Published on

NVIDIA's AI Revolution: Blackwell GPUs, Project DIGITS, and the Future of AI Agents

Authors
  • avatar
    Name
    Ajax
    Twitter

Jensen Huang: A Disruptor of the AI Era

At CES 2025, Jensen Huang's appearance in his signature alligator jacket was a spectacle, but the innovations he unveiled were far more impactful. NVIDIA's announcements went beyond even their own conferences, signaling a true disruption in the tech world. Let's delve into what NVIDIA is truly revolutionizing.

RTX Blackwell Series GPU: The New 'Alchemy Tool'

NVIDIA's RTX Blackwell series GPU, headlined by the RTX 5090, is making waves. While we won't delve into the detailed specs, it's worth noting that the entry-level 5070 model rivals the previous generation's 4090 in performance, but at a third of the cost. These consumer-grade GPUs are particularly suitable for locally deployed open-source models, earning the RTX 5090 the moniker of a new-generation 'alchemy tool' for AI development.

  • Performance Boost: The Black Forest Studio collaborated with NVIDIA to optimize the FLUX model, achieving significantly faster inference speeds on the 50 series GPUs.
  • DEV Model Acceleration: The DEV model’s inference speed on the 5090 is twice that of the 4090.
  • FP4 Quantization: An FP4 quantized version of the FLUX model is slated for release in February.

The pre-sale frenzy for the 5090 indicates a potential boom in AI design, AI studios, AI comics, and AI short drama production this year.

Project DIGITS: A Desktop Cloud Revolution for Large Models

Can large models with over 13 billion parameters be deployed locally? Jensen Huang answers affirmatively with "Project DIGITS," a desktop cloud platform computer capable of running 200 billion parameter models from a standard power outlet.

This technology allows seamless deployment of models developed or inferred on desktop systems to accelerated cloud or data centers. This opens up the possibility for specialized models based on personal datasets. Developers could potentially deploy 8-13 billion parameter models locally, replicating the impact Stable Diffusion had on individual creators. The $3,000 price point makes this technology accessible.

NVIDIA GB200 NVL72: A Data Center Superchip

NVIDIA has introduced the NVIDIA GB200 NVL72, a data center superchip featuring 72 Blackwell GPUs, 1.4 exaFLOPS of computing power, and 130 trillion transistors. Huang has even compared this chip to Captain America's shield.

The sheer power of this chip is staggering. Six of these chips held in Huang's hands can match the computing capabilities of entire server rooms used by many Chinese AI and automotive companies for autonomous driving. For context, the intelligent driving computing power of Li Auto is 8.1 EFLOPS. The continuous deployment of data centers utilizing these superchips promises to alleviate the shortage of computing power for the next generation of large language models, end-to-end autonomous driving, and robot world models.

Cosmos Model: Enabling AI to Understand the Physical World

NVIDIA's Cosmos model is a world model development platform designed to "teach AI to understand the physical world." It comprises a world foundation model, Tokenizers, and video processing workflows, making it a boon for robotics and AV labs.

Cosmos can accept text, image, or video prompts to generate virtual world states. This enables machines to build and understand the world conceptually. As an open-source video world model with open weights, Cosmos is trained on 20 million hours of video, with weights ranging from 4 billion to 14 billion.

While world models have various definitions, Cosmos's 4D simulation capability is unique. The immediate revolutionary impact of this technology is that synthetic data will address the big data shortage faced by physical AI. NVIDIA is already using Cosmos for large-scale synthetic data generation for robotics and autonomous driving, and has opened it up for developers to fine-tune data and train robots and AI.

Betting on Physical AI: Autonomous Driving and Robotics

NVIDIA is strategically investing in computing power, models, and data, betting that autonomous driving and robotics will be the first sectors to experience explosive growth. Jensen Huang even predicts that Robotaxis will be the first trillion-dollar robotics industry.

  • Autonomous Driving: NVIDIA has launched the next-generation automotive processor, "Thor Blackwell," which is 20 times more powerful than its predecessor and can also be used in humanoid robots.
  • Robotics: NVIDIA IsaacGroot provides developers with four key supports: foundation robot models, data pipelines, simulation frameworks, and the Thor robot computer.

NVIDIA is laying the groundwork for the "GPT moment for robots." It is anticipated that the embodied intelligence and autonomous driving sectors in China will see a surge in funding in 2025.

AI Agents: A Multi-Trillion Dollar Industry

Jensen Huang also forecasts a multi-trillion dollar AI Agent industry. A related product is Agentic AI with "Test-Time Scaling" functionality, which supports tools such as calculators, web search, semantic search, and SQL search. If NVIDIA partners with the Swarms framework for GPU-accelerated computing and AI integration, Swarms could become the dominant platform for all AI Agents. Swarms has the potential to become a trillion-dollar giant, with a current market cap of just $540 million, indicating significant growth potential.

NVIDIA's Four Stages of AI Development

Compared to OpenAI Sam's five stages of AGI development, NVIDIA's four stages of AI development are more macro and ambitious:

  1. Perception AI: Speech recognition, depth recognition
  2. Generative AI: Text, image, or video generation
  3. Agent AI: Programming assistants, etc., to help humans complete tasks
  4. Physical AI: Autonomous vehicles, general-purpose robots

This categorization clearly illustrates the development path and industrial patterns of AI. From his humble beginnings on stage a decade ago to becoming a $3.6 trillion giant, Jensen Huang's future seems limitless.