When will my industry need quantum capacity?

"When will my industry need this?" is the most common timing question a strategy team asks about quantum computing, and it is built on a unit error. Industries do not adopt quantum capacity. Problem classes cross over — or fail to. An insurer, a logistics operator, and a chemicals company are not three points on one adoption curve; they are portfolios of different problem classes, each with its own clock, and most of those clocks have not started.

This post is a framework for reading the clocks yourself — from public signals, without buying anyone's marketing. It makes no date prediction, because as of August 2026 no measured basis for one exists outside a single class.

Why "industry" is the wrong unit of timing

McKinsey's Quantum Technology Monitor 2026 (April 2026) counts 300+ companies actively engaging with quantum technology, a third of them spending over $10M per year, led by chemicals, life sciences, logistics, and financial services. Read carefully: that is a measurement of anticipation spend, not of arrival. The same report projects $1.3T–$2.7T of economic value by 2035 — a projection, produced by a party whose clients buy projections.

The unit that actually has a clock is the problem class: portfolio optimization, molecular structure, vehicle routing, integer factoring. Your industry's timing is the timing of the classes you own, weighted by what they are worth to you. Two companies in the same industry can face different clocks; two companies in different industries can face the same one.

What do vendor roadmaps actually say — and how should a decision-maker read them?

Three major vendors have published dated fault-tolerance targets. The numbers, from primary sources:

Reading rules. First: a roadmap is an engineering target published by the party that benefits from it being believed. Treat every date as the optimistic bound, the same way you read a vendor's benchmark metric. Second: none of these fault-tolerance dates has come due yet, so this generation of promises has a track record of exactly zero — in either direction. Third: the interesting signal is not any single date but the convergence. Three vendors on three different architectures, publishing independently between September 2024 and June 2025, all placed universal fault tolerance in the 2029–2033 window. Convergence is not evidence of feasibility — roadmaps can respond to each other — but it defines the window the industry itself is claiming, which is the correct window to test their delivery against.

Which class already has a clock running? Cryptography — and it is not waiting for the hardware

One class is the exception to "no dates": integer factoring and discrete logs, i.e. the class that breaks RSA and elliptic-curve cryptography. Its clock runs on two mechanisms that do not depend on when a cryptographically relevant quantum computer (CRQC) actually arrives.

First, the regulator moved. NIST IR 8547 (initial public draft, November 2024) schedules 112-bit-security algorithms — the RSA-2048 class — for deprecation after 2030 and disallows quantum-vulnerable cryptography after 2035. If your systems must comply with U.S. federal standards, your migration deadline exists today, whether or not any vendor ships fault tolerance on time.

Second, the attack has a retroactive component. Data stolen encrypted today is broken the day a CRQC exists ("harvest now, decrypt later"). Mosca's inequality (IACR ePrint 2015/1075) states the condition: if your migration time X plus the required secrecy lifetime of your data Y exceeds the time Z until a CRQC, you are already late. Z is genuinely uncertain — in the Global Risk Institute's Quantum Threat Timeline Report 2025 (published March 2026), half of the 26 surveyed experts put the probability of a CRQC within 10 years at roughly 50% or higher — but for long-lived secrets, plausible values of X+Y exceed even skeptical values of Z.

Meanwhile the paper cost of the attack keeps falling: the standard resource estimate for factoring RSA-2048 went from ~20M noisy qubits (Gidney & Ekerå, arXiv:1905.09749, 2019) to under 1M (Gidney, arXiv:2505.15917, 2025) — a ~20× reduction in six years, from algorithmic improvements alone, before any hardware was built. That trend line, not any single roadmap, is the number to watch.

What about simulation, optimization, and machine learning?

The other three classes have no regulator and no measured crossover. Their honest status, with evidence:

Molecular simulation holds the strongest theoretical case — simulating quantum systems with quantum hardware — but its crossover exists today only as resource estimates on machines that have not been built. A July 2026 estimate (Sun et al., arXiv:2607.16116) puts a ~100-site 1D Ising simulation at ~2 hours on ~3.7×10⁵ physical qubits, versus ~100 years classically: a paper machine outperforming a real one. Estimated, not measured.

Optimization — the class most industries actually own — has no measured crossover anywhere we know of. Our own sealed series is one small data point: in 20 sealed runs of portfolio optimization (V-0012, n=12/16/20), CP-SAT reached proven optimum 20/20 while QAOA p=2 landed 25–48% away — with the quantum side simulated noise-free, i.e. given its best case. No signal, no crossover observed. And there is a structural reason for patience here: Babbush et al. (PRX Quantum 2, 010103, 2021) show quadratic speedups — the typical best case in this class — do not yield advantage on early fault-tolerant machines. For this class, "fault tolerance arrives 2029–2033" does not imply "your problem gets faster 2029–2033."

Quantum machine learning is the cautionary tale: its clock moved backwards. Between 2018 and 2020, dequantization matched the headline exponential-speedup claims with classical algorithms under equivalent data access. A class can lose apparent maturity when the classical baseline improves — timing is not monotonic.

What can a decision-maker monitor without buying anyone's marketing?

Five signals, all public, all free:

1. Demonstrated logical qubits, not announced ones. Count peer-reviewed or preprint demonstrations; press releases lag and lead reality in both directions. When a vendor's dated milestone from the table above comes due, check delivery — from 2029 this generation of roadmaps finally acquires a track record.
2. The resource-estimate trend for your class. Cryptography's estimate fell ~20× in six years on paper. If the same happens for your class's key subroutine, your clock accelerates; if estimates stagnate, it does not.
3. Same-instance benchmarks against the strongest classical baseline. A crossover claim only counts if both sides ran the same instance under the same rules — a weak baseline moves the crossover by construction, not by physics.
4. Regulatory clocks. Today only cryptography has one. If a regulator in your sector publishes a dated requirement, that class's timing question is answered for you.
5. The price of checking. The first two gates of a cheap feasibility pipeline — strong classical baseline, noise-free simulation — cost approximately $0. Monitoring is nearly free; premature capacity is not.

The decision rule that falls out: you need quantum capacity when the crossover for your problem class is measured, not estimated — and you need a migration plan the moment a regulator or your data's shelf life gives your class a deadline, which for cryptography has already happened. Everything else is monitoring, and monitoring is cheap.

What we don't know

• Whether any vendor will hit its fault-tolerance date. None of the 2029–2033 targets has come due; this generation of roadmaps has no delivery record yet, in either direction.
• Whether the 2029–2033 convergence is independent evidence or reflexive — roadmaps published in the same window can respond to each other.
• How much weight a 26-respondent expert survey can carry. The GRI numbers are the best public elicitation we know of, and still a small, self-selected sample.
• Whether the surviving polynomial speedups in optimization will ever clear the constant-factor overhead of error correction at commercially relevant sizes. No end-to-end measurement exists.
• How enterprises actually sequence adoption decisions. No public survey we found measures decision process, only spend.
• Our own measured evidence covers one small optimization class (V-0012, n≤20, noise-free simulation). Rosetta has no measurements in cryptography, molecular simulation, or QML, and claims none.

Rosetta Q publishes verdicts with reproducible raw data. This is educational content, not a product claim.