Jensen Huang on Why Unstructured Data Is the Next Foundational Platform

A clip of NVIDIA CEO Jensen Huang explaining why unstructured data -- PDFs, videos, speeches, and everything else that does not fit into rows and columns -- is about to become useful at scale. About 90% of the data generated each year is unstructured, and until now, most of it has been effectively dead weight. You store it, maybe read it, but you cannot query it or search it in any systematic way.

The Problem with Unstructured Data

Structured data has always been easy to work with. You put it in a database, you index it, you query it. Unstructured data has no natural indexing scheme. A PDF does not have a primary key. A video does not have columns. To make unstructured data searchable, you have to understand its meaning and purpose, and that was not computationally feasible at scale until recently.

AI Solves the Indexing Problem

Huang's argument is that the same multimodal perception and understanding capabilities that let AI interpret images, text, and audio can be applied to the indexing problem. AI can read a PDF, understand what it means, and embed that meaning into a searchable structure. That turns the 90% of the world's data that was previously inert into something you can query.

NVIDIA's Two New Foundational Libraries

NVIDIA created two libraries that Huang positions as foundational for this shift:

cuDF -- for structured data frames. Think of it as the GPU-accelerated layer for traditional tabular data processing.

cuVS -- for vector stores and semantic data. This is the library for unstructured and AI-native data, handling the embedding and retrieval operations that make semantic search work at scale.

Huang draws a direct parallel to RTX for 3D graphics: just as RTX became foundational infrastructure for rendering, cuDF and cuVS are intended to become foundational infrastructure for the world's data processing systems.

Key Takeaways

• About 90% of data generated each year is unstructured, and until now it has been largely unsearchable
• AI multimodal perception solves the indexing problem by understanding the meaning of unstructured content
• NVIDIA's cuDF (structured data) and cuVS (vector/semantic data) are positioned as foundational libraries for this new data processing layer
• The shift from storing unstructured data to querying it at scale represents a significant platform opportunity