AI and quantum computing: the convergence only runs one way (for now)

THESIS: The convergence between artificial intelligence and quantum computing is real and accelerating, but it is deeply asymmetric. Today AI is delivering concrete, measurable gains to quantum —above all in error correction and calibration— while the reverse path, a "quantum AI" that beats the classical kind, still has no demonstrated practical advantage. Anyone selling the loop as symmetric and already operational is confusing aspiration with capability.

Start with the direction that actually works: AI accelerating quantum. The flagship example is AlphaQubit, the error decoder built by Google DeepMind together with Google Quantum AI and published in Nature. It's a transformer neural network trained to spot the "noise" that corrupts quantum computations, and in experiments on the Sycamore processor it made 6% fewer errors than tensor-network methods (the most accurate, but too slow for real-time use) and 30% fewer than the fast reference technique. Its Achilles' heel is still speed: decoding must happen in microseconds, at the pace of superconducting qubits, and that's where the work is now. But the underlying signal is unmistakable: one of the hardest problems in information physics —deciding which error occurred from ambiguous signals— turns out to be exactly the kind of problem deep learning excels at.

And it's not an isolated case. AI is seeping into the entire quantum stack: reinforcement learning to optimize circuits of up to 60 qubits and cut their depth, graph neural networks to design superconducting circuit parameters, and even language agents that autonomously orchestrate the bring-up and calibration of a 112-qubit processor —what its authors half-jokingly dubbed "vibe calibration." Our reading: qubit calibration and control are tedious, extremely high-dimensional, thankless tasks —exactly the work intelligent automation should absorb. That AI is maturing here first, in the lab's back room, is consistent with what we see elsewhere: technology reaches the boring before the spectacular.

The hardware, meanwhile, has taken a leap worth acknowledging without euphoria. Google's Willow chip (105 qubits) demonstrated "below-threshold" error correction for the first time: as you enlarge the code, the logical error rate falls instead of rising, with a 2.14× suppression each time you increase the code distance, and a logical memory that outlives the best physical qubit by 2.4×. It's a milestone 30 years in the making since the 1990s. IBM has published an unusually concrete roadmap toward Starling in 2029 —200 logical qubits, 100 million gates— built on qLDPC codes that, the company says, cut the number of physical qubits needed by up to 90%. Quantinuum boasts 99.9% fidelities on its trapped ions with Helios; QuEra and Quantinuum itself already show dozens of logical qubits with far more efficient encoding ratios than once assumed. This is the end of the "quantum is always twenty years away" narrative.

Now the uncomfortable part, because our line is never to sugarcoat the short term. First, hype with consequences: Microsoft's Majorana 1 chip, unveiled in February 2025 with the claim of eight ultrarobust "topological" qubits, has drawn harsh and sustained criticism from the condensed-matter physics community; Nature's own editorial note warned that the published results do not constitute evidence of topological modes, and the debate was still open in 2026. We cite this as an allegation and a caveat, not a verdict —Microsoft stands by its interpretation— but it illustrates why in this field you must demand peer review and hard benchmarks before buying the headline.

Second, and more important for dismantling the magic loop: quantum applied to AI still isn't delivering. Quantum machine learning (QML) runs into "barren plateaus," where gradients vanish exponentially with system size and training becomes unfeasible. And there's a devastating irony, documented by Los Alamos researchers: many of the architectures designed to dodge those plateaus turn out to be classically simulable —meaning that if the quantum model can be trained, an ordinary computer could probably have done it too. Add the cost of loading classical data into quantum states, which tends to devour any theoretical advantage. As of today there is no demonstrated practical advantage of QML over the best classical methods. The most credible horizon for quantum advantage isn't machine learning but chemistry and materials: simulating molecules with more than ~100 active orbitals, something classical methods can't reach and that will require hundreds or thousands of logical qubits, foreseeably in the 2030s.

The money, of course, is already betting. According to McKinsey's Quantum Technology Monitor, investment in quantum startups reached $12.6 billion in 2025 —6.3 times more than in 2024— with 90% aimed at computing; the firm projects up to $2.7 trillion in economic value by 2035. A telling sign: the share of public capital fell from a third in 2024 to just 3% in 2025, and 60% was concentrated in the ten largest deals (IonQ buying Oxford Ionics for $1.1 billion, PsiQuantum's $1 billion round, Quantinuum nearing a $10 billion valuation). Our reading: private capital smells a "commercial tipping point," but that same concentration and enthusiasm are the perfect breeding ground for inflated claims. The discipline of separating demo from product has never been more necessary.

IMPLICATIONS. In the short term (this decade), the sensible expectation is real but bounded progress: better decoders, more logical qubits, first niche advantages in optimization and chemical simulation, and an urgent need to migrate cryptography to post-quantum standards —the "harvest now, decrypt later" risk is today's, not tomorrow's. None of this justifies either panic or euphoria. In the long term, the convergence does point toward the horizon we champion: if AI keeps greasing the quantum machinery and it matures toward fault tolerance, the prize is accelerating the discovery of drugs and materials —batteries, catalysts, superconductors, molecules against today's untreatable diseases. That's the direct link to abundance and longevity: not an AI that thinks faster thanks to qubits, but tools that compress decades of R&D into years. Targets like DARPA's utility-scale quantum computation by 2033 fit our framework of a hard but well-being-oriented transition.

In short: the AI↔quantum convergence is not a self-propelling rocket; it is, for now, AI pushing quantum uphill while quantum can't yet return the favor. Recognizing that asymmetry isn't pessimism —it's the precondition for investing, regulating and getting excited with a cool head. The horizon of eradicating diseases and freeing up human time remains intact. It's just conquered with verified physics and honest benchmarks, not headlines.