Quantum in Finance: A Reality Check
A practitioner study, August 2026. Every factual claim is cited to a named, dated source. Not investment advice.
Two things about quantum computing are true at the same time, and holding both is the whole point of this piece.
The first: the hardware is genuinely advancing. In 2024 Google's Willow chip crossed "below threshold" — the point where adding more physical qubits makes the logical error rate fall instead of rise (Google Quantum AI, Nature, Dec 2024). IBM has published a concrete path to a fault-tolerant machine, Starling, targeted for 2029 with 200 logical qubits (IBM Quantum, Jun 2025). This is real engineering, not a press cycle.
The second: for the finance problems people actually want to run today, at today's sizes, classical methods keep winning — and the gap isn't closing on the schedule the roadmaps imply.
The thesis of this study is the narrow, defensible version of that tension:
For near-term finance specifically, improving classical heuristics and machine learning keep clearing the bar quantum would need to beat, so the practical adoption timeline for quantum-in-finance keeps slipping — even while the hardware genuinely advances.
I'm scoping this tightly on purpose. The broad claim — "AI is delaying quantum" — is wrong, and I'll say why in plain terms before building anything on top of it.
Intro: what a qubit buys you, in one paragraph
A classical bit is 0 or 1. A qubit holds a superposition — a weighted combination of both — and a register of n qubits represents a probability distribution over all 2ⁿ outcomes at once. The catch is measurement: you don't read out that whole distribution, you get one sample, and the amplitudes collapse. So quantum computing isn't "compute everything in parallel and read the answer." It's a narrower craft: arrange interference so that the outcomes you want become likely, and take many shots to estimate a distribution. Almost every finance use case below lives or dies on that one constraint — how expensive it is to load your data into those amplitudes, and how many shots and how much circuit depth you need before noise eats the result.
The state of AI, compressed to what matters here
Only one property of the current AI wave feeds this argument: classical methods are still improving quickly, and cheaply, on exactly the problems quantum was supposed to win. Better heuristics, GPU-scaled Monte Carlo, and modern ML keep raising the bar a quantum machine has to clear. That's the moving target. When people say "quantum is five years away," they usually mean five years from a fixed classical baseline — but the baseline isn't fixed.
The state of quantum, honestly current (2026)
2026 is being framed as an inflection year, and fairly. Error correction has gone from theory to demonstrated: neural-network decoders like Google DeepMind's AlphaQubit cut logical error rates roughly 30% below the best classical matching decoders (Google DeepMind, Nature, Nov 2024). The field is converging on qLDPC codes to shrink the physical-qubit overhead per logical qubit. Roadmaps now assume hybrid quantum-classical as the default operating mode.
All of that is upstream of finance. It's about making a reliable logical qubit and running a deep circuit at all. None of it, yet, changes which tool a bank should reach for to price a book of options this quarter.
The thesis, and the honest part first
Here is the bidirectional reality the broad "AI delays quantum" take gets wrong, stated out loud because the rest of the argument isn't credible without it:
- AI is accelerating quantum, not delaying it. The same machine-learning wave is producing better error-correction decoders (AlphaQubit above) and better circuit design. The two fields are collaborators as much as competitors.
- Classical gains are polynomial; quantum advantages are often exponential. Where a genuine exponential separation exists, classical catching up buys years, not a permanent reprieve — the gap re-widens at scale. So "classical keeps winning" is a statement about problem sizes and timelines, not about the physics forever.
Scope is what makes the thesis survive both objections. I'm not claiming quantum won't matter. I'm claiming that for near-term horizons, on finance problems, at today's sizes and today's P&L, the classical incumbent keeps clearing the bar — and every year it does, the adoption date moves right.
The mechanism behind that has a name: dequantization. In 2018 Ewin Tang took a quantum recommendation-system algorithm that was considered one of the strongest candidates for an exponential speedup, and wrote a classical algorithm that matched it to within a polynomial factor (Nature Reviews Physics, 2022). The lesson generalized: many claimed quantum-ML speedups quietly assumed the data was already loaded in a convenient quantum form, and once you charge for that loading, the advantage evaporates. Dequantization is classical methods clawing back ground that was announced as quantum territory — and it keeps happening.
Futuristic finance applications — the receipt
This table is the centerpiece of the study. I built it by hand, one row at a time, from the primary literature. Each verdict is what the current evidence actually supports — not what a vendor deck claims, and not a blanket "quantum is hype." The point is that the answer is different for each use case, and you can't get the rows right without understanding both sides.
| Finance use case | Classical incumbent | The bar quantum must clear | Honest near-term verdict |
|---|---|---|---|
| Portfolio optimization | Mixed-integer solvers + problem-tailored heuristics on constrained mean-variance | Beat them on realistic constrained sizes (hundreds–1,000 assets), not toy instances | Classical wins today. A 2025 benchmark of 250 instances up to 1,000 assets found all solved to proven optimality by MIP in seconds, with heuristics consistently beating quantum at fixed runtime — "only very limited room for a potential quantum advantage." Quantum was testable only up to ~30 assets.¹ |
| Option / derivative pricing | Monte Carlo + variance reduction, GPU-accelerated | A quadratic speedup that survives data-loading and noise | Real in theory, not near-term. Quantum amplitude estimation gives a genuine quadratic speedup over Monte Carlo, but the result needs fault-tolerant hardware and efficient state preparation; on today's noisy machines the depth required to realize it is where noise takes over.² |
| Risk / stress simulation | Classical Monte Carlo on HPC | A real wall-clock win at production scale | Same speedup, same blocker. The identical amplitude-estimation advantage applies to Value-at-Risk, but "NISQ noise entirely destroys the advantage at circuit depths needed for meaningful amplification." Production scale here waits on fault tolerance.³ |
| Fraud / ML tasks | Modern classical ML on GPUs | An advantage not erasable by dequantization | Weakest quantum case. This is exactly where dequantization bites: the flagship exponential-speedup QML claim was reduced to a polynomial gap once classical sampling was applied and data-loading costs were charged honestly.⁴ |
| Cryptography (the real one) | RSA / ECC today → NIST PQC standards | N/A — this is a threat and migration, not a speedup | The one urgent-today item. Not a place quantum helps finance — a place it threatens it. A future machine running Shor's algorithm breaks RSA/ECC, and "harvest now, decrypt later" means long-lived financial data recorded today is already exposed. Standards are final and migration is underway.⁵ |
The rows don't rhyme, and that's the finding. Optimization: classical wins on the merits, now. Pricing and risk: quantum has a real edge on paper that is gated behind hardware that doesn't exist yet. ML/fraud: the quantum case is actively eroding. Cryptography: the only row that's real, dated, and urgent — and it's a defensive problem, not a speedup.
How to be ready: the cryptography row is the actionable one
Everything above says "not yet" or "watch it." This row says "act now," and it's squarely in the readiness-and-governance lane rather than the speculative one.
The finalized standards landed on 13 August 2024: NIST published FIPS 203 (ML-KEM, for key exchange), FIPS 204 (ML-DSA, for signatures), and FIPS 205 (SLH-DSA, a hash-based backup) — the end of an eight-year process (NIST / CSRC, Aug 2024). These are the algorithms banks migrate to.
The urgency isn't the arrival of a code-breaking machine — it's harvest now, decrypt later. An adversary can record encrypted traffic today and decrypt it whenever a cryptographically-relevant quantum computer arrives. For data with a long confidentiality life — settlement instructions, correspondent-banking messages, cardholder flows — the clock started years ago. A US Federal Reserve working paper examines exactly this exposure for the financial system (Federal Reserve, FEDS 2025-093). Consensus estimates for a code-breaking machine cluster around 2030 give or take a couple of years, and 2025 algorithmic work on quantum factoring pulled resource estimates down, not up (The Quantum Insider, Aug 2026).
The migration is not hypothetical. SWIFT is targeting a PQC-enabled SwiftNet by 2027; the EU roadmap wants critical financial systems transitioned by 2030. A practical readiness posture for a financial institution looks like: inventory where RSA/ECC lives, prioritize long-lived secrets first, adopt crypto-agility so algorithms can be swapped without re-architecting, and pilot hybrid (classical + PQC) key exchange now rather than at the deadline.
Honest conclusion: what's real, what's theatre, what's urgent
Real: the hardware progress. Below-threshold error correction and credible fault-tolerance roadmaps are not marketing. And where a true exponential separation exists, classical parity is temporary.
Theatre: any near-term "quantum advantage in finance" claim for optimization, pricing, risk, or fraud. The evidence today runs the other way — classical solvers close instances to optimality in seconds, and the quantum speedups that are real are gated behind machines that won't exist for years. A vendor demo on 30 assets is not a signal about a 1,000-asset book.
Urgent: post-quantum cryptography. This is the one item that is dated, standardized, and already exploitable through harvest-now-decrypt-later. It's also the least glamorous — a migration project, not a moonshot — which is probably why it gets the least attention relative to its actual risk.
The practitioner's move isn't to buy quantum optimization or to dismiss quantum entirely. It's to migrate cryptography now, keep a standing watch on the pricing-and-risk rows for the fault-tolerance milestone that would change the verdict, and treat any near-term finance-advantage claim as something to check against a classical baseline that keeps getting better.
Sources
- Quantum Portfolio Optimization: An Extensive Benchmark — arXiv:2509.17876 (Sep 2025, rev. Jul 2026).
- Stamatopoulos et al., Option Pricing using Quantum Computers — Quantum 4, 291 (Jul 2020).
- Quantum Amplitude Estimation for tail-risk pricing, NISQ noise analysis — arXiv:2603.15664 (2026); Woerner & Egger, Quantum risk analysis, npj Quantum Information (2019).
- Tang / dequantization — Dequantizing algorithms to understand quantum advantage in machine learning, Nature Reviews Physics 4, 692 (2022).
- NIST FIPS 203/204/205 — CSRC, 13 Aug 2024; Federal Reserve FEDS 2025-093, Harvest Now, Decrypt Later; PQC migration timelines, The Quantum Insider, Aug 2026.
This is a theoretical practitioner study — no code, no benchmark of my own. The contribution is the synthesis: reading both sides of each use case and stating a scoped, cited verdict. Field figures move monthly; dates are noted so they can be re-checked.
Have thoughts on this?
Let's continue the conversation on LinkedIn.
Get new posts by email
Occasional notes on the BA → AI PM journey. No spam.