San Jose was abuzz with excitement as AI enthusiasts gathered for the 2025 NVIDIA GTC AI conference. NVIDIA showcased its expanding data center offerings, along with a commitment to joint developments with server and network vendors. Everyone had high expectations, as this is a world-renowned AI infrastructure event, and this year, it did not disappoint.

Sovereign AI led off the agenda, with UK Secretary of State for Science, Innovation, and Technology Peter Kyle highlighting the UK’s ambitious AI strategy and representatives from Denmark, India, Italy, South Korea, and Brazil also sharing their sovereign AI initiatives. Italy’s Colosseum and Brazil’s WideLabs stood out as prime examples of innovative international AI applications.

Another highlight was the collaboration between DeepMind and Disney Research that demonstrated AI’s potential to revolutionize fields such as robotics, drug discovery, and energy grids, along with the introduction of Dynamo, both as an open-source project and a framework for NVIDIA’s hardware, which promises to accelerate industrywide advancements in AI infrastructure. GTC also brought forward NVIDIA’s news of the disaggregation of NVLink, partnerships with Cisco for future telecommunications, and the expansion of its hardware certification program. Here’s a roundup of some of the most notable announcements:

  • Vera Rubin and Rubin Ultra. Jensen Huang introduced the Vera Rubin architecture, named after astronomer Vera Rubin. This next-generation GPU, launching in 2026, is designed to significantly enhance system performance. Rubin Ultra, expected in 2027, will further boost these capabilities.
  • Disaggregated NVLink. NVIDIA’s NVLink72 is an advanced interconnect architecture that facilitates ultra high-speed communication between GPUs and CPUs in large-scale computing setups. It connects 72 NVIDIA Blackwell GPUs and 36 NVIDIA Grace CPUs within a single rack, enabling them to function as a unified, massive computational resource.
  • Partnerships with Cisco. NVIDIA and Cisco are collaborating to develop an AI-native wireless network stack, focusing on radio access networks for 6G technology. This partnership focuses on performance, efficiency, and scalability in telecommunications.
  • Expanded certification program. NVIDIA’s certification program validates servers equipped with NVIDIA GPUs to handle diverse AI workloads, including deep learning training and inference tasks. The rigorous testing ensures optimal performance, manageability, and scalability. Systems from Dell Technologies, HPE, and storage providers like NetApp and VAST Data have achieved NVIDIA-certified status.
  • AI Data Center Blueprint. Recognizing the unique requirements of AI data centers, NVIDIA is partnering with vendors like Cadence, Vertiv, and Schneider to develop AI Factory Blueprints. These blueprints streamline the design, testing, and optimization of AI data centers, creating visual models to simulate and refine aspects such as power, cooling, and networking before construction, ensuring efficiency and reliability.
  • Dynamo. NVIDIA released Dynamo, an open-source framework for scalable model inferencing. Although not every organization will be inferencing models directly on their own hardware, NVIDIA aspires to become to AI what Kubernetes is to cloud. Cohere is an early explorer of this project.

Some more tactical updates:

  • CUDA-X libraries. Powered by GH200 and GB200 superchips, these libraries accelerate computational engineering tools by up to 11x and enable 5x larger calculations. With over 400 libraries, key microservices include NVIDIA Riva for speech AI, Earth-2 for climate simulations, cuOpt for routing optimization, and NeMo Retriever for retrieval-augmented generation capabilities.
  • NVIDIA Llama Nemotron reasoning. This feature enhances multistep math, coding, reasoning, and complex decision-making with Llama models. It boosts accuracy by 20% and optimizes inference speed by 5x, reducing operational costs.
  • NVIDIA Cosmos World Foundation Models (WFMs). WFMs introduce customizable reasoning models for physical AI. Cosmos Transfer WFMs generate controllable photorealistic video outputs from structured video inputs, streamlining perception AI training.
  • NVIDIA Isaac GR00T N1. New models GROOT and Newton accelerate reliable robot deployment across various industries, using real and synthetic training data. These are enhanced by the latest Cosmos WFM.

As firms build agentic AI, the need for optimized hardware to run inferencing reasoning models becomes ever more critical. Targeted inferencing frameworks such as NVIDIA’s Dynamo that are released as open-source projects are very valuable for the early movers of the agentic world, allowing for broader community co-innovation.

What It Means

NVIDIA is driving a vertical integration story based on its prowess in AI hardware and is now extending this to libraries, open–source AI models (generic and industry-specific), edge, and robotics. This certainly is good news for organizations (the idea of a one-stop shop), but business and tech leaders must address challenges extraneous to their NVIDIA relationship, such as export controls, trade sanctions that limit infrastructure availability, power requirements, business cases for AI, skills, cost increases, and risks including security, privacy, and compliance.

Specifically, power requirements for AI ambitions remains an ongoing challenge. Jensen Huang talked about how AWS, Azure, GCP, and Oracle Cloud will procure nearly 3.6 million Blackwell GPUs in 2025. In another session, Schneider execs talked about additional 150-gigawatt capacity requirements now through 2030. For reference, one rack full of NVIDIA Blackwell servers with NVLink72 requires approximately 150+ kilowatt power (compared to 10–30 kWs for traditional systems). These massive deployments across the globe require thinking outside of the box to make it all sustainable.

We are looking forward to publishing a few research reports on this market very soon. If you’re exploring AI potential and want to discuss it further, please submit an inquiry request.