NVIDIA forecasts $1T AI demand
NVIDIA told GTC attendees it expects roughly $1 trillion in cumulative AI infrastructure demand through 2027 as workloads move toward 'agentic' systems—a shift that will reshape data‑center and edge hardware priorities. The talk also previewed new inference servers and vertically integrated CPU/GPU efforts that mirror industry-wide moves toward custom acceleration stacks. (geneonline.com)
NVIDIA said Vera Rubin ships with seven new chips now in full production, including Rubin GPUs, a Vera CPU, NVLink™ 6 switch, ConnectX‑9 SuperNICs, BlueField‑4 DPUs and Spectrum‑6 Ethernet — all packaged as configurable rack designs. (investor.nvidia.com) The Vera CPU is an 88‑core Armv9.2 design with Olympus cores that NVIDIA says supports FP8 precision and high single‑thread performance, and the company lists Vera CPUs as shipping in systems from Dell, HPE and Lenovo. (nvidia.com) NVIDIA’s flagship NVL72 rack combines 72 Rubin GPUs and 36 Vera CPUs connected via NVLink 6, with company figures citing up to 3.6 EFLOPS NVFP4 inference and 2.5 EFLOPS training in the NVL72 configuration. (nvidia.com) The company struck a roughly $20 billion licensing-and‑asset arrangement with Groq that includes a talent transfer of founder Jonathan Ross and president Sunny Madra to NVIDIA while Groq continues operating independently under new leadership. (techcrunch.com) NVIDIA announced a Groq‑3 LPX inference‑accelerator rack alongside the Rubin and Vera racks, and said those inference racks will enter customer availability in the second half of 2026. (investor.nvidia.com) Microsoft’s Azure teams are reported as the first cloud provider to validate NVL72 systems, and both OpenAI and Anthropic were named as early partners in NVIDIA’s platform briefings at GTC. (winbuzzer.com) NVIDIA presented efficiency and cost metrics for Rubin NVL72 — including up to 10× lower token inference cost versus prior Blackwell racks and 20.7 TB of HBM4 capacity plus 54 TB of LPDDR5x per rack in its published spec sheets. (videocardz.com)
