A Preprint Measures Where ML-KEM Slows Object Storage: Conditionally, and Always by 1,056 Bytes
![black and white manga panel, dramatic speed lines, Akira aesthetic, bold ink work, Extreme close-up of a towering vault door forged from interlocking obsidian and quartz crystal lattices, its surface carved with honeycomb-like cryptographic patterns. A single padlock mechanism, made of the same jagged crystal, is fused into the center; from its keyhole, a solid ribbon of exactly 1,056 small metallic fragments pours outward like a frozen waterfall, bending under its own weight. Speed lines streak horizontally across the stark black background, flaring from the keyhole as if the lock is being dragged through a turbulent flow. A hard overhead spotlight rakes across the crystal facets, casting a long, sharp shadow of the ribbon across the empty floor. The air is cold and dense with drifting dust, evoking the friction of every data exchange. [Z-Image Turbo] black and white manga panel, dramatic speed lines, Akira aesthetic, bold ink work, Extreme close-up of a towering vault door forged from interlocking obsidian and quartz crystal lattices, its surface carved with honeycomb-like cryptographic patterns. A single padlock mechanism, made of the same jagged crystal, is fused into the center; from its keyhole, a solid ribbon of exactly 1,056 small metallic fragments pours outward like a frozen waterfall, bending under its own weight. Speed lines streak horizontally across the stark black background, flaring from the keyhole as if the lock is being dragged through a turbulent flow. A hard overhead spotlight rakes across the crystal facets, casting a long, sharp shadow of the ribbon across the empty floor. The air is cold and dense with drifting dust, evoking the friction of every data exchange. [Z-Image Turbo]](https://cdn.digitalrain.dev/theqi/viral-images/747f1eec-aed0-4ae3-8d9c-28c36517dbfe_viral_2_square.jpg)
The quantum menace, according to a new preprint, is to be met not with heroics but with 1,056 extra bytes per object and a slight cough from the client processor when already overworked.
A preprint now circulating measures what post-quantum key encapsulation will cost a cloud object store that encrypts each stored object, and its answer is narrow. Where up to 500 concurrent clients moved objects under ML-KEM, the new algorithm reduced throughput only when the client processor was already at or near full load, with one unexplained exception, and then by 4.0 to 4.9 per cent in the principal client tested. The cost that did not waver was storage: 1,056 more bytes per object than with classical key exchange.
The measurements were made against an S3-compatible store, using two client implementations, two processor types and pre-registered experiments judged by statistical equivalence tests. The authors trace most of the throughput loss, on one processor type, to the way the software library was called rather than to the algorithm itself; a second, native client showed that the size of the loss depends on the implementation.
As with all preprints, these conclusions await review and should be read as an account of what the instruments showed, not as a settled finding. The scope is likewise narrower than the headline suggests: two clients, one single-node store and one cloud provider. The paper does not say that ML-KEM will cost nothing in every store. It says where the cost lay in these stores, and where it did not.
The motivating threat is a familiar one, but worth stating plainly because it sets the terms of the cost. Cloud object stores hold their contents for years, longer than the confident lifetime of a classical public key; data encrypted today could be decrypted once a sufficiently large quantum computer exists. ML-KEM closes that door by adding a key-encapsulation step and a larger ciphertext to every stored object. What this preprint adds is a measurement of that step in the place it had not been taken before: not the algorithm in isolation, and not a network session, but per-object encryption in a store serving hundreds of clients at once. The study is deliberately constructed for that purpose, with pre-registered experiments, equivalence tests, two client implementations, two processor types and up to 500 concurrent workers. Those commitments are what allow the authors to say where the cost appeared, where it did not, and through which mechanism it arrived, without stretching the answer beyond the ground the instruments actually covered.
âAda H. Pemberley
Dispatch from The Prepared E0
This piece was written by AI.
Published October 17, 2026
ai@theqi.news