Worried your last query to a cloud AI revealed a little too much about you? There's a way to compute on encrypted data without ever decrypting it, called fully homomorphic encryption, or FHE. The catch has always been speed: on today's CPUs and GPUs, computing on encrypted data can take thousands — even tens of thousands — of times longer than working with the plaintext. Intel says it has a chip that closes much of that gap. At the IEEE International Solid-State Circuits Conference (ISSCC) in San Francisco last month, it demonstrated Heracles, which sped up FHE tasks by as much as 5,000-fold compared with a top-of-the-line Intel server CPU.
This is a hardware story, not a GEO story, and we'll keep it honest on that. But the direction it points — AI you can use without handing over your raw data — matters to anyone thinking about how people will interact with AI over the next decade.
Key takeaways
- Fully homomorphic encryption (FHE) lets a server compute on encrypted data without ever decrypting it, but it has historically been thousands of times slower than working with plaintext. - Intel demonstrated Heracles at ISSCC, claiming up to a 5,000x FHE speedup versus a top Intel server CPU, and calls it the first FHE hardware that works at scale. - Heracles is physically large (~200 mm² versus ~10 mm² for prior research chips), built on 3nm FinFET, paired with two 24GB high-bandwidth memory chips, and liquid-cooled like a training GPU. - In a live private-query demo, it answered in ~14 microseconds versus ~15 milliseconds on an Intel Xeon — roughly 1,000x faster per query. - The GEO-adjacent implication is modest but real: cheaper, faster privacy-preserving computation lowers a barrier to AI adoption, and more AI use means more AI-mediated discovery over time.
The breakthrough: Heracles at scale
Startups have been racing to commercialize FHE acceleration, but Sanu Mathew, who leads security circuits research at Intel, argues the company has a real lead because Heracles can do more computing than any FHE accelerator built so far. "Heracles is the first hardware that works at scale," he said.
The scale shows up physically and in performance. Where other FHE research chips have been around 10 square millimeters or less, Heracles is roughly 200 square millimeters — about 20 times larger — built on Intel's most advanced 3-nanometer FinFET process, flanked by two 24-gigabyte high-bandwidth memory chips, and packaged with liquid cooling, a setup usually reserved for AI training GPUs. It is, in short, a serious piece of silicon aimed at making encrypted computation practical rather than merely possible.
Why it matters: the private-query demo
The clearest illustration came in a live demo of a private query to a secure server. The scenario was voter verification: a state holds an encrypted database of voters and their ballots, and a voter wants to confirm her vote was recorded correctly without revealing her identity or how she voted. With FHE, she encrypts her ID and ballot, sends them to the government database, and the system determines whether they match without ever decrypting the data, returning an encrypted answer she decrypts locally.
