Intel Built a Chip That Computes on Encrypted Data 5,000 Times Faster. It Could Flip Surveillance Capitalism Upside Down.

The Heracles chip - five years in the making under a DARPA program - just demonstrated the first fully homomorphic encryption hardware that works at real-world scale. The implications stretch far beyond cryptography labs.

PRISM — BLACKWIRE Tech & Science — March 10, 2026

Semiconductor fabrication. Photo: Unsplash

The fundamental bargain of the internet has always been this: if you want a service to work for you, you have to let it see your data. Your search queries, your medical history, your financial records, your location - they all have to travel to a server in plaintext so the server can actually compute with them.

That bargain is broken. And Intel just demonstrated the hardware that could finally replace it.

At the IEEE International Solid-State Circuits Conference (ISSCC) in San Francisco last month, Intel unveiled Heracles - a purpose-built chip accelerator for fully homomorphic encryption (FHE), a mathematical technique that allows computation directly on encrypted data without ever decrypting it. The chip is 1,074 to 5,547 times faster than Intel's own top-of-the-line Xeon server processors at FHE operations. It's the first hardware to demonstrate FHE working at genuine scale, and it was built under a classified-adjacent DARPA program that's been running for five years without public attention.

The timing is not coincidental. FHE has gone from theoretical curiosity to urgent infrastructure priority as governments worldwide build mandatory surveillance systems disguised as child-safety laws, and as AI inference pipelines route billions of daily queries - each one exposing intimate personal data to cloud operators, law enforcement, and hackers. Heracles is a direct answer to a problem that most people don't yet know they have.

What Is Fully Homomorphic Encryption, and Why Has It Taken This Long?

FHE is, at its core, a mathematical transformation - conceptually similar to the Fourier transform but applied to encryption. It wraps data in a quantum-resistant cryptographic scheme while preserving mathematical relationships that allow operations to be performed on the ciphertext and produce correct results when decrypted.

The concept was first theorized in 1978 by Rivest, Adleman, and Dertouzos, the same team behind RSA encryption. But it wasn't until 2009 that Craig Gentry, then a PhD student at Stanford, actually proved FHE was mathematically possible with a practical construction. The problem was speed: his scheme was so computationally expensive it would take millions of years to process meaningful data on contemporary hardware.

The core challenge is data bloat combined with computational complexity. When you encrypt data for FHE, you aren't working with bytes - you're working with polynomials over massive number rings. A single bit of plaintext can expand to kilobytes of ciphertext. FHE also requires exotic operations: "bootstrapping" to cancel accumulated computational noise, "twiddling" for polynomial coefficient manipulation, and "automorphism" operations that rotate or permute encrypted data. None of these have analogues in standard CPU instruction sets, which is why even modern Xeon chips running at 3.5 GHz are brutally slow at them.

For 15 years, FHE remained a cryptographer's toy - provably correct but practically useless. DARPA decided in 2021 to change that.

Cloud data centers handle billions of daily queries - each one currently requiring plaintext access. Photo: Unsplash

Inside Heracles: A Chip Built for Math That Doesn't Exist on CPUs

Intel's Heracles chip is physically unusual. While most FHE research chips fit comfortably in 10 square millimeters, Heracles is roughly 200mm² - about 20 times larger. It's fabricated on Intel's most advanced 3-nanometer FinFET process and packaged in a liquid-cooled module alongside two 24GB high-bandwidth memory (HBM) chips, giving it 48GB of attached memory with 819 gigabytes per second of bandwidth. That's the memory configuration you see in Nvidia's AI training GPUs, not in cryptography chips.

That bandwidth isn't wasteful - it's the minimum necessary. FHE's data expansion problem means a simple query that would fit in a few kilobytes of RAM balloons into gigabytes of ciphertext. Without massive memory bandwidth, the compute cores would starve waiting for data.

At Heracles' heart are 64 compute cores, called tile-pairs, arranged in an eight-by-eight grid. Each is a single-instruction-multiple-data (SIMD) engine purpose-built for polynomial arithmetic, bootstrapping, twiddling, and the other FHE-specific operations. They're connected by a 2D mesh network with 512-byte-wide buses, allowing 9.6 terabytes per second of data flow across the chip. The chip also carries 64 megabytes of on-chip cache - slightly more than an Nvidia Hopper-generation GPU.

One key architectural decision: Intel's team made a calculated bet early in development to use 32-bit arithmetic chunks rather than 64-bit chunks for the enormous numbers FHE requires. While FHE math operates on numbers far larger than 64-bit words, breaking them into 32-bit fragments that compute independently provides parallelism and halves the size of arithmetic circuits. Sanu Mathew, who leads security circuits research at Intel, describes the chip's challenge as "balancing the movement of data with the crunching of numbers" - the chip runs three synchronized instruction streams simultaneously: one for loading and offloading data, one for internal data movement, and one for actual computation.

The project ran for five years under a DARPA initiative to accelerate FHE through purpose-built hardware. Ro Cammarota, who led the Heracles project at Intel until December 2025 and is now at UC Irvine, calls it "a whole system-level effort that went all the way from theory and algorithms down to the circuit design." The chip is not a one-trick laboratory prototype - it's engineered for the kind of real-world scale at which Intel historically operates.

The Startup Race Intel Just Fired a Warning Shot At

Intel's Heracles announcement landed in a market where at least five startups have been quietly racing to build the first commercially viable FHE accelerator. None of them have shipped. Heracles just raised the stakes dramatically.

Niobium Microsystems, spun out of Galois - a DARPA contractor in Portland, Oregon - is furthest along in the commercial pipeline. Last month, Niobium disclosed a deal worth 10 billion South Korean won (approximately $6.9 million USD) with Semifive, a Seoul-based chip design firm, to develop an FHE accelerator using Samsung Foundry's 8-nanometer process technology. Niobium has positioned itself as the first commercially viable FHE chip. That claim just got harder to defend.

Duality Technology, which holds FHE software patents and develops encrypted-query products, was also part of the original DARPA program that spawned Heracles. Its CTO Kurt Rohloff offered measured praise: "It's very good work. When Intel starts talking about scale, that usually carries quite a bit of weight." But Duality is taking a different approach - focusing on software products for the kinds of encrypted queries Intel demonstrated, arguing that at today's scale, specialized hardware isn't necessary. "Where you start to need hardware is emerging applications around deeper machine-learning oriented operations like neural nets, LLMs, or semantic search," Rohloff told IEEE Spectrum.

That caveat is telling. Duality already demonstrated an FHE-encrypted version of BERT, the Google language model that preceded GPT-era AI. The catch: FHE-encrypted BERT is only one-tenth the size of even the most compact LLMs available today. Scaling FHE to modern AI model sizes has remained the key unsolved problem. Heracles, with its HBM memory and parallelized polynomial engines, is specifically architected to attack that problem.

Other players in the space include Fabric Cryptography and Cornami, both of which have received DARPA or venture funding but have not yet publicly disclosed working silicon. Niobium's John Barrus, vice president of product, agrees that encrypted AI inference is the endgame: "There are a lot of smaller models that, even with FHE's data expansion, will run just fine on accelerated hardware."

Intel has made no commercial announcements for Heracles. No pricing, no availability timeline, no roadmap. That's either because the chip is still fundamentally a research project, or because Intel is deciding how to productize it in a market it doesn't yet fully understand. Either way, the demonstration at ISSCC was unmistakably a signal to the startup field: Intel is in the game, and Intel's version works at scale.

Encrypted computing promises computation without exposure. Photo: Unsplash

The Surveillance Problem That Makes This Urgent

FHE is not just an interesting cryptography problem. It's a direct technical response to a surveillance architecture that is expanding rapidly and deliberately.

Right now, in 2026, roughly half of all U.S. states have enacted or are advancing laws requiring online platforms to verify users' ages before granting access to social media, adult content, gaming, and a growing list of other services. The laws are framed as child safety measures. The mechanism they require is the construction of a massive, distributed identity-verification infrastructure that touches hundreds of millions of Americans.

The industry that has grown up to meet this demand includes companies like Jumio, Socure, and dozens of smaller vendors. Their systems typically work like this: a user approaches a platform, the platform routes them to a verification vendor, the vendor captures a selfie or government ID scan using AI facial recognition and age-estimation models, and returns a pass-fail signal. The problem is the data that flows through and often stays with these vendors.

Socure, one of the largest players, retains some adult verification data for up to three years under its standard contracts. Discord announced mandatory global age verification in February, then almost immediately delayed it to the second half of 2026 after user backlash - and then disclosed a separate data breach that had already exposed ID images belonging to approximately 70,000 users through a compromised third-party service. A Virginia age-verification law was blocked by a federal court last month after a First Amendment challenge from a tech industry trade group.

"Age verification strikes at the foundation of the free and open internet. It ties users' most sensitive and immutable data - names, faces, birthdays, home addresses - to their online activity." - Molly Buckley, legislative analyst, Electronic Frontier Foundation

Heidi Howard Tandy, a partner at Berger Singerman specializing in internet law, put it more bluntly: "If they say they're holding it for three years, that's the minimum amount of time they're holding it for. I wouldn't feel comfortable trusting a company that says, 'We delete everything one day after three years.' That is not going to happen."

Law enforcement access is the other dimension that verification contracts rarely foreground. Vendors include provisions in their terms of service allowing them to respond to law enforcement requests - meaning the government can, with the right paperwork, map verified identities to the content they accessed. This is surveillance infrastructure being built under the cover of protecting children.

FHE is the technical solution that the policy debate has not yet discovered. A properly implemented FHE-based age verification system could confirm that a user is over 18 without the verification service ever seeing the user's name, face, or ID document in plaintext. The computation happens on encrypted data. The vendor receives an encrypted query and returns an encrypted result. No plaintext identity is ever exposed. No database of faces and government IDs is ever assembled.

Why This Matters for AI - The Second-Order Effect Nobody Is Talking About

The age verification use case is important, but it's actually the smaller story. The larger story is AI inference privacy.

Every query you send to ChatGPT, Gemini, Claude, Copilot, or any cloud AI system travels to a server where it is processed in plaintext. The AI company can see your query. So can any hacker who compromises the server. So can any government that serves the company with a legal demand. If you ask an AI chatbot about a medical symptom, a legal situation, a mental health struggle, or a business strategy, you're exposing that information to every entity that can access the server infrastructure - now and in the future, because logs are often retained.

FHE-encrypted AI inference flips this model. Your query would be encrypted before leaving your device. The cloud server would run the AI model against your encrypted query using FHE operations. The result would come back encrypted. Only your device could decrypt the answer. The AI provider would never see what you asked or what the model said.

This is not currently possible at useful AI model sizes. A BERT-sized model with FHE is the state of the art for encrypted AI inference, and BERT is roughly 100x smaller than GPT-3, which itself is considered modest by 2026 standards. Heracles is the first hardware that could realistically close this gap, because it was designed specifically for the polynomial operations that FHE AI inference requires, and it operates at a scale that can actually handle the data expansion FHE introduces.

Duality's Rohloff confirmed this trajectory: the next frontier for FHE hardware is "neural net, LLM, or semantic search" applications - exactly the AI inference pipeline that currently forces billions of people to expose their most sensitive queries to cloud operators daily.

Nvidia has dominated the AI compute market by building chips optimized for the matrix multiplication operations that make neural networks run. What FHE requires is fundamentally different - polynomial arithmetic at massive scale, with bootstrapping and automorphism operations that GPUs handle poorly. This creates an opening. If FHE inference becomes practically viable through chips like Heracles or its successors, it could carve out an entirely new segment of the AI compute market that Nvidia does not currently own.

The DARPA Bet That Took Five Years to Pay Off

Five years ago, when Intel started building Heracles under the DARPA program, FHE was still widely considered too slow to matter. The academic consensus was that algorithmic improvements and software libraries would be sufficient to bring FHE performance to acceptable levels. DARPA disagreed, and funded a hardware-first approach.

The DARPA program that birthed Heracles also produced competing designs from other DARPA contractors - including Galois (which later spun out Niobium) and a team that worked with Duality Technology. The program's hypothesis was that no amount of software optimization would overcome the fundamental mismatch between FHE's mathematical requirements and the architecture of general-purpose processors. You cannot retrofit a CPU to efficiently do polynomial arithmetic over gigabyte-scale number rings. You have to build something new.

Heracles validates that bet. Five years of DARPA funding, Intel's most advanced 3nm process, a liquid-cooled packaging configuration borrowed from GPU design, 48GB of HBM memory, and a custom 2D mesh interconnect - all of it converging on a chip that finally makes FHE fast enough to use.

"Heracles is the first hardware that works at scale." - Sanu Mathew, Security Circuits Research Lead, Intel

Ro Cammarota, who led the five-year project before moving to UC Irvine, is direct about the accomplishment: "We have proven and delivered everything that we promised." For DARPA, the promise was a specific speedup target. The delivered range of 1,074x to 5,547x across FHE's seven key operations exceeds most of what was publicly projected.

The Political Economy of Privacy Computing

FHE's commercial success is not guaranteed, and not everyone is rooting for it.

The entire surveillance capitalism model - the business model that funds Google, Meta, and most of the ad-supported internet - depends on the ability to observe user behavior in plaintext. FHE-encrypted AI inference would break that model at the technical level. If Google's AI assistant cannot see your queries, it cannot build an advertising profile from them. If Facebook's recommendation algorithm processes your interests as encrypted polynomials it never decrypts, it cannot sell that interest graph to advertisers.

This is why the commercial deployment of FHE in consumer-facing AI is not primarily a technical problem - it's a political and economic one. The companies with the resources to deploy FHE at scale are the same companies whose business models it would destroy. Intel, as a chip vendor, has no such conflict of interest. But the cloud hyperscalers that would deploy Heracles-class accelerators in their data centers absolutely do.

The more likely early deployment path runs through regulated industries where the business case for FHE is already clear: healthcare (where HIPAA liability is enormous and FHE-encrypted genetic analysis could be genuinely transformative), finance (where encrypted query databases could satisfy compliance requirements without creating surveillance liabilities), and government (where the ballot verification demo Intel showed is directly applicable to real election infrastructure).

In all of these cases, the FHE pitch is not about privacy as a value - it's about liability reduction. A healthcare company that processes patient genetic data using FHE cannot be breached in any meaningful sense: the attacker gets ciphertext that decrypts only on the patient's device. A financial institution that runs credit checks on encrypted income data cannot be sued for leaking that data to a third party that later gets hacked. The liability math is compelling before the civil liberties argument even enters the room.

Homomorphic encryption wraps data in mathematical transformations that allow computation without decryption. Photo: Unsplash

What Comes Next: The Road from ISSCC to Infrastructure

Intel's Heracles demonstration at ISSCC was a proof of concept, not a product announcement. The chip showed what's possible. The commercial path from here is murky.

Intel is currently navigating a turbulent period in its own history - a leadership transition, manufacturing execution challenges, and market share losses in both server and desktop chips to AMD and Arm-based competitors. Heracles is a genuine technical achievement, but Intel's ability to convert technical achievements into commercial products has been inconsistent over the past several years. The question is whether the company has the organizational focus to productize FHE acceleration or whether it will be outmaneuvered by the startups it just publicly embarrassed.

Niobium's Samsung 8nm partnership is meaningful: 8nm is a mature, well-understood process with high yield rates and low cost relative to Intel's leading-edge 3nm. A commercially available FHE chip on 8nm would be significantly slower than Heracles, but also significantly cheaper and faster to produce. For many applications, a 1,000x speedup over a CPU is plenty - it doesn't need to be 5,000x.

The AI inference angle may actually be the forcing function that accelerates the whole market. As AI systems become embedded in healthcare diagnosis, legal research, financial advisory, and other high-stakes domains, the liability exposure of running inference on plaintext sensitive data becomes a board-level risk management issue. One large breach - a cloud AI provider's servers compromised, millions of medical queries exposed - could shift enterprise procurement toward FHE-capable inference infrastructure overnight.

DARPA's investment is paying off in ways beyond the specific chips it funded. The program has built a community of researchers, validated the architectural approaches that work, and created a public benchmark - Heracles - that the commercial field can measure against. The next five years in FHE hardware will be driven by companies trying to build the commercial version of what Intel just proved is possible.

The bargain of the internet is changing. Not because regulators demanded it, not because tech companies had a change of heart, but because a chip built in a lab under a defense research program just made the alternative mathematically viable. Encrypted computation at scale is no longer a theoretical promise. It's a 200-square-millimeter piece of silicon running at 1.2 gigahertz, and it just verified 100 million ballots in 23 minutes without ever learning who voted for whom.

The question now is who builds the next one, how fast they can get it into production, and whether the companies whose business models depend on seeing your data in plaintext can find a way to slow it down.

Sources: IEEE Spectrum / ISSCC 2026 proceedings; CNBC, "Online age-verification tools spread across U.S. for child safety, but adults are being surveilled," March 8, 2026; The Verge, AI daily coverage, March 10, 2026; Thinking Machines Lab / Nvidia press release, March 10, 2026; Niobium Microsystems press release; Duality Technology; Electronic Frontier Foundation commentary.

← All BLACKWIRE Reports