The only number that matters in Google's homomorphic encryption push is 10-50x, and it is unlabelled

A post on dev.to reports that Google has announced a homomorphic encryption result that "could make private AI inference practical at scale," citing a Google security blog post and a Hacker News thread with 270 points and 163 comments ^[2]. For anyone weighing whether to run inference on encrypted data, exactly one number decides the question, and the post gives it as a range with no labels attached: overhead reduced from 1000x to roughly 10-50x, "depending on the model" ^[6].

That phrase is where the deployment case lives or dies. A 5x spread separates the top and bottom of the claimed range ^[8], which is the difference between a request budget you can absorb and one you cannot. The post's own illustration shows the problem: a 10ms plaintext inference becomes 100 to 500ms encrypted ^[7], numbers that are simply the range multiplied out rather than measurements of anything ^[9]. Nothing in the write-up says which model gives 10x and which gives 50x, at what batch size, or whether the plaintext baseline ran on the same silicon ^[13].

The mechanism described is plausible enough on its face: optimise the scheme for the matrix operations that dominate neural network work, accelerate homomorphic operations on TPUs, and combine homomorphic encryption with secure multi-party computation ^[6]. The author treats the TPU element as the decisive part, calling custom HE hardware "a game-changer for performance" ^[16]. That is an assertion about hardware, not a throughput figure, and it is the kind of sentence that tends to appear where a benchmark table should be.

Credit where due: the post does not oversell the ceiling. It notes that even at 10-50x, real-time applications may struggle, that models with billions of parameters remain challenging, that losing a key destroys the data, and that no standard HE protocol exists, so implementations are proprietary ^[10]. It also situates the work against Microsoft SEAL, available since 2015, IBM HELib, the ML-focused startup Zama, and DARPA-supported OpenFHE ^[11]. The historical baseline it cites is that HE has been possible since Craig Gentry's 2009 paper and has run thousands to millions of times slower than plaintext ever since ^[4][5].

What is absent matters more than what is present. There is no accuracy delta, no cost per query, no memory or ciphertext-size figure, and no throughput number as distinct from latency ^[14]. The applications sketched for healthcare, banking, personal assistants and GDPR or CCPA scope are the author's inferences about what would follow if the numbers held ^[15], and the closing prediction that HE-accelerated inference becomes a standard cloud offering in two to three years is also the author's ^[12]. The provenance signal offered for the whole account is a Hacker News score ^[2].

Watch for one artefact and one only: a per-workload benchmark that names the model, the batch size, the hardware on both sides of the comparison, the accuracy loss from any polynomial approximation, and the cost per thousand queries. Until a vendor publishes that table, the honest planning assumption is the old one, because a range that moves 5x on unnamed variables ^[8] cannot be put into a capacity model. The second thing to track is standardisation ^[10]; proprietary schemes mean any pilot you run is a bet on one provider's key format as much as on its speed.