What Are the Four Waves of AI and How NVIDIA Is Shaping the Future?

On March 19, NVIDIA CEO Jensen Huang delivered the keynote at the NVIDIA AI conference GTC 2025 in San Jose, California. Semianalysis provides an in‑depth interpretation of the speech, outlining NVIDIA’s latest advances in AI inference performance.

Key Points

Artificial intelligence (AI) development can be divided into four waves: perception AI, generative AI, agentic AI, and physical AI, each contributing uniquely.

Generative AI is currently active but its compute demand is 100 times higher than expected, putting huge pressure on data‑center infrastructure.

Data‑center capital expenditures are projected to exceed $1 trillion by the end of 2028, supported by data trends.

NVIDIA’s Blackwell chip is in full production and an upgrade to Blackwell Ultra is planned, with reports showing significant performance gains.

New AI computers such as DGX Spark and DGX Station deliver high compute power for desktop AI development, confirming market potential.

Dynamo, an open‑source software, accelerates AI inference and improves efficiency in AI factories.

NVIDIA collaborates with Google DeepMind and Disney on the Newton robot platform, showcasing the "Blue" prototype and rapid progress in physical AI.

General Motors uses NVIDIA’s platform to optimize factories and autonomous driving, with recent partnership details released.

NVIDIA partners with T‑Mobile and others to develop AI‑native 6G networks, indicating future potential.

Artificial Intelligence Four Waves: From Perception to Physical

AI is experiencing four successive waves—perception AI, generative AI, agentic AI, and physical AI. Each wave represents a leap in AI capability, expanding from data understanding to interaction with the physical world. This article explores the definitions, current status, and impact of each stage, incorporating NVIDIA’s latest technological advances.

Perception AI: The Starting Point

Perception AI enables machines to sense and understand data, such as recognizing images through computer vision or interpreting speech via speech recognition. It forms the foundation of AI, relying on supervised learning and large labeled datasets. The 2012 AlexNet breakthrough marked the deep‑learning era, with applications in security monitoring, medical imaging, and virtual assistants like Siri. However, perception AI is limited to static data processing and cannot generate new content or perform reasoning.

Generative AI: A Revolution in Content Creation

Generative AI can produce new content—text, images, video—based on inputs, with core technologies including GANs and large language models (LLMs) such as ChatGPT and DALL‑E. Over the past five years, generative AI has transformed media, design, and education, but its compute demand has surged, consuming up to 100 times the expected resources. Challenges include ethical concerns (bias) and high costs, requiring data centers to invest heavily to support its growth.

Agentic AI: The Rise of Autonomy

Agentic AI grants autonomy, allowing systems to reason, plan, and act, such as using tools to process multimodal information. Current advances feature "chain‑of‑thought" techniques that decompose complex problems to improve answer accuracy. Applications span customer service, software development, and healthcare, with huge potential yet raising data‑privacy and security issues. Industry analysts view agentic AI as the next major breakthrough that will reshape work.

Physical AI: Interaction with the Physical World

Physical AI enables understanding of the physical world, supporting robotics and autonomous driving. NVIDIA contributes open physical‑AI datasets and the Isaac GR00T model, and co‑develops the Newton platform, showcasing the "Blue" robot prototype. Use cases include manufacturing and logistics, while challenges involve handling physical complexity; the technology is expected to profoundly influence an intelligent society.

Data Center and AI Infrastructure

Data‑center capital expenditures are projected to exceed $1 trillion by the end of 2028, reflecting the surge in AI compute demand. AI factories are shifting from traditional retrieval to generative computing, requiring more resources, a trend supported by industry reports.

Blackwell Chip and Upgrade

The Blackwell chip is now in full production, delivering a 25× performance boost while reducing power consumption. An upgrade, Blackwell Ultra, is slated for release in the second half of 2025. Its NVLink 72 architecture supports AI inference, and industry dynamics indicate it is critical for data‑center transformation.

Next‑Generation AI Computers

DGX Spark, based on the GB10 Superchip, offers 1,000 TOPS of compute power and is priced at $3,000, targeting desktop AI development. DGX Station provides 784 GB of memory and 800 Gb/s networking, with manufacturers such as ASUS and Dell. These systems make AI computing more accessible.

Dynamo: AI Inference Accelerator

Dynamo is an open‑source library that optimizes AI inference, delivering a 30× token‑generation efficiency boost and supporting frameworks like PyTorch, thereby lowering costs. Its dynamic resource management suits AI factories, and analysts see strong future potential.

Robotics and Physical AI

NVIDIA collaborates with DeepMind and Disney on the Newton platform, unveiling the "Blue" robot whose physics engine enables high‑fidelity simulation. The Isaac GR00T model gives humanoid robots generalization abilities for manufacturing and medical applications, highlighting rapid progress in physical AI.

Autonomous Driving: GM and NVIDIA Collaboration

General Motors uses NVIDIA’s platform to optimize factories and autonomous‑driving systems, leveraging Omniverse and DRIVE AGX. Recent partnership announcements underscore the strategic importance of this collaboration, which will accelerate GM’s autonomous‑vehicle fleet development and impact the automotive industry.

6G Networks: AI‑Native Wireless

NVIDIA partners with T‑Mobile and other telecom leaders to develop AI‑native 6G networks, aiming to improve spectral efficiency and support billions of connected devices. AI‑RAN technology integrates AI with radio access networks, and industry analysis points to significant future potential.

Conclusion

The four AI waves each contribute uniquely, and NVIDIA’s technologies drive their evolution. Advances in data centers, Blackwell chips, next‑gen AI computers, Dynamo, robotics, autonomous driving, and 6G networks demonstrate AI’s broad impact. Future work must address ethics and privacy to ensure AI benefits society.

Key References

NVIDIA Unveils Open Physical AI Dataset to Advance Robotics and Autonomous Vehicle Development

NVIDIA Advances Robot Learning, Humanoid Development With New AI and Simulation Tools

NVIDIA Blackwell Platform Arrives to Power a New Era of Computing

NVIDIA Contributes Blackwell Platform Design to Open Hardware Ecosystem

NVIDIA Announces DGX Spark and DGX Station Personal AI Computers

NVIDIA Dynamo Open‑Source Library Accelerates and Scales AI Reasoning Models

General Motors and NVIDIA Collaborate on AI for Next‑Generation Vehicle Experience and Manufacturing

NVIDIA and Telecom Industry Leaders to Develop AI‑Native Wireless Networks for 6G

What Is Nvidia's DGX Station? A New, Specialized Desktop Line for AI Work