FHE: The Next Layer of Computing Infrastructure?

Over the past two decades, the digital economy has built massive layers of infrastructure including cloud computing, networking, and artificial intelligence. As data becomes the most valuable asset in the digital economy, protecting that data during computation becomes increasingly important. For decades computing has been built on the assumption that data must be decrypted to be useful. Databases decrypt it to process queries, cloud servers decrypt it to run workloads, and AI models train on plaintext data.The moment of decryption is also the moment when data becomes vulnerable. Breaches, insider access, and regulatory exposure all stem from this fundamental design constraint.

Fully homomorphic encryption (FHE) challenges that assumption. It allows computers to perform computations directly on encrypted data, producing encrypted outputs that only the data owner can decrypt. If this capability becomes practical at scale, it could fundamentally reshape how data is used across the digital economy.

The Breakthrough: FHE Meets Hardware Acceleration

FHE has been known in cryptography circles for years, but only recently has the technology started moving toward practical deployment. Intel’s announcement of the Heracles chip during the 2026 IEEE International Solid-State Circuits Conference represents one of the first serious attempts to solve this problem through specialized hardware. Instead of relying on traditional CPUs, which are inefficient for the mathematical operations required by FHE, Heracles is designed specifically to focus on massively parallel computing optimized for homomorphic encryption.

Early demonstrations show dramatic improvements. Intel reports that Heracles can perform key FHE operations between roughly 1,074× and 5,547× faster than a 24-core Intel Xeon processor (Xeon servers represent current real-world hardware used for FHE workloads). The chip is designed to accelerate the types of calculations that dominate encrypted computing, which traditional processors handle very inefficiently.

Another key challenge with encrypted computing is that encrypted data becomes much larger than normal data, which means far more information must be moved around during calculations. In many FHE workloads, moving encrypted data can become as expensive as performing the computation itself. To address this, Heracles includes large amounts of high-bandwidth memory so encrypted datasets can be processed much more quickly.

Heracles is not meant to replace traditional processors. Instead, it works as a specialized accelerator installed alongside a normal server, similar to how GPUs are used to accelerate artificial intelligence workloads. The main processor runs the application, while the Heracles chip handles the most demanding encrypted computations.

Importantly, Heracles is still a research prototype rather than a commercial product. Its purpose is to demonstrate that hardware acceleration can dramatically reduce the performance barriers that have historically limited homomorphic encryption. By approaching encrypted computation as a hardware problem rather than purely a software challenge, Intel is attempting to close the enormous performance gap that has prevented FHE from becoming practical. If this approach proves successful, encrypted computation could move from academic research to real-world infrastructure, allowing sensitive data to be analyzed without ever being decrypted.

Total Addressable Market

Historically, running calculations on encrypted data could be thousands or even millions of times slower than running the same calculation on normal data. Even simple tasks could take enormous amounts of computing power and memory. This made FHE impractical for most real-world applications. Companies could theoretically use it, but the cost and latency were far too high. This is why hardware acceleration matters so much. If specialized chips like Heracles can reduce the performance gap enough, encrypted computing could suddenly become economically viable. If FHE becomes practical, the range of potential applications is enormous, as many of the most valuable datasets cannot currently be shared safely. Just as GPUs unlocked modern AI by making machine learning practical at scale, specialized hardware could unlock encrypted computation.

The opportunity for FHE exists at two distinct levels that must be analytically separated.

the direct market: software, hardware, and services built for homomorphic encryption. Estimates place the current market at approximately $200–300 million, with projections reaching $500 million to $1 billion over the next decade, depending on assumptions around performance improvements. At this stage, the market reflects experimentation rather than scaled deployment.

The broader privacy-enhancing technology (PET) stack, which includes FHE alongside MPC, trusted execution environments, and related technologies, is already more substantial. This market is estimated at roughly $3–5 billion today, with forecasts suggesting expansion toward $10–30 billion over the next decade, driven by regulatory pressure, cloud adoption, and increasing demand for secure data processing.

The real thesis lies in the enabled market: the value of computation that becomes possible once sensitive data can be used without being exposed. FHE does not create a new market from scratch. It unlocks access to high-value, previously unusable data within existing markets.

Healthcare is a clear example. Global healthcare spending exceeds $10 trillion annually, yet data remains fragmented due to privacy constraints. Enabling secure, cross-institutional analysis of medical data could materially expand research productivity and analytics capabilities. Financial services represent another major opportunity. Institutions collectively manage hundreds of trillions in assets, yet cannot easily collaborate on fraud detection, risk modeling, or market intelligence without exposing sensitive customer data. FHE enables shared computation without shared data. Cloud computing is the most direct monetization pathway. With global cloud spending already exceeding $600 billion annually and projected to approach $1 trillion, even limited adoption of privacy-preserving computation, on the order of 5–10% of workloads, would imply a tens of billions of dollars annual infrastructure opportunity. Government and defense add an additional layer of demand, driven not by cost optimization but by classification and security requirements.

Healthcare and life sciences are the most likely early adopters, driven by strong regulatory constraints and relatively higher tolerance for latency. Financial services follow, particularly in collaborative analytics use cases. Cloud infrastructure represents the long-term structural prize, where encrypted computation could evolve into a standard service layer.

The current TAM should be viewed as a floor, not a ceiling. The relevant question for investors is not the size of today’s encryption market, but the size of the compute market that unlocks if FHE crosses a performance threshold. If encrypted computation becomes sufficiently efficient to support real-world workloads, it could begin to capture a share of privacy-constrained data processing across cloud, AI, healthcare, and finance. Even limited penetration, on the order of 1–5% of these markets, would translate into hundreds of billions of dollars in economic value over time.

The inflection point is not technical feasibility, but economic viability. When the cost of encrypted computation falls within an acceptable range relative to the value of the underlying data, adoption can accelerate rapidly. In that scenario, FHE should not be viewed as a niche security tool, but as a foundational layer in the next generation of computing infrastructure.

Artificial intelligence

Artificial intelligence may become one of the strongest drivers of homomorphic encryption adoption. AI systems depend heavily on access to large datasets, but many of the most valuable datasets are private and cannot easily be shared. At the same time, the newest generation of AI tools increasingly requires deep access to sensitive environments. AI assistants and autonomous agents often need permission to read documents, analyze internal databases, access emails, or interact with company codebases in order to be useful. This creates a growing trust problem: the more powerful AI becomes, the more sensitive data it must access.

Fully homomorphic encryption could enable a new model of AI development in which data remains encrypted throughout the training process. Organizations could contribute data to machine learning models without revealing the raw information. This could dramatically expand the amount of data available for AI training while preserving confidentiality.

Another important use case is privacy-preserving inference. Today, when users interact with AI systems, their data must often be sent to centralized servers for processing, where the model provider can theoretically access or store the inputs. With FHE, users could submit encrypted inputs to AI models and receive encrypted results, ensuring that the service provider cannot access the underlying data.

This capability could be especially important for enterprise AI. Many companies hesitate to adopt large language models or other AI tools because they fear exposing sensitive corporate information such as internal documents, financial data, or proprietary code. Privacy-preserving AI could remove one of the biggest barriers to enterprise adoption by allowing organizations to use powerful models without revealing their underlying data.

The Privacy Stack

Fully homomorphic encryption is often discussed as if it must compete with other privacy technologies. In reality, most of these systems are complementary rather than competitive. Technologies such as trusted execution environments (TEEs), multi-party computation (MPC), homomorphic encryption (FHE), and zero-knowledge proofs (ZKPs) all aim to enable secure computation, but they approach the problem from different angles and come with different trade-offs.

Multi-party computation, for example, allows multiple participants to jointly compute a result without revealing their individual inputs. Each participant splits their data into encrypted pieces and distributes them across several computing nodes. These nodes perform computations on the pieces and return encrypted results that the participants can reconstruct. The advantage of MPC is that it avoids a single point of failure. However, it requires large amounts of communication between nodes. As the number of participants grows, the amount of data exchanged can increase rapidly, which limits how many nodes can efficiently participate in the system.

Fully homomorphic encryption solves a different problem. Instead of distributing computation across many parties, a user can encrypt data and send it to a server that performs computations directly on the encrypted data. The server never sees the underlying information. This approach requires far less communication than MPC, but it demands far more computational power. Encrypted computations can still be significantly slower than normal computations, although hardware acceleration such as Intel’s Heracles chip may reduce this gap over time.

Zero-knowledge proofs play yet another role. rather than enabling general computation on private data, they allow proof that a computation was performed correctly. This makes them particularly useful for verification systems, such as blockchain scaling technologies, where proving correctness is more important than performing the computation itself. However, the entity generating the proof often still sees the original inputs during the computation process.

Trusted execution environments take a different approach altogether. Instead of relying purely on cryptography, TEEs rely on secure hardware environments where code can run in isolation from the rest of a system. Technologies like Intel SGX allow computations to occur inside protected enclaves where even the operating system cannot see the underlying data. TEEs are generally much faster than cryptographic approaches, but they rely on trust in hardware security and have historically faced vulnerabilities such as side-channel attacks.

Because these technologies solve different problems, they can often be combined to build stronger systems. For example, TEEs can help manage cryptographic keys or handle certain performance-intensive tasks while homomorphic encryption protects sensitive data during computation. MPC can distribute trust across multiple machines instead of relying on a single secure environment. Zero-knowledge proofs can then verify that the final result was produced correctly without revealing the underlying data.

Rather than competing architectures, these technologies are increasingly forming what could be described as a privacy-computing stack. Each layer addresses a different part of the security challenge: protecting data during computation, enabling collaboration between parties, verifying results, or isolating sensitive workloads.

As data becomes more valuable and AI systems demand access to increasingly sensitive information, combining these technologies may become the most practical path toward secure computing. In the long run, the most robust systems will likely integrate several of these approaches rather than relying on a single method alone.

Intel’s Strategic Pivot: Privacy Computing After Losing the AI Race

The emergence of hardware acceleration for encrypted computation may also carry strategic implications for the semiconductor industry. Over the past decade, Intel lost its dominant position in the most important new data center workload: artificial intelligence. While Intel remained dominant in traditional server CPUs, NVIDIA captured the most important new data-center workload: artificial intelligence.

Homomorphic encryption represents an entirely different class of computing workload. Instead of matrix computations used in machine learning, encrypted computing relies on large-scale calculations and heavy data movement. These workloads require architectural optimizations that differ significantly from those used in AI accelerators.

Intel’s work on homomorphic encryption hardware may therefore represent an early attempt to position itself in a new category of computing infrastructure. The key strategic question is whether encrypted computation will remain a niche security capability or evolves into a fundamental layer of cloud infrastructure. Companies such as Amazon, Microsoft, and Google are already investing heavily in custom silicon for AI. If encrypted computation becomes more widely adopted, particularly for enterprise AI workloads, these companies may also begin developing hardware optimized for privacy-preserving computation.

If the latter occurs, the race to build hardware for secure computation could become the next major battleground in data-center infrastructure.

Measuring Progress: When Does FHE Become Real?

The first dimension to monitor is performance improvement. As mentioned earlier, historically, encrypted computations have been orders of magnitude slower than normal computation. Progress can be measured by how quickly this gap closes. If FHE workloads move from being thousands of times slower to within a range where latency and cost become manageable for specific applications, the technology transitions from research to infrastructure.

The second dimension is real-world deployment. Early adoption will not appear in academic papers, but in production environments. Signals include enterprises running encrypted analytics, financial institutions collaborating on private datasets, or healthcare organizations using FHE for sensitive data processing. The shift from experimentation to production use is one of the clearest indicators that the technology is maturing.

The third dimension is integration into existing infrastructure, particularly cloud platforms. If major cloud providers begin offering encrypted computation as a service, it suggests that FHE has reached a level of maturity where it can be deployed at scale. This would mark a transition from niche applications to a broader infrastructure layer.

The fourth dimension is integration with artificial intelligence workflows. As AI systems increasingly require access to sensitive data, the demand for privacy-preserving computation may grow. The emergence of encrypted model training or encrypted inference would indicate that FHE is becoming relevant to one of the largest and fastest-growing compute markets.

Finally, regulatory pressure may act as an accelerant. Stricter data privacy requirements could force organizations to adopt technologies that allow computation without exposing underlying data. In such a scenario, adoption would not be driven purely by performance improvements, but by necessity.

Taken together, these dimensions provide a practical framework for evaluating progress. The inflection point for FHE will not be a single breakthrough, but the moment when performance, deployment, and infrastructure integration converge. At that point, encrypted computation moves from theoretical possibility to economically viable technology.

Public Companies Positioned for FHE

While several companies are exploring privacy-preserving computation, Intel currently represents the only meaningful public market entry point into fully homomorphic encryption.

Through its work on hardware acceleration, Intel is directly targeting the core bottleneck that has historically limited FHE adoption. However, this exposure should be viewed as optionality rather than a core investment thesis. FHE remains an early-stage technology and contributes little to Intel’s current business.

FHE should be viewed as a long-duration optionality within Intel’s broader strategy. It is not part of the current investment case, but rather a potential future driver if encrypted computation becomes a meaningful data center workload. The key for investors is not to underwrite Intel on FHE today, but to monitor the progression of the technology toward economic viability. If performance improves to the point where real-world applications become cost-effective, the next step is to assess where adoption emerges and how large the addressable market could become.

At this stage, buying Intel purely as an FHE play is premature. Instead, it should be seen as a broader semiconductor turnaround story with a potential embedded option on a new computing paradigm, one that may only become relevant if and when encrypted computation moves from research to commercially viable infrastructure.