Contrary to investors' concerns that tightening security requirements would reduce demand for equipment, leading developers of advanced models at the AI Infra Summit stated that ensuring AI security and consistency are extremely resource-intensive processes that will drive infrastructure spending growth. This is according to a research report by Citi. The conclusion was voiced on the first day of the conference in Santa Clara: analysts noted that providing large-scale computing power, energy supply, and production facilities, not just innovations in model architecture, has become the main bottleneck for the industry. Next-generation data centers are being designed to accommodate 30–40% more GPUs, and orchestration platforms like Astra and hardware updates, particularly Nvidia Vera Rubin, are aimed at optimizing end-to-end throughput.
Ian Buck, Vice President of Nvidia for Hyperscale and HPC, stated that the transition from chatbots to autonomous agent AI dramatically increases hardware requirements. According to him, agent workloads are about 100 times more demanding than chat applications of 2023 due to the exponential growth in input sequence lengths, KV-cache sizes, multi-turn interactions, and the generation of sub-agents.
Intel CEO Lip-Bu Tan outlined the company's strategic pivot from a traditional processor supplier to a full-stack AI infrastructure provider. According to him, as competition shifts from raw GPU power to system-level optimization, processors will remain key for reinforcement learning, orchestration layers, and agent workflows. To accelerate solutions, Tan simplified Intel's management structure, reducing the number of hierarchy levels from 10–12 to 4–5. In terms of production, he confirmed that the Intel 18A process technology has entered serial production, and the 14A node is approaching market release with PDK 0.9 support in October and PDK 1.0 in the next quarter.
Related articles
