Fair Math, a Polymorphic Capital portfolio company, has published a joint research paper with NVIDIA and Duality Technologies: "Efficient Large-Integer Arithmetic for FHE." The paper is deeply technical. What it signals is a business story, and this is what we want to share with you today.
"Privacy is quickly becoming a critical part of the modern AI stack - especially as more sensitive data moves into AI inference and computation," says Eldar Khamitov, Partner at Polymorphic Capital. "Excited to see our portfolio company Fair Math collaborating with NVIDIA on the fundamental challenges behind privacy-preserving computation."
The short version: the most valuable data in the world, in healthcare, finance, and enterprise systems, largely cannot be used by AI today, because using it means handing it to someone else's infrastructure in readable form. Fully homomorphic encryption (FHE) removes that requirement. It lets a server compute on data while the data stays encrypted the entire time. The catch has always been speed. This paper, written by three organizations with very different vantage points, maps out exactly where the speed problem lives and how it is being solved.
We asked Gurgen Arakelov, founder and CEO of Fair Math, why privacy remains one of the major open challenges for AI.
"Privacy is becoming a fundamental constraint on where and how AI can be deployed. Many of the most valuable AI use cases involve highly sensitive data - healthcare, finance, enterprise data, personal information - and in many of these areas, sending raw data to a third party for inference is simply not acceptable.
This means privacy is not just a security feature. It directly determines which data AI can access and, ultimately, which problems AI can solve. If we want AI to become part of critical infrastructure and operate on the world's most sensitive data, we need a way to compute without exposing that data."
Read that as an investor and the framing changes. Privacy technology is usually priced as a cost center, a compliance line item. Gurgen is describing it as the thing that decides how large the market for AI services can get. Every hospital, bank, and government dataset that cannot legally or commercially leave its owner is revenue that current AI infrastructure cannot touch.
Plenty of technologies already claim to protect data during computation, so we asked founder of Fair Math why he considers the problem unsolved.
"Because every practical solution available today requires some form of compromise or trust assumption.
Trusted Execution Environments are a good example. They can provide excellent performance, but their security ultimately depends on hardware, firmware, vendors, and a relatively large trusted computing base. We have also seen multiple classes of attacks against TEEs over the years. So while TEEs are extremely useful, the security model is fundamentally different from cryptographic protection.
Other approaches introduce different trade-offs - restrictions on computation, interaction between multiple parties, changes to the threat model, or assumptions about who can be trusted.
For many applications these compromises are perfectly reasonable. The open problem is how to achieve strong cryptographic privacy with minimal trust assumptions while preserving the flexibility and performance developers expect from ordinary computation."
FHE is the one approach where the guarantee comes from mathematics instead of from a vendor's hardware. Data leaves the user encrypted, gets processed encrypted, and comes back encrypted. There is nothing for the infrastructure operator, or an attacker inside it, to see.
Encrypting data this way forces computers to do arithmetic on numbers thousands of bits long, while processors are built for 64-bit numbers. Closing that gap is the single biggest cost in every FHE system, and it is the subject of the paper. The authors, including the researchers behind the algorithms used in today's leading FHE libraries, trace 15 years of progress on this problem and show how far GPU acceleration has already pushed it. One result cited in the paper: encrypted inference on Llama3-8B, an 8-billion-parameter language model, in 134 seconds on a cluster of NVIDIA GPUs. Operations that were once the technology's showstopper now run in milliseconds.
We asked Gurgen where FHE fits and what the collaboration contributes.
"FHE is unique because it allows computation to happen directly on encrypted data. The data can remain encrypted from the moment it leaves the user until the result comes back. You don't have to trust the infrastructure performing the computation with the plaintext.
The challenge is that FHE still faces two major barriers to broad adoption: performance and programmability.
Our work with NVIDIA and Duality Technologies is focused on the first one. We are working together on some of the fundamental computational challenges behind FHE and exploring how advances in algorithms, implementations, and hardware acceleration can make encrypted computation significantly more efficient and practical at scale.
At Fair Math, we have also been working on the second barrier for quite some time: making encrypted computation substantially easier to program, compile, and deploy. We've made important progress there.
Ultimately, these two directions have to converge. FHE needs to become both fast enough and easy enough to use that developers don't have to be cryptographers to build privacy-preserving applications. That's when privacy-preserving AI can move from a specialized cryptographic capability to a native part of AI infrastructure."
Polymorphic Capital invested in Fair Math's pre-seed round in February 2024. Our thesis then was that FHE would follow the path of other infrastructure technologies: dismissed as impractical until performance crosses a threshold, then adopted quickly because the demand was there all along. AI is now supplying that demand.
What this paper confirms for us is where Fair Math sits in that shift. A pre-seed company co-authoring foundational research with NVIDIA and with the team behind the open-source OpenFHE library is operating at the layer where the performance roadmap of encrypted computation is actually decided.
For a broader look at what this research means for private AI and its commercial adoption, Inside Deep Tech breaks down the paper from an AI infrastructure and business perspective.
The full paper, "Efficient Large-Integer Arithmetic for FHE," by researchers from Fair Math, NVIDIA, and Duality Technologies, is available on the IACR ePrint archive.
Fair Math is a research-driven company building cryptographic infrastructure for privacy in AI and decentralized systems, based on fully homomorphic encryption. The company develops FHERMA, a challenge platform that crowdsources an open-source library of FHE components so that application developers can build encrypted computation without being cryptographers. Fair Math was founded by PhD holder in mathematics and computer science.
Polymorphic Capital is a Web3-focused venture fund investing primarily in application-layer companies, with an emphasis on businesses using blockchain infrastructure to solve real economic problems rather than speculation or liquid-token trading. Fund has a strong exposure to stablecoins, payments, fintech infrastructure, RWAs/tokenization, agentic commerce, identity/data and privacy.
FHE is a cryptographic technique that allows a computer to run calculations directly on encrypted data. The data is never decrypted during processing, so the party running the computation never sees it.
The most valuable AI use cases involve sensitive data in healthcare, finance, and enterprise systems. FHE lets AI models process that data while it stays encrypted end to end, removing the need to trust the infrastructure provider with readable information.
A research paper, "Efficient Large-Integer Arithmetic for FHE," covering the core computational bottleneck of FHE and the algorithmic and GPU-acceleration techniques that address it.
Fair Math was founded by Gurgen Arakelov, who holds a PhD in mathematics and computer science. Polymorphic Capital invested in the company's round in February 2024.
