What GTC2025 Reveals About AI's Next Leap: Agentic AI, Blackwell GPUs, and Emerging Technologies

The GTC2025 conference highlighted the evolution of artificial intelligence from Perception AI to Generative AI, and now towards Agentic AI and Physical AI, emphasizing that Agentic AI will require orders of magnitude more computation than current Generative models.

Scaling Law, which has driven the growth of large language models, will continue to push model size and compute requirements, with Agentic AI expected to increase computational demand by a factor of one hundred.

NVIDIA introduced the Blackwell GPU architecture, featuring the Grace Blackwell NVLink 72 chip that integrates 72 Blackwell GPUs and 18 NVLink switches on a single wafer, delivering up to 1.4 EFLOPS at FP4 precision. Compared to the Hopper architecture, Blackwell offers a 25‑fold overall performance boost and up to 40‑fold improvement in inference workloads.

The company also announced Dynamo, an open‑source distributed inference service library described as an “AI factory operating system,” aimed at alleviating token supply constraints. Early partners include Microsoft and Perplexity.

In the hardware line‑up, NVIDIA revealed the Blackwell Ultra NVL72 chip, designed specifically for AI inference. It improves AI performance by 1.5 × over the previous GB200 NVL72 and offers up to 11 × faster large‑model inference and 7 × higher compute capacity compared to the Hopper generation.

Looking ahead, NVIDIA outlined the VeraRubin platform, slated for release in late 2026 and 2027, featuring the VeraCPU and delivering 50 petaflops (VR200) and 100 petaflops (VR300) of inference power, along with 288 GB HBM4 memory to address memory bottlenecks.

Manufacturing technology will transition from 4N/4NP nodes (2022‑2025) to more advanced 3NP nodes (2026‑2027). HBM capacity will grow from 80 GB (Hopper) to 1025 GB (VR300) with stack layers increasing from 5 to 16 and bandwidth rising from 3.35 TB/s to 32 TB/s. SerDes speeds will double to 448 Gb/s by 2027.

NVIDIA also introduced silicon‑photonic network switches: Spectrum‑X (Ethernet‑based) and Quantum‑X (InfiniBand‑based), both employing co‑packaged optics to integrate optical communication directly into the switch chassis.

In the robotics domain, NVIDIA highlighted its Physical AI initiatives, including the Cosmos physics‑world model for synthetic data generation, the GROOT N1 humanoid foundation model, and the Omniverse real‑time 3D simulation platform.

Finally, NVIDIA announced the establishment of the NVIDIA Accelerated Quantum Computing Center (NVAQC) in Boston, a research‑focused facility intended to advance quantum computing architectures and algorithms.