For years, the biggest challenge in privacy-preserving computing was performance.
Fully Homomorphic Encryption (FHE) promised something extraordinary: the ability to compute on encrypted data without ever decrypting it. The concept was powerful, but for much of its history, it remained largely confined to research labs because the computational overhead was simply too high for most real-world applications.
That is beginning to change.
Recent advances in cryptography, open-source software, hardware acceleration, and cloud infrastructure have transformed Privacy-Enhancing Technologies (PETs) from an academic pursuit into an emerging enterprise capability. A recent OpenFHE webinar examining DARPA’s DEPRIVE program highlighted just how much progress has been made over the past decade, from early government-funded research initiatives to sophisticated hardware architectures capable of dramatically accelerating encrypted computation.
But the most important takeaway isn’t about hardware.
It’s about what happens next.
The conversation around AI is shifting from model performance to data access. Organizations increasingly understand that their biggest competitive advantage isn’t necessarily a larger model or a new algorithm. It’s access to data they couldn’t use before.
That reality is elevating Privacy-Enhancing Technologies from a specialized security tool to a foundational layer of secure AI infrastructure.
DARPA‘s investments helped turn encrypted computation into a practical engineering challenge rather than a theoretical one. Programs such as PROCEED and DEPRIVE brought together researchers, software developers, hardware designers, and cryptographers to address one of the most persistent obstacles in FHE: performance. These efforts contributed to advances in schemes, libraries, hardware acceleration approaches, and eventually open-source initiatives such as OpenFHE that are now widely used throughout the industry.
The results are significant. What once required impractical amounts of computing power can now be accelerated through specialized architectures, optimized software libraries, GPUs, and dedicated hardware designs. The ecosystem surrounding privacy-preserving computation has matured considerably, attracting startups, enterprise investment, government interest, and growing commercial adoption.
Yet the maturation of FHE and related technologies has revealed a deeper challenge.
Organizations rarely suffer from a lack of data. More often, they struggle to access the data they need.
The most valuable information frequently sits outside organizational boundaries. It may belong to a partner, a customer, a government agency, a healthcare institution, or a coalition member. It may be protected by privacy regulations, data sovereignty requirements, security classifications, or intellectual property concerns. In many cases, the data cannot be moved, centralized, or shared without introducing unacceptable risk.
As AI initiatives become increasingly dependent on larger and more diverse datasets, these constraints become more problematic.
Healthcare researchers want to collaborate across institutions without exposing patient records. Financial institutions want to detect fraud across organizational boundaries without sharing customer data. Governments want to analyze information across domains while maintaining strict security controls. Enterprises want to evaluate third-party AI models without exposing sensitive data or prompts.
Traditional approaches were never designed for these challenges.
Historically, organizations solved data access problems by copying information into centralized environments. Security was enforced through perimeter controls, permissions, and governance processes. But as data volumes grow and regulatory requirements become more stringent, centralization is increasingly becoming a barrier rather than an enabler.
This is where Privacy-Enhancing Technologies are creating a new path forward.
Importantly, the future will not be built on a single PET.
One of the clearest lessons emerging from both research programs and enterprise deployments is that different technologies solve different problems. Fully Homomorphic Encryption enables secure queries against sensitive datasets while protecting the requester’s query and results. Federated learning allows organizations to train models across decentralized datasets without moving underlying data. Trusted Execution Environments provide secure processing environments for workloads that require access to raw data inside protected enclaves. Differential Privacy helps protect individual privacy when sharing analytical outputs.
Together, these technologies create a framework for secure collaboration that was previously impossible.
This is particularly important for AI.
The next generation of AI systems will not operate inside a single organization or a single cloud. They will span jurisdictions, institutions, partners, and data owners. They will require access to information that cannot be shared through conventional means. Success will depend not only on model quality but on the ability to securely use data wherever it resides.
The organizations that solve this challenge will gain access to insights their competitors cannot reach.
That is why PETs are increasingly becoming part of enterprise AI strategy discussions. They are no longer viewed solely as privacy technologies. They are becoming business-enablement technologies that unlock access to previously inaccessible information while preserving security, compliance, and governance.
The significance of DARPA’s early investments, and the progress demonstrated by communities like OpenFHE, is not simply that encrypted computation became faster. It is that they helped establish the technological foundation for a future in which organizations can securely collaborate, analyze, and build AI on data that cannot be moved or exposed.
The next era of AI will not be defined solely by bigger models.
It will be defined by trust.
And increasingly, Privacy-Enhancing Technologies are becoming the foundation on which that trust is built.
