Why Quantum Won’t Save Wall Street (Yet): An Honest Assessment of Quantum Computing in Finance

This article is part of my Quantum Utility Map Deep Dive series. It expands on the finance section of Quantum Computing by 2033 with a detailed examination of the evidence, the resource estimates, and what financial institutions should actually be doing with their quantum budgets.

The pitch and the problem

Every major bank has a quantum computing team. JPMorgan Chase, Goldman Sachs, Barclays, HSBC, Crédit Agricole, and dozens of others have invested hundreds of millions of dollars collectively in quantum research over the past decade. The pitch is compelling: Monte Carlo pricing of complex derivatives, portfolio optimization across thousands of assets, real-time risk computation that currently requires hours on classical clusters. If quantum computers can provide even a modest speedup on these calculations, the financial return would be enormous. The notional amount outstanding in the global OTC derivatives market alone is approximately $846 trillion (BIS, end-June 2025), with a gross market value of about $21.8 trillion.

The speedup quantum computers offer for these tasks is quadratic. Goldman Sachs’ own resource estimates show that beating classical Monte Carlo with it would take logical clock rates a thousand times or more above what those papers cite as feasible.

The numbers circulating in the opposite direction deserve the same scrutiny the resource estimates get. McKinsey’s Quantum Technology Monitor projects 400 to 600 billion dollars of value created in financial services by 2035. I went through the model behind it and could not reproduce the arithmetic from the assumptions on McKinsey’s own slide, and the projection never engages a single one of the resource estimates in the next section. Allianz Research, working the same question from the insurer’s side, reached the opposite conclusion: a race between two clocks in which the cryptographic one strikes first, because a CRQC needs only the machine while financial advantage needs the machine and an algorithm that beats a classical method still improving every year.

This article explains why the quadratic speedup falls short, what the actual resource estimates say, and what financial institutions should be doing instead. I am writing it because the gap between quantum finance marketing and quantum finance evidence has grown wide enough to be doing real harm: misallocating research budgets, distracting from urgent cryptographic migration, and creating false expectations in boardrooms where the real quantum risks and opportunities are poorly understood.

The quadratic speedup, explained without the jargon

Classical Monte Carlo pricing of a complex derivative works by simulating thousands or millions of random market scenarios, computing the payoff in each scenario, and averaging the results. The accuracy improves with the number of scenarios, but slowly: to halve the error, you need four times as many scenarios. This is the square-root scaling of classical Monte Carlo, and it is the fundamental bottleneck.

Quantum amplitude estimation, the core quantum algorithm for financial Monte Carlo, reverses this relationship. To halve the error, you need only twice as many quantum operations. This is a square root improvement over the classical approach. In complexity theory terms, quantum provides an O(1/ε) convergence rate versus classical O(1/ε²), where ε is the target precision.

The improvement is polynomial, not exponential, and its size depends on the target precision. The quantum algorithm does the same fundamental thing as the classical algorithm (sample paths, compute payoffs, aggregate results), just with a more efficient convergence rate. The classical algorithm can be parallelized trivially across thousands of GPUs. The quantum algorithm in the resource estimates below runs as one coherent computation with a long serial critical path, and error correction adds its overhead to that path. That critical path is a property of the circuits those papers compile and not a law. Low-depth and iterative amplitude-estimation schemes trade depth against sample count, and extra qubits can fund factories or batching. What none of them has yet produced is a useful financial crossover.

This structural asymmetry (unlimited classical parallelism versus serial quantum coherence) is what makes the quadratic speedup insufficient in practice. I have written about the same structural problem in the context of Grover’s algorithm versus AES encryption, where the quadratic quantum speedup is similarly eroded by error correction overhead and classical parallelism. The finance case is structurally identical.

What the resource estimates actually say

Three papers define the current state of quantum finance resource estimation, and all three come from Goldman Sachs researchers, two of them jointly with IBM, the institutions with the strongest motivation to find quantum advantage in finance.

Chakrabarti, Krishnakumar, Mazzola, Stamatopoulos, Woerner, and Zeng (Goldman Sachs/IBM, 2021) provided the first complete resource estimate for useful quantum derivative pricing. Using autocallable and TARF (Target Accrual Redemption Forward) derivatives as benchmark use cases, they found that achieving quantum advantage requires approximately 8,000 logical qubits and a T-depth of 54 million, with the full program executing on the order of a second, the authors’ estimate of the classical Monte Carlo pricing time a quantum machine must beat.

Stamatopoulos, Mazzola, Woerner, and Zeng (Goldman Sachs/IBM, 2022) extended the resource analysis from pricing to market risk and restated the 2021 requirement in throughput terms: a processor executing roughly 10⁷ T-gates per second (10 MHz), rising to 50 MHz to match a strict one-second classical pricing time. Using quantum gradient methods, they cut the clock rate needed for computing greeks by a factor of about seven and cite contemporary estimates of feasible logical clock rates of around 10 kHz.

Stamatopoulos and Zeng (Goldman Sachs, 2024) applied quantum signal processing (QSP) to encode derivative payoffs directly into quantum amplitudes, eliminating costly quantum arithmetic. This brought the requirements to approximately 4,700 logical qubits and 10⁹ T-gates, with a required T-gate throughput of approximately 45 MHz. Against the arithmetic-based state of the art they benchmark, Stamatopoulos and Zeng cut T-gates by 16×, logical qubits by 4× and the required logical clock rate by 5×.

These are impressive algorithmic improvements. The reduced requirements still exceed both what anyone has demonstrated on hardware and the logical rates the papers themselves treat as foreseeable. The Goldman Sachs papers cite feasibility estimates of around 10 kHz for logical clock rates. No vendor roadmap includes a sustained logical gate throughput at all. The finance applications require 10–45 MHz: a gap of roughly three orders of magnitude.

Even with the most optimized algorithms available, a quantum computer fast enough to price derivatives competitively with classical Monte Carlo would need logical gate rates a thousand times or more beyond the feasibility estimates the Goldman Sachs papers themselves cite, on machines whose public roadmaps do not specify logical gate throughput at all.

The three barriers, and why they are structural

Barrier 1: The clock rate gap

Published fault-tolerant roadmaps specify logical qubit counts and total operation budgets, not clock rates; the error-correction cycle times of demonstrated hardware imply logical clock rates in the kilohertz range. Finance applications require 10–45 MHz. This gap cannot be closed by adding more qubits. It is a function of the physical speed of error correction cycles: how fast syndrome measurements can be performed, decoded, and fed back to the quantum processor. Improving clock rates by three orders of magnitude requires fundamental advances in decoder speed, measurement fidelity, and control electronics throughput. These advances may come, but they are engineering challenges distinct from the qubit-scaling progress that dominates quantum computing roadmaps.

Barrier 2: The parallelism asymmetry

A classical Monte Carlo simulation distributes trivially across GPU clusters. A bank running derivative pricing on a cluster of 10,000 NVIDIA H100 GPUs achieves massive parallelism at commodity cost. The quantum algorithm must run as a single coherent computation; you cannot split a quantum amplitude estimation circuit across multiple processors the way you split classical Monte Carlo across GPUs. This means the quantum computer must be faster per-unit-of-computation than the entire classical cluster, not faster than a single classical core. The quadratic speedup must overcome the parallelism of the classical alternative, and at current and projected GPU cluster sizes, it does not.

One quantum advantage sidesteps this barrier by changing the axis. The April 2026 memory separation does not race the classical cluster on time; it uses exponentially less space, and a bank’s first objection is that memory is cheap. Nobody knows what the quantum algorithm costs to run. Zhao and colleagues report fewer than 60 logical qubits but no gate count, depth or runtime, and no fault-tolerant compilation of their algorithms exists.

The nearest costed relative in the same family offers a caution and nothing more. Pradeep Niroula, Shouvanik Chakrabarti, Ruslan Shaydulin and colleagues at JPMorganChase with Quantinuum compiled the pair sketch behind Hidden Matching, a different streaming task, in a preprint posted on 5 November 2025, five months before Zhao’s. At that algorithm’s break-even point their Table II gives 581 logical qubits and 8.68×10¹⁴ Toffoli gates, which is 27.5 years if you serialize it at one Toffoli per microsecond and 2,750 years at the 10 kHz rate the Goldman Sachs papers cite as feasible. Those figures belong to Hidden Matching and have not been shown to apply to Zhao’s tasks. The serialization is also an illustration, not a schedule, since nobody has worked out how far either algorithm parallelizes; Zhao and colleagues note that their oracle sketching is largely built from commuting operations. What it tells a bank is that the only member of this family anyone has costed came back at decades to millennia of serial runtime, and that the one with the theorem attached has not been costed at all.

Barrier 3: The moving classical baseline

Classical finance computation is improving continuously. GPU-accelerated Monte Carlo, AI-driven surrogate models, better variance reduction techniques, and improved numerical methods all erode the quantum advantage over time. The quantum speedup is defined relative to a classical baseline that does not stand still. A quadratic improvement over today’s best classical approach may be negligible against the approach available in 2033, because classical methods will have improved by then as well.

This is a structural difference from quantum chemistry, where the classical difficulty for strongly correlated systems comes from exponential scaling in the simulated system rather than from the cost of more GPUs. In chemistry, there are selected active spaces where the best available classical methods cannot reach the accuracy the application requires, and adding hardware does not obviously close the gap. Electron count alone does not settle that. Basis size, symmetry, the active space and which observable you want all affect it, so there is no clean cutoff. In finance, more classical hardware always helps, and the cost per unit of compute falls every year.

Portfolio optimization: same barrier, different disguise

Portfolio optimization (mean-variance, CVaR minimization, index tracking) has been extensively promoted as a quantum computing application. The proposed quantum approaches use either QAOA (Quantum Approximate Optimization Algorithm) or quantum annealing to search for optimal portfolios across combinatorial spaces.

The proposed quantum approaches rely on heuristic algorithms (QAOA, quantum annealing) that lack even a proven quadratic scaling advantage over advanced classical solvers. And the classical alternatives (mixed-integer programming solvers, heuristic algorithms, classical simulated annealing) are mature, parallelizable, and improving. Published QAOA runs on hardware have so far used problem sizes that tuned classical heuristic solvers handle easily, and the 2024 community review of quantum optimization by Abbas and colleagues says more tools are needed to realize a quantum advantage in optimization.

Bravyi, Kliesch, Koenig, and Tang showed that the locality and symmetry of QAOA severely limit it: on certain MaxCut instances, the classical Goemans-Williamson algorithm outperforms QAOA at any constant circuit depth. The easy instances can be solved classically; the hard instances resist both classical and quantum approaches.

Risk analysis and VaR

Value-at-risk (VaR) computation and stress testing use Monte Carlo at massive scale, making them natural candidates for quantum speedup. The same barriers apply: quadratic improvement, clock rate requirements, parallelism asymmetry. The Bank for International Settlements’ Quantum Readiness Roadmap (July 2025) recommends that systemically important financial institutions support post-quantum cryptography by 2030, and makes no recommendation about quantum computational advantage for risk analysis. The paper’s subject is cryptographic readiness, so its silence is a matter of scope and not a verdict on the finance algorithms; what it does show is where a central-bank body thought the actionable risk was.

The regulators that followed it through 2026 did the same thing. The Hong Kong Monetary Authority published the most granular supervisory assessment to date, scoring its Authorized Institutions 2.3 out of 10 on its new Quantum Preparedness Index against a stated objective of full preparedness by 2030, which is a supervisory target, not a deadline for completing every migration. The Saudi Central Bank issued a circular requiring cryptographic inventory procedures by December 31, 2026 and an enterprise quantum risk assessment by March 31, 2027. Neither sets a target for quantum computational advantage in risk analysis, and neither makes its cryptographic timetable depend on one.

What would change the picture

Quantum computing may still help in finance, but the current evidence does not support near-term advantage, and strategic planning should reflect that. Three developments would change my assessment.

A super-polynomial speedup for a financial problem class. If someone discovers a quantum algorithm with an exponential speedup for a specific class of financial computation, the entire analysis above becomes obsolete. A cubic speedup would also help, though cubic is still polynomial. Whether any polynomial speedup is useful depends on the problem size at which the crossover occurs. When I first wrote this signal, I named structured problems as the most promising direction and listed sparse or low-rank covariance matrices first.

In April 2026 Zhao, Zlokapa, Neven, Babbush, Preskill, McClean, and Huang proved an unconditional exponential quantum memory advantage for estimating a quadratic form of the solution to a sparse, well-conditioned linear system, from streamed samples of its entries and with no QRAM. The result is unconditional in a strong sense: it holds even if BPP = BQP, and the dequantization arguments that undid the 2018 generation of quantum machine learning claims do not remove it.

The nearest portfolio quantity is μᵀΣ⁻¹μ, a quadratic form of the solution of Σx = μ that sets the best attainable Sharpe ratio. Covariance matrices are not generally sparse, and a low-rank one is singular, the opposite of well conditioned. A portfolio’s covariance matrix therefore does not meet the theorem’s preconditions. The access model is the harder obstacle: each sample is read once and discarded, which is not how any risk engine reads a covariance matrix.

What the theorem does is pose a question finance did not have before. Can a useful risk task be mapped into a sparse, well-conditioned streaming model while keeping the memory separation? Nobody has shown that it can. A bank that found such a mapping would still save only memory, the resource it is least short of. So the signal is not hit. For the first time since I wrote it, though, a proven result exists for an adjacent problem. What to watch now is a published mapping of a named financial task into that model, then a crossover with a stated problem size, gate count and logical clock rate, benchmarked against a tuned classical baseline, not a provable-guarantee one. I track the same question across the series in The Quantum Utility Ladder.

A breakthrough in logical clock rates. If fault-tolerant quantum hardware achieves 10+ MHz logical clock rates through advances in decoder design, measurement speed, or alternative error correction architectures, the clock rate barrier dissolves. Recent advances in algorithmic fault tolerance cut the number of error-correction rounds per logical operation, for the operations the theorem covers, from about the code distance to one, which raises logical throughput by a factor of roughly 20 to 30 at practical distances. A factor of 20 to 30 closes part of a thousandfold gap. Closing the remainder requires faster physical cycles, measurement and decoding.

A classical plateau in financial computation. If GPU-based Monte Carlo and AI surrogate models hit fundamental accuracy or scalability limits for complex derivatives, the relative value of even a quadratic quantum speedup increases. This seems unlikely given current trends in classical hardware and software, but it cannot be ruled out.

Until at least one of these developments materializes, quantum advantage in finance remains speculative.

What financial institutions should actually be doing

Priority 1: Post-quantum cryptography migration

The highest-value quantum investment for any financial institution today is post-quantum cryptography (PQC) migration. The reason is straightforward: the cryptographic threat from quantum computing is certain, near-term, and directly relevant to financial services, while the computational advantage is uncertain, far-term, and structurally weak.

The harvest-now-decrypt-later (HNDL) threat is active today. Adversaries with access to encrypted financial data (transaction records, client communications, trading strategies, M&A negotiations) can store that data now and decrypt it later when a cryptographically relevant quantum computer (CRQC) becomes available. For financial data with long confidentiality requirements (decade-plus), this means the window for protection is closing now, regardless of when a CRQC actually arrives.

NIST has finalized the post-quantum cryptographic standards ML-KEM (CRYSTALS-Kyber), SLH-DSA (SPHINCS+) and ML-DSA (CRYSTALS-Dilithium). Regulatory deadlines are already being set: the BIS recommends PQC readiness by 2030 for systemically important institutions. Forget Q-Day predictions; the deadlines are already set.

Your quantum budget should go to PQC migration first. Every dollar spent on PQC migration addresses a known, quantifiable risk with a clear regulatory driver. Every dollar spent on quantum finance advantage research addresses an uncertain, structurally challenged opportunity.

Priority 2: Monitor the field, don’t abandon it

Maintaining a small, expert quantum research team makes strategic sense even without near-term advantage. Algorithmic breakthroughs are unpredictable, and the institution that has quantum expertise when a breakthrough occurs will be positioned to exploit it faster than competitors who have to start from scratch. The key is right-sizing the investment: a monitoring and readiness posture, not a production deployment bet.

Track the three developments under “What would change the picture”: a peer-reviewed super-polynomial quantum speedup for a financial problem class; logical clock rates moving toward 10 MHz, with a demonstration above 1 MHz as the first milestone; and evidence that GPU Monte Carlo and AI surrogate models are reaching accuracy or scalability limits on complex derivatives. Add a fourth marker: a resource estimate that brings derivative pricing below 1,000 logical qubits, which would suggest a qualitative shift in algorithmic approach, not just incremental optimization.

Priority 3: Quantum-adjacent opportunities

Some financial applications are adjacent to quantum computing without requiring quantum advantage in finance itself. Quantum-resistant cryptographic infrastructure protects trading systems and client data. Quantum random number generation can supply entropy for key generation where the source and its implementation have been characterized and validated; a quantum origin does not by itself certify the entropy rate, the health tests or the extraction. And understanding quantum computing’s impact on other industries (pharma, materials, energy) informs investment analysis, venture capital strategy, and sector allocation decisions.

A bank that understands the Quantum Utility Ladder can make better investment decisions about quantum computing companies and more accurate risk assessments of quantum-exposed sectors, and can give clients better strategic advice during the transition. That analytical advantage may be worth more than any direct computational speedup.

The honest summary

Quantum computing in finance has been oversold. The quadratic speedup offered by quantum amplitude estimation and related algorithms is structurally insufficient to overcome the combined barriers of clock rate requirements, GPU parallelism, and a continuously improving classical baseline. The best resource estimates, produced by Goldman Sachs’ own research team, require hardware capabilities three to four orders of magnitude beyond current projections.

This does not diminish quantum computing as a technology. In chemistry, materials science, and physics, the quantum advantage is genuine, growing, and grounded in exponential classical difficulty. Finance simply faces a different mathematical structure: one where classical methods scale well and quantum methods face overhead that erodes their theoretical edge.

The right quantum strategy for financial institutions is to invest heavily in PQC migration (the certain, near-term threat), maintain a lean quantum research capability (the uncertain, far-term opportunity), and develop the analytical expertise to understand quantum computing’s impact on the industries that will actually be transformed. The quantum revolution in finance will come through understanding quantum’s impact on the real economy, not through pricing derivatives 2× faster.

For the full technical resource estimates underlying this analysis, see The Quantum Utility Ladder. For the broader competitive analysis across industries, see Quantum Computing by 2033. For practical guidance on starting PQC migration, see Practical Steps to Quantum Readiness and the PQC Migration Framework.

From the author