Data Infrastructure, Not Models, Is the AI Bottleneck

NVIDIA has expanded its Cosmos platform with new tools for training, evaluating, and deploying world models, including Cosmos Data Factory blueprints for data curation and synthetic data generation. The platform ranks in the top five on six open model leaderboards and now supports fine-tuning and optimization of compatible models like GR00T, Alpamayo, and QWEN. Early adopters including Porsche, ANYbotics, and Ultralytics are using the tools to reduce physical testing and accelerate model development.

A new market has emerged in which AI labs and data companies are purchasing datasets from defunct startups through intermediaries like SimpleClosure and Sunset. The practice, which became common only in the past year as AI training labs exhausted public datasets, has grown to include major deals such as Google's $10 million acquisition of Spirit Airlines' data during bankruptcy. The trend raises legal and privacy concerns, as the data often contains personally identifiable information embedded in emails, calendars, and other workplace communications.

Google is joining Apache Ossie, an industry group that includes Snowflake and Nvidia, to standardize how AI tools access data from applications and databases. The move aims to reduce friction in data integration for enterprise AI deployments. Apache Ossie is working to create common standards that make it easier for AI systems to understand and work with data across different sources.

Proximal, a San Francisco-based data curation startup, emerged from stealth with $15 million in seed funding led by General Catalyst at a $300 million valuation. The 40-person company, founded last fall by Calvin Chen and Justus Mattern, has reached $200 million in annualized revenue in just 10 months, capitalizing on AI developers' demand for high-quality training data in software engineering.