I remember the first time someone explained fully homomorphic encryption to me at a meetup — the guy basically said "imagine doing math on a locked box without ever opening it" and then immediately followed with "also it's completely impractical." That was seven years ago. Google apparently just decided "impractical" is a solvable UX problem, not a cryptography problem.
On August 14, 2026, Google dropped HEIR (Homomorphic Encryption Intermediate Representation) as part of its Private Computing Toolkit, and the pitch is almost absurdly simple: take a pre-trained model that normally chews on plaintext, run it through this open-source compiler, and get a version that can do inference on encrypted inputs. No cryptography PhD required. Google is literally calling it a one-click path for non-experts to bolt encrypted inference onto production apps.
That's a big claim. And Hacker News noticed — 485 points, 282 comments is not a normal reaction to a compiler toolchain announcement.
Here's what actually got people talking, though, and it wasn't the marketing copy. It was the numbers buried in the technical discussion:
- 24 seconds of latency on a UCI health prediction task
- 228 to 489 seconds — that's up to eight minutes — for a Criteo click-prediction task
- All of this on a single-threaded CPU
<> Even supportive commentary points to inference times measured in tens of seconds to many minutes, which is far slower than conventional AI serving./>
Eight minutes to predict whether someone clicks an ad. Let that sink in for a second. Somewhere a product manager is quietly closing this tab.
But I don't think that's the point, and I think a lot of the HN pile-on missed it. Nobody serious thinks HE-based inference is replacing your recommendation engine's hot path next quarter. The point is that Google just took something that used to require a research team fluent in lattice cryptography and turned it into a compiler flag. That's a categorically different kind of progress than
