Nestpoint Group builds and invests at the intersection of energy resilience, technology infrastructure, and capital markets. These are sectors where complexity demands conviction, and where the next wave of value creation will be driven by teams that understand both the underlying technology and the commercial frameworks required to scale it.
Last week, the team from Spatious Data, a Nestpoint portfolio company, completed a whirlwind series of meetings across Silicon Valley with leading players in artificial intelligence and robotics. The conversations spanned foundation model developers, autonomous systems companies, and research labs pushing the boundaries of embodied intelligence. One theme dominated every room: the ferocious appetite for high-quality spatiotemporal data to train the next generation of AI models. And one problem followed close behind: a staggering percentage of the available data simply is not good enough.
As we outlined in our recent analysis of the global robotics installation race, the country that deploys at the greatest scale accumulates the training data, manufacturing capacity, and cost advantages that compound over decades. China's two-million-unit operational robot stock is already generating spatiotemporal data at a volume that no other country can match. The question of who curates, validates, and makes that data accessible to American AI developers is not a secondary concern. It is central to competitiveness.
That gap between demand and quality represents both a critical bottleneck for the industry and a defining opportunity for Spatious Data.
A Primer on Spatiotemporal Data for Robotics and AI
To understand why the data problem matters, it helps to understand what these models actually consume. The term "spatiotemporal data" refers to information that captures both spatial relationships and how those relationships change over time. For robotics and embodied AI, this data is the raw material that teaches machines to perceive, reason, and act in the physical world.
The scale of data generation is directly tied to the scale of robot deployment. According to the International Federation of Robotics, 542,000 industrial robots were installed globally in 2024, more than double the figure from a decade ago. China alone accounted for 295,000 of those installations, representing 54% of global deployments. The United States installed 34,200 units, outpaced roughly nine to one. Every one of those machines generates spatiotemporal data during operation, and the gap in deployment scale translates directly into a gap in available training data.¹
The landscape of relevant data types is broad and growing more complex by the quarter.
Vision-Language-Action Models
Vision-Language-Action (VLA) models represent one of the most promising frontiers in AI research. These architectures combine visual perception, natural language understanding, and physical action planning into a single framework. Training them requires massive volumes of paired data: video streams annotated with language descriptions and corresponding physical actions. The quality bar for this data is extraordinarily high. Noisy labels, inconsistent action mappings, or poorly calibrated visual inputs can degrade model performance in ways that are difficult to diagnose after the fact.
Robot-Specific Sensorimotor Data
Robots generate two fundamental categories of sensory data during operation. Proprioceptive data captures the internal state of the machine: joint angles, torques, velocities, and forces measured by onboard sensors. This is how a robot understands its own body. Exteroceptive data captures the external environment: depth maps from LiDAR, point clouds, camera feeds, and tactile sensor readings from contact with objects and surfaces. Together, these streams form the sensorimotor foundation that allows a robot to interact with the world. The challenge is that this data is generated at high frequency, varies significantly across hardware platforms, and degrades rapidly when collection protocols are inconsistent.
Multi-Modal and Ego-Centric Data
Modern AI systems increasingly rely on multi-modal data, meaning information fused from multiple sensor types into a coherent representation. Ego-centric data is a particularly valuable subset. This data is captured from a first-person human perspective, typically through head-mounted or body-worn cameras that record how people navigate environments, manipulate objects, and perform everyday tasks. The resulting footage provides a naturalistic view of human interaction with the physical world, offering the kind of grounded, first-person context that third-person observation datasets cannot replicate. For robotics, ego-centric data is a powerful training signal because it demonstrates how intelligent agents actually solve physical problems. Collecting it at scale, with consistent quality and meaningful annotations, remains one of the hardest problems in the field.
Synthetic Data and Its Limitations
The industry has invested heavily in synthetic data generation as a way to supplement scarce real-world datasets. Physics-based simulations can produce controlled training scenarios. Generative world models can fabricate novel environments. Data augmentation techniques can expand existing datasets through transformation and variation. These approaches have genuine value, but they also have well-documented limitations. Simulated data inevitably reflects the assumptions and simplifications of the simulator. Models trained primarily on synthetic data often struggle to transfer their capabilities to real-world conditions, a problem the research community refers to as the "sim-to-real gap." The consensus from our Silicon Valley conversations was clear: synthetic data is a supplement, not a substitute, for high-quality real-world data.
How Spatious Data Addresses the Gap
The demand for this data is reflected in the market. The global AI training dataset market was valued at approximately $3.6 billion in 2025 and is projected to reach $23.2 billion by 2034, a compound annual growth rate of nearly 23%.² That growth is being driven by organizations across healthcare, automotive, retail, and industrial automation that require high-quality, diverse datasets to improve model accuracy and real-world performance. The spatiotemporal segment, data with both spatial and temporal dimensions, is among the fastest-growing categories within that market.
This is where Spatious Data enters the picture.
Spatious Data Inc. aggregates spatiotemporal data generated by robotic systems during real-world operation. The company applies proprietary algorithms for event tagging, labeling, and success/failure classification, transforming raw operational data into structured, validated datasets that are ready for model training and evaluation.
The value proposition is straightforward but difficult to replicate. Spatious Data serves as a trusted data intermediary between international robotics data sources and U.S.-based AI and robotics customers. Robots operating across global markets generate enormous volumes of spatiotemporal data every day. Much of this data is underutilized because it lacks the curation, validation, and compliance frameworks necessary to make it useful for American AI developers operating under increasingly rigorous data governance standards.
Spatious Data solves this through its Sovereign Shield contract frameworks, which ensure compliance and data integrity for internationally sourced data. This layer of trust and governance allows customers to access diverse, high-quality datasets from global sources without taking on the legal, regulatory, or quality risks that would otherwise make such data unusable.
In an industry where the difference between a model that works in the lab and one that works in the field often comes down to the quality and diversity of its training data, the role of a trusted curator and validator is not a convenience. It is a necessity.
Looking Ahead
The robotics and AI industry is moving fast, but by any honest measure, we are still early. Foundation models for embodied intelligence are advancing rapidly, yet the data infrastructure required to support them at scale remains underdeveloped. The companies that build reliable pipelines for high-quality, compliant, and diverse spatiotemporal data will hold a position of structural importance in the value chain.
That is the opportunity Spatious Data was built to capture, and the conviction behind Nestpoint's investment. We look forward to sharing more as this space continues to evolve.
Sources
¹ International Federation of Robotics, World Robotics 2025 Report (September 2025). Data visualized by Our World in Data.
² Fortune Business Insights, AI Training Dataset Market Size, Share & Industry Analysis (2025).
