Why The Sullivans are Investing Heavily in Privacy-Preserving LLMs

For most of the AI boom, scale has been the dominant narrative. Larger datasets, larger models, and larger compute clusters defined competitive advantage. But according to the Sullivans, the next phase of artificial intelligence will be shaped by something far less visible: privacy.

Their recent investments into privacy-preserving large language models (LLMs) signal a strategic bet that the future of AI adoption depends less on capability and more on confidence — specifically, whether organizations can deploy AI without exposing sensitive data.

The concern is not hypothetical. Modern LLMs are trained on vast datasets that may include proprietary information, personal communications, or regulated records. Researchers have demonstrated that models can unintentionally leak training data through techniques such as model inversion or membership inference attacks, raising serious security questions for enterprise adoption.

For industries like healthcare, finance, and government, this risk creates a paradox: AI offers enormous productivity gains, yet the data required to power it cannot safely leave secure environments. Privacy-preserving AI attempts to resolve this contradiction.

Privacy as Infrastructure, Not Feature

The Sullivan thesis treats privacy not as an add-on but as foundational architecture.

Emerging techniques such as federated learning allow models to train across decentralized devices without centralizing raw data. Instead of uploading sensitive information to a server, systems learn locally and share only encrypted updates — enabling collaboration while maintaining confidentiality.

Major technology companies already rely on similar methods. Smartphone keyboards and voice assistants improve collectively while keeping user inputs on-device, demonstrating that large-scale AI can evolve without direct access to private data.

For investors, this signals a structural shift: AI development is moving from data aggregation toward data minimization.

The Enterprise Adoption Bottleneck

Despite rapid advances in generative AI, enterprise deployment remains cautious. Companies fear exposing intellectual property to external models or cloud providers. Privacy-preserving LLMs address this hesitation by enabling secure training, encrypted inference, and controlled collaboration across organizations.

New research frameworks combine encryption, secure multi-party computation, and differential privacy to reduce leakage risks while preserving model performance.

This matters economically. If AI systems can operate directly inside private environments — hospitals, banks, or internal corporate networks — adoption expands dramatically. The Sullivans see this as the equivalent of the cloud transition moment, when security assurances unlocked enterprise-scale migration.

Regulation Is Quietly Driving the Market

Another force behind the investment surge is regulation. Global privacy laws increasingly restrict how data can be collected, transferred, and processed. Privacy-preserving machine learning allows organizations to comply with these frameworks while still benefiting from advanced analytics.

Startups focused on federated learning platforms have already emerged to meet demand from companies struggling to implement privacy-safe AI internally, highlighting a growing commercial opportunity.

Rather than slowing AI innovation, regulation may be accelerating a new category of infrastructure — trusted AI systems designed for compliance by default.

The Post-Scale Era of AI

The Sullivans’ broader view is that AI competition is entering a post-scale phase. Performance improvements from larger models are becoming incremental and expensive, while trust, security, and governance remain unresolved challenges.

Privacy-preserving LLMs shift value toward architecture rather than raw compute. Techniques such as encrypted training, decentralized learning, and synthetic data generation allow models to learn from sensitive environments previously inaccessible to AI developers.

Some experimental projects already train models across distributed networks rather than centralized data centers, hinting at a more decentralized AI ecosystem where users retain ownership of their data.

Betting on Trust as the Next Moat

Historically, technological revolutions reward those who remove friction. In early internet years, speed mattered most. In the cloud era, scalability dominated. In AI’s next chapter, the Sullivans believe trust will become the primary competitive advantage.

Privacy-preserving LLMs promise an environment where organizations can deploy intelligent systems without surrendering control over their most valuable asset — information itself.

If that vision proves correct, the biggest AI companies of the next decade may not be those with the largest models, but those capable of proving something far harder:

That intelligence can scale without sacrificing privacy.