AI and quantum computing: today AI pulls quantum forward, and the return trip hasn't started yet

THE THESIS. The convergence between AI and quantum computing is real, but today it runs one way. AI already speeds up quantum: it corrects its errors, calibrates its qubits and designs its experiments. Quantum has not yet sped up AI on any practical problem. The loop will close, but not through the much-hyped 'quantum machine learning'. It will close when fault-tolerant machines, which can correct their own errors while computing, simulate molecules and materials that no supercomputer can handle today. That data will feed the AI models searching for new drugs and materials. On an optimistic reading, that window opens around 2029-2030. Meanwhile, the money is moving much faster than the physics.

WHAT AI ALREADY DOES FOR QUANTUM. A qubit is the basic unit of quantum information. It is extremely fragile: heat, radiation or vibration corrupts it within microseconds. The fix is error correction. The information of one reliable 'logical' qubit is spread across many noisy physical qubits, which are monitored constantly. A 'decoder' reads those signals and works out which error occurred. That is essentially pattern recognition, which is exactly what neural networks do best. Google DeepMind's AlphaQubit appeared in Nature in November 2024. It made 6% fewer errors than tensor-network methods and 30% fewer than correlated matching, the standard technique. Its weakness was speed. A second version, posted as a preprint in December 2025, claims to decode in under one microsecond per cycle. It does so for logical qubits built from up to 241 physical qubits, running on commercial accelerators. That is the speed a real machine needs. Industry is already wiring it in. NVIDIA offers NVQLink, a link between GPUs and quantum processors. It also ships a convolutional neural-network pre-decoder in its CUDA-Q QEC library. Quantinuum says it decoded errors in real time in 67 microseconds, 32 times inside the two-millisecond window its Helios machine requires.

THE 2026 LEAP: AGENTS IN THE LAB. This year's shift is that language models no longer just analyse data: they also run the equipment. In September, OpenAI and MIT published a case study. GPT-5.6 Sol, working through Codex, calibrated a six-qubit superconducting chip. Calibrating means finding its frequencies, tuning the control and readout signals and measuring how long the qubits hold information. According to one secondary report, it completed 40 measurements and researchers stepped in on only four. The company itself acknowledges that performance drops when signals are weak or noisy. A June preprint nicknamed 'Vibe Calibration' goes further. It claims autonomous calibration of a processor with about a hundred qubits. There is no independent review yet. Our reading: this is where the real acceleration lies. Manual calibration takes days of expert time per chip. Automating it lets quantum scale from tens to thousands of qubits without an army of PhDs. A caveat: these results come from single labs or single companies. That is promising capability, not capability proven at scale.

THE HARDWARE: FROM NOISE TO FAULT TOLERANCE. The industry still lives in the NISQ era: machines with hundreds of noisy qubits and no full error correction. The goal is fault tolerance. Google set the first serious milestone in December 2024. Its 105-qubit Willow chip showed that adding qubits to an error-correcting code lowers errors instead of raising them. The fine print matters: it was a 'memory' experiment with a single logical qubit and no operations. In October 2025 Google published 'Quantum Echoes' in Nature. The company calls it the first verifiable quantum advantage: a calculation 13,000 times faster than on the Frontier supercomputer. It tied the result to a molecular magnetic-resonance study with UC Berkeley. On 15 October 2026 Google reported another milestone in Science. Its 'magic state cultivation' technique cut errors 40-fold in special states needed for computation, reaching 99.99% fidelity. According to Scientific American, it used 54 of the 105 qubits and discarded about 92% of attempts. It is a proof of concept, not a factory. IBM promises Starling for 2029: 200 logical qubits able to run 100 million operations. Its Kookaburra module, due in 2026, has not been unveiled yet. The next checkpoint is IBM's conference on 11-13 November. Quantinuum, which uses trapped ions, says it is close to 'five nines' of logical fidelity (99.999%), though experimental details are still missing. Google has also moved into neutral atoms: these arrays already reach about ten thousand qubits, but their cycles are a thousand times slower. Microsoft is the counterpoint. Its Majorana 1 chip, announced in 2025, remains under suspicion. Several physicists argue its data do not demonstrate a topological qubit, and Microsoft already had to retract a similar paper in 2018.

WHAT QUANTUM CAN'T YET DO FOR AI. Quantum machine learning (QML) promises to train models faster or with less data. Recent scientific literature is sober about it. First hurdle: 'barren plateaus'. In many trainable quantum circuits, the gradient, the signal that tells the model how to improve, shrinks to almost zero as the system grows. Without it, the model cannot learn. A December 2025 paper puts it bluntly. When a quantum model suffers from this, a classical program that always answers 'zero' matches it almost surely. Second hurdle: loading classical data into a quantum computer is expensive. That cost can wipe out any speed-up. Third hurdle: 'dequantization'. Again and again, classical algorithms turn up that match quantum ones once billed as superior. A review by Eisert and colleagues (October 2025) concludes that broad, practically meaningful end-to-end advantages remain undemonstrated. Real ML data is noisy and weakly structured. Quantum algorithms, by contrast, shine on highly structured problems. Our reading: quantum's big contribution to AI will not be training chatbots faster. It will be producing data that does not exist today: exact simulations of molecules, catalysts or superconductors. That data will train the AI models that design drugs and batteries. Quantum will be a microscope, not an engine.

THE MONEY IS OUTRUNNING THE PHYSICS. According to McKinsey's Quantum Technology Monitor 2026, investment in quantum start-ups reached about $12.6 billion in 2025. That is roughly six times the 2024 level. 97% was private, and ten deals took 60% of the total. Sector revenue topped $1 billion, against a potential economic value of $1.3-2.7 trillion by 2035. Quantinuum listed on 4 June at $60 a share. It raised about $1.68 billion, the largest IPO in the sector's history. Its quarterly revenue, according to trade press, is around $8 million. IonQ has bought the SkyWater foundry for $1.8 billion. Infleqtion and Pasqal are now public too. Our reading: there is solid science underneath, but valuations price in a timeline nobody has demonstrated. One more data point is worth keeping in mind. In 2025 Google's Craig Gidney estimated that breaking RSA-2048 encryption would take under one million physical qubits running for under a week. His 2019 estimate was 20 million. No machine today comes close to that figure. But the trend is clear: the required resources are falling faster than expected.

IMPLICATIONS. In the short term, there are three real risks. The first is a financial bubble that could hurt retail investors if timelines slip. The second is encryption: data stolen today could be decrypted tomorrow, so migrating to post-quantum cryptography is not optional. The third is talent: AI is automating routine calibration, the entry-level work of many young physicists, in line with our view that AI 'changes which work matters'. In the long term, the horizon is worth it. AI building better quantum computers, and quantum computers giving AI exact chemistry data, form the loop with the most potential for medicine and energy. Drugs designed atom by atom, fertilisers that use less energy, better batteries: that is real material abundance. What to watch: whether IBM unveils Kookaburra in November, whether anyone runs logical operations rather than just memory, and whether each 'quantum advantage' is independently replicated. No euphoria: today the arrow points from AI to quantum. No doom: it is the right arrow to start with.