Lux Capital is a science and technology investor whose AI portfolio connects software with infrastructure and robotics. Its disclosed investments span coding agents, data platforms, model infrastructure and cloud execution, covering tools used to build AI systems as well as applications of those systems.
Databricks, Hugging Face, Modal and Together AI occupy different positions around data, models and execution. Their inclusion supports an infrastructure view of AI adoption, where useful products depend on more than a model’s headline capability.
Physical Intelligence adds a robotics dimension to the selected sample. Embodied AI introduces requirements around real-world feedback, hardware and safe operation that software-only comparisons can miss.
A company serving model developers faces different constraints from one building a coding agent or robot model. Understanding the bottleneck each company addresses makes the portfolio more informative than treating every position as the same AI bet.
Selected disclosed relationships, reviewed November 4, 2026. Each source supports the investment relationship; this is not a complete portfolio or a statement of current holdings.
Disclosed in Lux’s official company portfolio.
Source
A data-platform investment in the official portfolio.
Source
Listed in Lux’s portfolio as Hugging Face.
Source
A cloud-compute company included in the official portfolio.
Source
The official portfolio describes its work on AI for robotics.
Source
Listed as Together.xyz in Lux’s official portfolio; resolved to Together AI’s company identity.
SourceAI systems depend on data platforms, model tooling and execution infrastructure as well as applications. Databricks, Hugging Face, Modal and Together AI provide different examples of those enabling layers. Their inclusion broadens the AI view without implying that they have identical products, revenue models or exposure to every part of the model market.
They occupy related parts of the AI ecosystem but should be assessed separately. Model discovery, collaboration, hosting and inference can involve different users and workflows. A shared infrastructure label is a starting point for comparison, not evidence that the companies provide interchangeable services or face the same economics.
Cloud execution concerns running workloads reliably with suitable resources, scheduling and developer interfaces. Training a foundation model concerns producing model capabilities from data and computation. The activities can interact, but infrastructure quality and model quality require different evaluation methods; success in one does not automatically establish success in the other.
Physical systems must respond to environmental variation, hardware limits and the consequences of actions in the real world. Data collection and evaluation can require physical interaction rather than only offline tests. Physical Intelligence makes this distinction relevant to the selected portfolio, without allowing software benchmark results to stand in for robotics deployment evidence.
Executing code and managing an organization’s data workflows are different tasks, even when both support AI applications. A team needs to understand where its workload runs, how data is governed and which operational responsibilities each service handles. Grouping both under infrastructure should not obscure those separate purchasing and integration decisions.
Data platforms, model tooling, inference services and cloud execution address different customer problems. Their costs, switching barriers and purchasing decisions can differ substantially. Comparing the specific workflow and bottleneck is more useful than grouping all revenue under AI infrastructure and assuming that growth or margins will follow the same pattern.