The challenge of keeping data secure has shifted from storage to the moment it's actually being used, especially as artificial intelligence (AI) systems become more sophisticated at uncovering security weaknesses. While encrypting data on hard drives has been standard practice for years, protecting sensitive information as it flows through networks and gets processed by AI models remains a critical gap. Government regulations in the US and Europe now require full data protection in transit by 2030, pushing chip designers and data center operators to innovate faster .

Why Is Data Security During Processing So Difficult?

The problem sounds simple but is technically complex: once data is decrypted for processing, it becomes vulnerable. AI models like Anthropic's Mythos can now identify software vulnerabilities in weeks that previously took months to discover, making the window of exposure even more dangerous . This is particularly critical for enterprises handling financial records, health data, and proprietary information that could be catastrophic if exposed during AI training or analysis.

Broadcom's Emulex Connectivity division has emerged as a leader in addressing this challenge. The company developed post-quantum encryption technology for networking cards that protects data running on Fibre Channel connections, a high-speed protocol used extensively in data centers . Unlike traditional encryption approaches, this solution operates at line speed, meaning it doesn't slow down data transfers.

"We started this effort five years ago and we are way out in front of any adapter vendor. There are two key pieces. The PQC algorithms, no other vendor has this today, and every adapter vendor will have to add this," said Jeff Hoogenboom, General Manager of the Emulex Connectivity division at Broadcom.

Jeff Hoogenboom, General Manager of the Emulex Connectivity division at Broadcom

Broadcom's approach uses post-quantum cryptography (PQC), which is designed to resist attacks from future quantum computers. The company built this capability into its Gen7 silicon, which already performs real-time ransomware detection. The technology adds PQC algorithms to 64-gigabit-per-second Fibre Channel adapter cards that link storage to servers. Broadcom has already taped out 128-gigabit-per-second silicon this year, with plans to launch in 2027 for even faster connections .

How Are Companies Implementing Encrypted Data Processing?

• Host Bus Adapters (HBAs): Broadcom's SecureHBA technology encrypts data at the connection level between storage and servers, with server vendors already integrating it and storage manufacturers shipping it as standard equipment.
• Fully Homomorphic Encryption (FHE): Niobium Microsystems is developing specialized chips that allow computation on encrypted data without ever decrypting it, keeping sensitive information protected throughout the entire process.
• Optical Processing: UK-based Optalysys is using field-programmable gate arrays (FPGAs) with optical acceleration to speed up fully homomorphic encryption in cloud environments.

The market for these technologies is expanding rapidly. Currently, there are approximately 250,000 ports using secure Fibre Channel connections in data centers, but Hoogenboom estimates that 98% of servers will ship with HBAs by 2026 and 2027 . This could grow the installed base to 1 million to 2 million ports within a couple of years.

"We negotiate a random key for every connection, and then encrypt all the data running between the end points. The AES-GCM256 algorithm is in silicon while the PQC algorithms are done in firmware," explained Dale Kaisner, Principal Architect at Broadcom.

Dale Kaisner, Principal Architect at Broadcom

The latency overhead is minimal, running through the gates on the application-specific integrated circuit (ASIC) at around 2 microseconds, which is negligible for most data center operations .

What Makes Fully Homomorphic Encryption Different?

Fully homomorphic encryption represents a more radical approach. Instead of just protecting data in transit, FHE allows computation to happen on encrypted data without ever exposing the plaintext. Niobium Microsystems, founded by engineers with experience at Groq (an AI inference chip company) and Google's cloud TPU division, is pushing this technology forward .

The company taped out a prototype chip in mid-2024 on a 12-nanometer process at GlobalFoundries and is now working with chip designer SemiFive on an 8-nanometer version to be built by Samsung . Rather than waiting for the latest manufacturing process, Niobium's approach focuses on specialized polynomial processing, which delivers significant performance gains without requiring cutting-edge silicon technology.

"We are trying to do a repeat of the TPU and the Groq LPU. One of the biggest challenges of bringing a chip to market is you need to have the whole stack and you also have to convince people to develop applications," noted John Barrus, Vice President of Product at Niobium Microsystems.

John Barrus, Vice President of Product at Niobium Microsystems

Niobium has implemented its design in an FPGA and is building a cloud service that processes encrypted data. The company calls this approach a "Fog" rather than a cloud, because the encrypted data remains opaque to the service provider . When the custom ASIC becomes available, it is expected to be up to 50 times faster than the FPGA version.

The security model is elegant: customers encrypt their data, Niobium performs computations on the encrypted information, and sends back results that only the customer can decrypt using their private key. The company never sees the decryption key, making data exposure virtually impossible .

What Real-World Applications Are Emerging?

Three primary use cases are being implemented with FHE technology. The first is secure search, particularly retrieval-augmented generation (RAG) systems that work with sensitive documents without exposing their contents. The second is deep learning recommendation engines that operate on encrypted customer data. The third is federated learning for classification and predictive maintenance tasks using sensitive information from multiple parties .

For healthcare and medical records, the implications are significant. According to Barrus, 95 to 98% of healthcare data could now be used in cloud AI applications while remaining fully encrypted, opening possibilities that were previously impossible due to privacy regulations .

Broadcom is also preparing for the next generation of security requirements. The company's Gen8 silicon will maintain line-rate performance while enabling homomorphic encryption, with server silicon expected in mid-2027 . This represents a coordinated effort across the industry to close security gaps before regulatory deadlines arrive.

The race to secure data during processing is not just a technical challenge; it is becoming a competitive advantage. Companies that can offer transparent, high-speed encryption without sacrificing performance will be positioned to capture the growing market for AI applications that handle sensitive information. As regulations tighten and AI models become more capable at finding vulnerabilities, the infrastructure to keep data protected throughout its lifecycle is no longer optional, it is essential.