Quantum AI Research

Quantum computing in 2026 is in the regime John Preskill named NISQ — noisy, intermediate-scale, with no clean separation yet between demonstration and useful application. We work this regime with discipline. The hard question is not whether quantum hardware is improving — it is — but which problems, on which timeline, repay the integration cost. The technical questions split into four. The first is whether variational and hybrid quantum-classical algorithms can produce useful results inside the noise-and-depth constraints of current hardware, despite the barren-plateau and noise-accumulation pathologies that limit them. The second is whether quantum error correction has crossed the threshold at which adding more physical qubits reliably reduces logical error rates, and at what overhead. The third is whether quantum-enhanced sampling and quantum chemistry produce useful advantages on problems that survive sustained classical-baseline pressure, rather than on contrived problems where the comparison is biased toward the quantum side. The fourth is the institutional question: whether the field’s claims of quantum advantage are evaluated against the best available classical methods, with the methodological honesty that the comparison being made is the relevant one. We invest in the variational and hybrid-algorithm side, in the error-correction substrate that will determine how quickly the picture changes, and in the small set of learning-system problems where quantum methods are plausibly competitive within the decade.

The four questions are different

Preskill’s 2018 framing of the NISQ era¹ remains the most honest summary of where quantum computing sits relative to its promise: hardware that is interesting enough to study but not yet capable of cleanly outperforming classical methods on problems anyone cares about, with error rates that constrain useful circuit depth to tens or low hundreds of gates. The variational algorithm family — VQE, QAOA, and their successors — was developed to fit inside this constraint by offloading much of the computation to a classical optimizer; Cerezo and colleagues 2021 is the current review of record.² The promise of these methods is real but is bounded by the barren-plateau phenomenon, which limits the trainability of variational circuits at scale; McClean, Boixo, Smelyanskiy, Babbush, Neven 2018 is the foundational reference.³

The error-correction picture has changed substantially in the last two years. Google’s surface-code work — Acharya and colleagues 2024 — demonstrated that increasing the code distance reduces logical error rates below the physical error rate, the threshold above which scaling is the bottleneck rather than the device physics.⁴ The bottleneck is now overhead: producing logical qubits with low enough error rates for useful algorithms requires thousands of physical qubits per logical qubit at current physical error rates. Project Q-Core is our internal investment in raising the operating temperature of quantum error correction, which would change the cost structure of large-scale fault-tolerant systems if it works.

The relevance to AI is narrower than the term “quantum AI” sometimes suggests. The HHL algorithm for linear systems — Harrow, Hassidim, Lloyd 2009⁵ — is asymptotically exponentially faster than classical methods on certain problems, but the assumptions required (quantum-accessible data, well-conditioned matrices, sparsity) limit its practical scope. AlphaTensor demonstrated that reinforcement learning could discover novel matrix-multiplication algorithms — Fawzi and colleagues 2022⁶ — and analogous techniques for quantum-circuit discovery are an active area. Quantum-enhanced sampling is the area where the case for near-term advantage is strongest, but the workloads on which it is competitive are narrow.

A discipline of the field, which we adopt: claims of quantum advantage are evaluated against the best available classical method, not against an unconditional asymptotic separation. Several published advantage claims have been overturned within a year by improvements in the classical baseline. This is not a failure of quantum hardware; it is a feature of the comparison being made. The serious near-term question is which problems retain a quantum advantage under sustained classical pressure, on the empirical hardware that can be built within a five-to-ten-year horizon, with honest accounting for the data-loading and read-out overheads that pure-circuit comparisons sometimes elide.

The reason this work matters: the timeline on which quantum methods become useful for problems we care about is uncertain by a decade or more, but the integration cost on the day they do is high. We invest at a level proportionate to that uncertainty, with a strong preference for the substrate work where the option value is clearest.

What the quantum-AI program is, technically

We organize this work along four sub-strands.

Variational algorithms

Variational quantum algorithms — VQE for ground-state estimation, QAOA for combinatorial optimization, and their successors — are the dominant near-term paradigm because they tolerate noise by construction. Cerezo et al.’s review² is the current reference. Our internal work focuses on ansatz design that mitigates the barren-plateau pathology,³ on the integration of variational algorithms with classical optimizers that exploit the structure of quantum gradients, and on benchmarking against the best available classical methods — the latter being the discipline that distinguishes serious quantum-algorithm work from demonstration-class results.

The discipline points include explicit-classical-baseline characterization (so that the quantum advantage claim is evaluated against the best available classical method rather than against a strawman), explicit-noise-budget analysis (so that the algorithm’s noise tolerance is characterized rather than assumed), and a preference for ansatz architectures whose trainability properties are analytically characterizable rather than only empirically observable.

Quantum-enhanced sampling

Sampling from distributions that are hard to sample classically is the area where near-term quantum advantage is most plausible. The application surface includes Boltzmann machines, certain classes of generative models, and Monte Carlo methods for combinatorial optimization. The technical question is whether the samples produced by a noisy quantum device retain enough fidelity to the target distribution to be useful, and whether the classical sampling baseline can be beaten on problems of practical size.

We work on this question empirically rather than asymptotically: the asymptotic separation results are well-established, and the practical question is now the bottleneck. Several recent advantage claims in the sampling regime have survived improved classical baselines; several have not. The discipline is to run the comparison fairly and to report the results regardless of which side wins.

Hybrid classical-quantum loops

The deployment unit for near-term quantum is a hybrid system: classical compute that issues quantum sub-routines, receives results, and adapts. The scheduling, latency, and data-movement problems in such systems are non-trivial, and they intersect the Cognitive Computing work on heterogeneous-architecture inference. We work on the systems-engineering side of hybrid loops — how to integrate quantum sub-routines with classical inference pipelines without the round-trip latency dominating the runtime — and on the algorithmic side, where the choice of which sub-routine to send to the quantum backend is itself an optimization problem.

The discipline points include explicit-latency-budget characterization across the hybrid loop, explicit-bottleneck identification (so that the quantum-versus-classical decision is informed by the actual bottleneck rather than by aesthetic preference), and a preference for hybrid architectures whose performance properties are analytically characterizable.

NISQ-era applications

The set of problems on which quantum hardware is plausibly useful within the NISQ regime is small. Quantum chemistry is the most often cited; certain combinatorial-optimization problems are candidates; sampling problems are the third major category. Our internal work focuses on the application-area selection question — which problems repay the integration cost on which timeline — and on the small set of learning-system problems where quantum methods may be competitive.

The connection to Economic Orchestration is direct: certain combinatorial-auction and matching problems are candidate workloads. The connection to Project Synthesis is also direct: closed-loop materials discovery includes quantum-chemical sub-problems whose accurate simulation is one of the most credible near-term application areas, and the substrate question — what materials enable better quantum hardware — is itself a target for the same loop. The discipline points include explicit application-area characterization (so that the application’s quantum-advantage profile is verified rather than assumed), and engagement with the broader application-area communities (chemistry, materials science, optimization) on the methodological questions.

Definitional bounds

Before moving to the open problems, four exclusions are worth being explicit about.

Quantum AI does not mean exponential speedups on arbitrary problems. The asymptotic separation results require specific problem structures and specific data-access assumptions. The popular framings of “quantum supremacy on machine learning” are not the program’s research substrate; the program’s framing is on the small set of problems where quantum advantage is plausible under realistic assumptions.

Quantum AI does not mean quantum machine learning replaces classical machine learning. The program is on quantum-enhanced sub-routines within classical learning pipelines, not on quantum replacements for classical learning. The hybrid-classical-quantum architecture is the deployable substrate.

Quantum AI does not mean quantum advantage on contrived problems. The methodological discipline of evaluating quantum advantage against the best available classical method on problems of practical interest is load-bearing. Demonstrations on contrived problems where the classical baseline is artificially weak are research debt rather than research result.

Quantum AI does not mean near-term deployment. The timeline on which quantum methods become useful for problems we care about is uncertain by a decade or more. The program invests at a level proportionate to that uncertainty, with a strong preference for substrate work where the option value is clearest. The framing is long-arc rather than near-term.

Open problems

1. Noise-resilient algorithms for NISQ. Variational algorithms tolerate noise but pay for it in trainability. The frontier of noise-resilient algorithms that remain trainable at useful scale is moving, but slowly relative to hardware improvements.³
2. Demonstrating advantage on practical problems. Quantum advantage demonstrations on contrived sampling problems do not establish advantage on problems anyone wants to solve. The methodological problem of constructing fair comparisons against the best classical methods on practical problems is open.
3. Hybrid scheduling. When to dispatch a sub-problem to a quantum backend versus solve it classically, given the round-trip latency and the noise budget, is an optimization problem we do not yet have a clean theory for.
4. Decoder-latency bounds. Real-time error correction requires syndrome decoders that operate within the coherence-time budget. Recent surface-code decoders are competitive,⁴ but the latency budget tightens as code distance grows.
5. The classical-quantum interface at scale. A fault-tolerant quantum computer is not a black box. The data-movement, calibration, and orchestration problems at the interface between classical compute and large-scale fault-tolerant quantum compute are open.
6. Higher-temperature error correction. Operating quantum error correction above millikelvin temperatures would change the cost structure of large-scale fault-tolerant systems substantially. This is the central question of Project Q-Core.
7. Quantum-circuit discovery via reinforcement learning. AlphaTensor demonstrated that RL can discover novel matrix-multiplication algorithms;⁶ the analogous quantum-circuit-discovery question is open.
8. Application-area-specific quantum advantage characterization. Chemistry, optimization, and sampling are the candidate application areas; the application-specific advantage characterization is open and in active research.

Three risk scenarios

Scenario A — Classical-baseline catches up

The first failure mode is the classical-baseline-catches-up scenario. Quantum advantage claims are made against weak classical baselines; the classical methods improve in response; the quantum advantage erodes; the quantum-AI investment thesis becomes harder to defend. The mitigation is the methodological discipline of evaluating quantum advantage against the best available classical method, with explicit characterization of the comparison conditions.

Scenario B — Error-correction overhead remains prohibitive

The second failure mode is the error-correction-overhead-remains-prohibitive scenario. The threshold-crossing has been demonstrated; the overhead remains at thousands of physical qubits per logical qubit; the deployable fault-tolerant systems remain a decade or more away; the near-term quantum-advantage thesis depends on NISQ algorithms that the barren-plateau and noise-accumulation pathologies limit. The mitigation is investment in the substrate (Project Q-Core’s higher-temperature work, broader hardware advances) and continued investment in NISQ-class algorithms with clear-eyed acceptance of the limitations.

Scenario C — Successful staged deployment in narrow application areas

The third scenario, which we treat as the base case if the substrate and algorithm work are competent, is staged deployment in which quantum methods produce clear advantages in narrow application areas (chemistry simulations, certain combinatorial optimization problems, sampling for specific generative-model classes), the application-area-specific advantages compound into a broader quantum-method-deployment substrate over a decade-plus timeline, and the substrate-investment-cost is justified by the long-arc option value rather than by near-term returns.

What technical work bears on this

Project Q-Core is our internal investment in higher-temperature quantum error correction, and is the most direct application of the work described here. This pillar connects to Cognitive Computing on the architectural-alternatives side: both pillars investigate substrates that depart from conventional digital CMOS. It connects to Economic Orchestration on the application side, where certain large-scale combinatorial problems are candidate workloads for quantum-enhanced optimization. The honesty-about-uncertainty stance that this pillar requires is shared with the broader AI Safety program. The connection to Project Synthesis is direct: closed-loop materials discovery includes quantum-chemical sub-problems whose accurate simulation is one of the most credible near-term quantum application areas.

Where to read further

Project Q-Core treats the higher-temperature error-correction substrate. Cognitive Computing treats the complementary architectural-alternatives bet. Economic Orchestration treats the combinatorial-optimization application surface. Project Synthesis treats the closed-loop materials-discovery infrastructure that intersects the quantum-chemistry application area.