What problem are we solving?

Unlocking Underutilized and Distributed Private Data for AI Training

In Silicon Valley, a core belief has guided AI development: “throw more data and more compute” and models will continue to improve. However, this strategy faces an impending limit.

Currently, AI models are trained on three major categories of data:

• Internet Data (publicly available information)

• Proprietary Data (privately owned datasets)

• Synthetic Data (artificially generated examples)

Recent research suggests that the supply of high-quality human-generated data could be exhausted as early as 2027 (source). As models scale, simply relying on existing data sources will no longer sustain progress.

Core Hypothesis:

The most valuable, high-quality data for the next wave of AI training is currently locked away:

• On edge devices like smartphones, laptops, and IoT devices

• Inside institutional silos (e.g., healthcare, finance, energy companies)

For example, clinical data in healthcare institutions contains deep, rich insights that could significantly advance AI capabilities. Yet today, this data remains untapped due to privacy concerns, regulatory barriers, and infrastructure limitations.

Unlocking this distributed private data — safely, ethically, and efficiently — could be key to fueling the next era of AI innovation.