Fifteen years ago, Nexdata was founded with a clear belief: every step forward in AI begins with better data.

At that time, AI models were learning to recognize speech, identify images, and understand text. Training data was becoming one of the most important foundations for AI development.

Today, AI is moving into a different stage.

Large language models are evolving into multimodal systems. Intelligent agents are moving beyond digital interfaces and into the physical world. In Physical AI scenarios, models need to perceive environments, understand spatial relationships, interact with objects, reason over long-horizon tasks, and execute physical actions.

For Nexdata, our fifteenth anniversary is more than a company milestone. It also marks a new stage in our development: from providing AI training data to building data infrastructure for the Physical AI era.

Fifteen Years of Building AI Data Foundations

Every new generation of AI brings new requirements for data.

Over the past fifteen years, the industry has moved from speech recognition and computer vision to natural language processing, autonomous driving, multimodal foundation models, and now Physical AI.

Through these changes, Nexdata has remained focused on one thing: providing high-quality data that supports AI model training, validation, and improvement.

Today, Nexdata provides more than 1,500 licensed datasets, including over 10 million hours of speech data across 200+ languages, as well as more than 800 TB of computer vision datasets for AI development across different industries.

These numbers represent more than scale.

They represent fifteen years of accumulated experience in data collection, preprocessing, annotation, quality management, and large-scale data production. They also reflect Nexdata’s ability to support different model types, application scenarios, and project requirements.

Over the years, Nexdata has not only witnessed the growth of the AI data industry. We have also participated in its development by continuously improving data quality standards, engineering workflows, and delivery capabilities.

Following Where AI Needs Data Next

Nexdata was founded to solve one of the most common challenges in machine learning: data preparation.

“Our founding team came from machine learning backgrounds,” said one of Nexdata’s founders. “In our own projects, we found that a large amount of time was spent preparing data instead of building models.”

That experience shaped the company’s early direction.

AI models need high-quality data, but collecting, preprocessing, labeling, and structuring data remains complex and time-consuming. Nexdata was created to make AI data collection and annotation more professional, scalable, and reliable, so AI teams could spend more time on model development and innovation.

Fifteen years later, that belief has not changed.

What has changed is the type of data AI now requires.

“AI learns from data much like people learn from experience,” the founder said. “The quality, diversity, and consistency of data directly influence what models are able to learn.”

Over the past fifteen years, model architectures, application scenarios, and industry priorities have changed significantly. But one question has always guided Nexdata:

What kind of data will the next generation of AI need?

Today, the answer is increasingly clear: Physical AI.

Unlike traditional AI, which mainly learned from digital information, Physical AI needs to understand the physical world. Models must perceive environments, recognize spatial relationships, interact with people and objects, perform dexterous manipulation, reason over long-horizon tasks, and complete physical actions.

This requires a different kind of data.

Instead of isolated images, short video clips, or single-step annotations, Physical AI needs multimodal data that captures continuous actions, human-object interactions, temporal sequences, environmental changes, spatial context, and complete task execution processes.

“For a company focused on AI data, supporting Physical AI is a natural continuation of what we have been doing for the past fifteen years,” the founder said. “As AI moves into the physical world, data infrastructure must evolve with it.”

From Customer Requirements to Physical AI Data Solutions

The development of AI technologies has also changed how customers define their data needs.

In earlier stages, many projects focused on speech, image, text, and standard annotation tasks. With the growth of autonomous driving, multimodal AI, and VLA/VLM systems, customers began to require multi-sensor synchronization, complex scene understanding, and task-oriented datasets.

Now, Physical AI is bringing another shift.

Customers no longer ask only whether data is available. They want to know whether the data reflects target tasks, whether the collection environment supports model development, whether the workflow is stable and consistent, and whether the data can support long-horizon reasoning and embodied task execution.

To support these more complex requirements, Nexdata uses an AI-powered data platform that connects data cleaning, filtering, annotation, quality control, and delivery within one production workflow. AI-assisted pre-labeling can improve production efficiency by about 30%, supporting standardized large-scale data production for Physical AI, autonomous driving, multimodal AI, and other next-generation AI projects.

This reflects a broader shift in the industry.

Future AI datasets will not simply become larger. They will become more scenario-oriented, task-oriented, and process-oriented, helping models move from perception to reasoning, planning, and physical execution.

From AI Training Data to Physical AI Data Infrastructure

As AI moves into the physical world, data requirements are changing as well.

Models are no longer learning only from speech, images, videos, and text. They are learning from continuous interaction with environments, objects, and tasks.

For Nexdata, this is not a sudden change in direction.

It is the natural extension of fifteen years of experience in AI data.

Today, Nexdata operates an 8,000+ square-meter Physical AI data collection facility, equipped with 300+ humanoid robots, 200+ robotic arms and dexterous hands, and simulated environments covering home, factory, retail, and office scenarios.

Together with capabilities in task design, scenario construction, multimodal data collection, annotation, and quality management, this enables Nexdata to provide end-to-end data production services for next-generation Physical AI, robotics, and intelligent systems.

This investment reflects Nexdata’s belief that Physical AI will require a new generation of data infrastructure.

And Nexdata is committed to helping build it.

Fifteen Years Forward

Over the past fifteen years, AI has moved from perception to understanding, from single-modality to multimodal intelligence, and from the digital world toward physical interaction.

Throughout these changes, one thing has remained constant:

High-quality data is still the foundation behind major progress in AI.

Looking ahead, AI will continue moving beyond screens and into real-world interaction. As models become more capable of perception, reasoning, manipulation, and autonomous task execution, the need for richer, more structured, and more scalable data will continue to grow.

Fifteen years ago, we believed better data could accelerate AI innovation.

Fifteen years later, we believe this even more strongly.

From AI training data to Physical AI data infrastructure, Nexdata will continue building the data foundations for the next generation of intelligent systems.

The next chapter has already begun. Nexdata is building the data infrastructure that will power it.