Quantum supremacy in finance: marketing, or reality?

Every major bank now seems to have a quantum computing team. IBM, Google, IonQ, D-Wave, Microsoft and a long tail of vendors publish regular announcements suggesting finance is one breakthrough away from a quantum rewrite. Goldman Sachs has published quantum pricing research; JPMorgan has built an internal quantum research capability; consortium memberships and vendor partnerships have become standard signaling devices for innovation budgets.

And yet: not a single production trading strategy runs on a quantum computer today. No market-making engine, no intraday risk stack, no XVA platform, no limit-order-book model, no factor timing pipeline, no high-frequency execution framework. That gap between narrative and reality is not an embarrassment for the field — it is the central fact any honest discussion has to start from.

Most quantum-finance commentary is directionally wrong in two ways simultaneously. It is overhyped about the near-term use cases that dominate press releases — option pricing, alpha generation, generic “portfolio optimization.” But it is also underappreciated where the first economically meaningful consequences may actually appear: crypto-agility, post-quantum migration, specific constrained-optimization niches, and the institutional option value of building technical readiness before the hardware becomes useful.

That distinction matters. If the wrong claims dominate, decision-makers either allocate capital to science projects with no realistic deployment path, or dismiss the field entirely and miss the one area where action is already warranted. The right posture is neither evangelism nor cynicism. It is technical discrimination — claim by claim, hardware threshold by hardware threshold.

The most abused term in the field remains “quantum supremacy.” Google's 2019 Sycamore result referred to a random-circuit-sampling task said to complete in around 200 seconds, against a classical supercomputer estimate measured in thousands of years. That was scientifically important — a milestone in controllable quantum hardware — but never an economically useful computation. The task was intentionally designed to be hard for classical simulation and irrelevant to business problems.

The term that actually matters for finance is quantum advantage: solving a practically useful problem faster, cheaper, or more accurately than the best classical alternative under realistic constraints. On that metric, as of 2026, no financial use case has demonstrated a defensible quantum edge in production. In finance, usefulness has a very specific meaning — lower total cost of computation for a required accuracy target, acceptable latency, integration into classical data systems, reproducibility, operational stability, and a workflow that survives model validation, audit, cyber review and production support.

Three paradigms matter for this discussion:

• Gate-based quantum computing — IBM, Google, IonQ, Quantinuum, Rigetti. The universal model; in principle it can run Quantum Amplitude Estimation (QAE), Shor's algorithm, HHL-type linear-system methods, QAOA and various quantum machine-learning routines. In practice, current devices remain noisy and limited.
• Quantum annealing — mainly D-Wave. Not universal quantum computing but a specialized optimization architecture for problems expressible in Ising / QUBO form. The right comparison is not “general quantum vs. classical” but “specific constrained-optimization heuristics vs. classical heuristics and commercial solvers.”
• Photonic / topological ambitions — PsiQuantum, Xanadu, Microsoft's topological path. Strategically important because they target fault tolerance differently, but not yet a source of near-term production finance workloads.

The hardware bottleneck remains decisive. IBM's public roadmap points toward increasingly modular systems and a fault-tolerant milestone highlighted for 2029. Google's Willow processor, announced in late 2024, featured 105 qubits and a benchmark linked to error-correction progress — yet still implies no readiness for large-scale commercial algorithms. Progress is still measured in physical-qubit quality, error suppression and modular scaling, not in deployable logical-qubit abundance.

That last phrase is the important one: logical qubits. A finance-relevant quantum computer is not “a machine with hundreds or thousands of physical qubits.” It is a machine with enough error-corrected logical qubits to run deep circuits reliably. Under conventional fault-tolerance assumptions, one logical qubit may require hundreds to thousands of physical qubits. That is why impressive physical-qubit counts can still correspond to effectively zero useful logical compute. A reasonable 2026 base case is blunt: current hardware is scientifically meaningful, commercially premature, and still several engineering layers away from routine workload migration.

C1“Quantum computers will price derivatives faster.”

This is the strongest claim in theory and one of the weakest in deployment timing. Quantum Amplitude Estimation can reduce sample complexity from classical Monte Carlo's roughly O(1/ε²) to O(1/ε) — a quadratic improvement in accuracy scaling. If realizable, that is a very serious asymptotic result, and it is why the Goldman Sachs / IBM / ETH Zürich line of work attracted attention.

But asymptotic speedup is not wall-clock superiority. Classical Monte Carlo in finance is not standing still: multi-core CPUs, GPUs, variance reduction, adjoint methods, quasi-random sequences and surrogate models. A vanilla option prices in milliseconds; even many exotics are dominated by model setup, calibration and market-data handling rather than raw sampling cost. The reality check comes from the same research line that motivates the excitement — Stamatopoulos et al. estimated meaningful quantum advantage for option pricing at roughly 7,500 logical qubits. That is not a minor engineering gap; it is a different era of computing. The first places QAE-like methods might matter are not daily listed-vanilla workflows but computationally severe tail cases: high-dimensional exposures, nested simulation, specialized exotics.

C2“Quantum optimization will revolutionize portfolio construction.”

The most common sales pitch, because it sounds intuitive: finance has many optimization problems, quantum devices are introduced as optimization machines, therefore portfolio construction should be a killer app. The argument is rhetorically elegant and technically incomplete. One must separate continuous convex optimization from combinatorial constrained optimization. Mean-variance optimization, risk-budgeting and most institutionally relevant allocation problems are already solved extremely well by classical convex solvers — for these, “quantum optimization” is solving the wrong problem.

The serious territory is the QUBO / Ising class: cardinality constraints, turnover penalties, lot sizes, sector caps, tracking-error objectives with discrete selection, sparse index replication, mixed-integer formulations — exactly what D-Wave and QAOA advocates target. The evidence so far is mixed at best. Portfolio studies on D-Wave annealers show feasibility and occasionally competitive solutions, but not robust superiority; comparative studies in 2024 found classical methods still outperform annealing and hybrid pipelines on most relevant benchmark families once quality, repeatability and end-to-end runtime are evaluated honestly. The gate-based story is more constrained still: QAOA circuit depths for nontrivial instances often exceed coherence windows, and parameter training is nontrivial. A 2024 JPMorgan / AWS decomposition collaboration is the right kind of research — it asks whether large problems can be restructured so a future quantum subroutine solves the most combinatorially expensive component while the classical stack does everything else.

C3“Quantum machine learning will improve alpha generation.”

This is where skepticism should be highest. Quantum ML has produced an enormous academic literature and an even larger marketing halo — the narrative being that quantum feature maps, kernels or variational circuits provide richer representation spaces that unlock superior structure in noisy financial data. The burden of proof should be correspondingly high, because alpha pipelines already sit on mature classical ecosystems: gradient boosting, random forests, factor models, transformers, graph and sequence models, Bayesian state-space systems, large-scale feature engineering on GPU clusters. Any credible quantum-ML claim must beat that stack, not a toy baseline.

So far it has not. There is no established body of evidence showing reproducible outperformance over strong classical baselines on real financial datasets. The reasons are structural: financial data is noisy, weak-signal, regime-dependent and small in true signal-to-noise once leakage is controlled; NISQ hardware adds noise exactly where one needs stable estimation; and the input bottleneck is severe — classical data must be encoded into quantum states, and that loading cost can erase theoretical gains. Even granting that some quantum kernels generate classically hard feature spaces, that does not convert into better out-of-sample Sharpe. Finance is not a contest in abstract expressivity; it is a contest in robust, transaction-cost-aware, adaptive prediction under strict controls.

C4“Quantum will break financial encryption — an existential risk.”

Mathematically the most robust claim, operationally the most misunderstood. Shor's algorithm can factor integers and solve discrete logarithms in polynomial time on a sufficiently powerful fault-tolerant machine, making RSA and elliptic-curve cryptography breakable in principle. The old comfort was timeline distance: Gidney & Ekerå's 2021 estimate put factoring RSA-2048 in ~8 hours at roughly 20 million noisy qubits. Then the estimates tightened. In 2025 Craig Gidney published updated work suggesting RSA-2048 could be factored with fewer than one million noisy qubits in under a week. The threat is not imminent next quarter, but the distance between “theoretically breakable” and “engineering-plausible within a strategic horizon” is shrinking.

More importantly, the relevant risk is not only the day a quantum computer decrypts live traffic — it is harvest now, decrypt later. Adversaries can capture encrypted data today and store it for future decryption. Institutions with long confidentiality horizons are already exposed: regulatory communications, client records, internal architecture documents, credentials, archival messages and research archives may retain value well into the quantum-relevant window. This has crossed from research curiosity into institutional planning. NIST finalized the first three post-quantum standards — FIPS 203, FIPS 204, FIPS 205 — in August 2024, standardizing ML-KEM, ML-DSA and SLH-DSA; the Federal Reserve published a 2026 paper explicitly discussing harvest-now-decrypt-later implications.

Five claims, evaluated against the same scale: theoretical basis, current reality, and the engineering horizon before each becomes a credible budget item.

One row is operational today. The others are research programs, scoring optionality, or category errors — and the only quantum task that belongs in a 2026 budget is the last one.

Three tiers. They are not equally urgent. They should not be equally resourced.

The serious immediate program in Tier 1 includes:

• cryptographic asset inventory across internal systems, client-facing APIs, key management and vendor channels;
• identification of datasets with long confidentiality tails — strategy IP, regulator-facing material, client communications, archived research;
• crypto-agility planning so algorithms can be swapped without rewriting the full platform;
• limited testing of NIST-standardized post-quantum schemes in low-criticality environments to build operational experience.

Among financial institutions, the genuinely serious actors are not hard to identify. Goldman Sachs has consistently published substantive work on quantum algorithms for pricing and risk, including the well-known option-pricing collaboration. JPMorgan has maintained a visible, technically credible effort — optimization work with Quantinuum, portfolio decomposition with AWS. These efforts matter because they build internal expertise, benchmark discipline and institutional memory.

But even the leaders are not “deploying quantum finance” in the popular sense. They are buying optionality — informed, relatively small bets on talent, partnerships, experimentation and IP so they are not structurally late if the hardware crosses a useful threshold. That is different from production adoption. At the other end sit overhyped announcements: generic MOUs, cloud-access partnerships framed as strategic transformation, token innovation narratives with little differentiated research. Many bank quantum teams remain small — often a handful of researchers — and are doing what they should in 2026: exploration, not deployment.

Even serious analysts often assume that once hardware matures, quantum will gradually absorb a wide range of financial workloads. That may be false. There are large parts of the stack where quantum may never be the right tool.

• Low-latency trading — the clearest case. Quantum computers run in cryogenic environments with gate times that make them fundamentally irrelevant for sub-millisecond execution. An FPGA next to an exchange gateway at room temperature wins this permanently; no roadmap changes that economic fact.
• Data-intensive systems — financial computation is often bottlenecked by data movement, normalization, cleaning, reconciliation and state management rather than arithmetic. Quantum algorithms usually assume data is already encoded into a quantum state at acceptable cost; in real systems that assumption is heroic, and the I/O bottleneck can erase the benefit before the algorithm starts.
• General-purpose platform workloads — order management, market-data plumbing, audit trails, surveillance, backtesting, risk aggregation. These remain classical. Even in a mature quantum future, the realistic architecture is a classical platform with rare calls to specialized quantum co-processors for well-defined bottlenecks.

There is also a real quantum winter scenario. If large-scale error correction proves materially harder than current roadmaps imply, useful fault-tolerant computing could be delayed another decade. Microsoft's 2025 Majorana announcement illustrates the point: the field remains capable of spectacular claims subjected to equally spectacular scrutiny. Timeline humility is not optional. It is rational governance.

Nyquist's position is straightforward: build classical infrastructure that is quantum-aware, not quantum-native. That means treating post-quantum cryptography as a present architecture problem rather than a future marketing theme; keeping optimization and analytics layers modular enough that, if a genuine advantage emerges in annealing or hybrid routines, workloads can be routed selectively without rebuilding the platform; and refusing to anchor product strategy to vendor roadmaps that remain highly uncertain.

That posture may sound conservative. It is actually the most aggressive version of realism. The wrong way to prepare for quantum is to pretend it is already here. The right way is to ensure today's infrastructure does not become tomorrow's security liability — and that tomorrow's compute primitives can be integrated without institutional replatforming.

A more honest quantum-finance vocabulary would also retire several phrases immediately: “quantum will soon replace Monte Carlo” (false in any operational sense for 2026); “quantum optimization beats classical portfolio construction” (false as a general statement, usually false even in narrow benchmarks); “quantum ML will unlock alpha” (unsupported by convincing evidence); “quantum is too far away to matter” (false in cyber and cryptographic planning); and “a bank quantum partnership implies strategic lead” (often just signaling). The path from theoretical fit to commercial edge is bottlenecked by hardware, error correction, systems integration and economics — not by ambition.

Is the firm's quantum agenda still sitting in the “interesting but not urgent” bucket — or has post-quantum migration already been elevated to an engineering and risk-management program? If the discussion is serious — around post-quantum cryptography, crypto-agile trading infrastructure and modular compute architecture for future optimization workloads — the conversation is worth having now, before the headlines and the real timelines converge.