Reuters reported the valuation had more than tripled since 2023 as investors poured money into companies supplying the hardware and software behind artificial intelligence systems. (cnbc.com, msn.com) The spending wave has shifted from chatbots and model demos toward the physical backbone underneath them: data centers, networking gear, chips, and storage clusters. Nvidia’s post framed that backbone as “AI factories,” a term the company uses for industrial-scale systems that generate tokens and predictions the way older data centers served web pages and databases. (blogs.nvidia.com) Google and Nvidia also tied the pitch to cost and power use. Nvidia said the Rubin-based A5X design would deliver up to 10 times lower inference cost per token and 10 times higher token throughput per megawatt than the prior generation, figures that matter as labs try to run larger models continuously instead of in short experiments. (blogs.nvidia.com) Vast said its systems already support environments “spanning millions of GPUs globally,” and named customers including CoreWeave, Lowe’s, the U.S. Air Force, and Cursor. That customer list shows how storage vendors are trying to move from a back-room role into a central position in the artificial intelligence stack. (vastdata.com) The immediate result is that compute access and data plumbing are being sold together, with cloud providers, chipmakers, and storage companies all promising bigger clusters and fewer slowdowns. On April 22, the money and the marketing moved in the same direction. (blogs.nvidia.com, vastdata.com)