Harvest now, decrypt soon: post-quantum migration is a funded program, not a research topic

Begin with the part of the case that does not depend on a hardware forecast. Harvest now, decrypt later assumes an adversary copies ciphertext today, stores it, and waits for a machine that can run Shor's algorithm against the key that protected it ^[10]. Karisik's revision, harvest now decrypt soon, is a claim about the wait being shorter than most CIOs budgeted for ^[11]. The operative variable sits on the other side of that equation: how long a given flow of traffic has to stay secret. Anything with a ten-year confidentiality requirement leaving the building today under RSA or ECC is already committed ^[7]. That is a records-retention question, and it can be answered without knowing when a cryptographically relevant quantum computer arrives.

Now the hardware figures, which do less work than their size suggests. Sycamore in 2019 had 53 qubits and a calculation the piece values at 10,000 supercomputer-years ^[3]. Willow has 105 qubits and a benchmark valued at 10 septillion classical years ^[4]. Qubit count roughly doubled ^[13]. The claimed classical-equivalent time rose by a factor of about 10^21 ^[14]. Those two numbers describe different benchmark problems, and neither run was an attempt to factor a key; the article names no qubit count and no date at which RSA or ECC would actually fall ^[15]. The figure in that set that bears on cryptanalysis is the third one, a logical error rate twenty times below Sycamore's ^[5], because factoring needs error-corrected logical qubits held stable through a very long circuit rather than a headline peak count. The piece's own nod to below-threshold error correction points at exactly that ^[11].

So the defensible version of the argument is narrower than the "mathematical cliff" framing ^[2], and more useful. The state space grows exponentially with each qubit ^[6], Shor's algorithm turns that into an attack on both factoring and discrete logarithms ^[8], and a state reported to have committed more than $15 billion, with dedicated campuses at Hefei, is a plausible funder of a decade-long build ^[9]. None of that produces a date. NIST produces a date: with ML-KEM and ML-DSA finalised ^[12], the schedule for a migration program is set by asset inventory, vendor support and integration testing rather than by a standards committee. The delay has moved into procurement, where the CIO owns it.

Worth noting who is making the case. It comes from the CTO for Intelligence and Cyber at Amentum, writing as a Forbes Tech Council contributor ^[1], which is an interested party arguing for urgency. The interest does not make the argument wrong. It does mean the load-bearing items for a board paper are the two that can be checked independently: the standards are published, and the traffic you are sending this quarter has a retention life you already know.