AI and quantum computing: the convergence is real — but it runs in the opposite direction to the one you've been sold

THE THESIS. The convergence between AI and quantum computing is real, but asymmetric. Today the actual flow runs in one direction: AI is accelerating the construction of a useful quantum computer. It corrects its errors, calibrates its chips and designs its circuits. The reverse path — quantum boosting AI — remains, on the evidence, aspiration rather than demonstrated capability. Our thesis: this is a flywheel that accelerates itself, but it spins in phases. First, AI builds quantum. Later, once fault-tolerant, quantum will return the favor with chemistry and materials simulations that feed AI itself. Anyone who confuses the order of those phases will be buying hype.

AI IS ALREADY WORKING FOR QUANTUM. A quantum computer's great enemy is noise. Its basic units, qubits, lose their information in fractions of a second. The fix is error correction: bundling many physical qubits into one sturdier 'logical qubit' and watching it constantly. That watching requires 'decoding' — interpreting millions of error signals per second and inferring what went wrong. This is where AI has stepped in. AlphaQubit, the Google DeepMind neural network published in Nature, makes 6% fewer errors than the best classical methods (tensor networks) and 30% fewer than the industry-standard algorithm, measured on real data from the Sycamore processor. Its successor, unveiled in December 2025, already decodes in real time and is orders of magnitude faster on the most promising codes.

AND IT IS NOT JUST GOOGLE. IBM says it has demonstrated real-time decoding of its qLDPC codes — an error-correction family that needs fewer qubits — in under 480 nanoseconds: ten times faster than previous approaches and a full year ahead of its own roadmap. NVIDIA has turned the trend into infrastructure: its NVQLink interconnect plugs quantum processors into GPUs so decoding runs on AI hardware, and Quantinuum already uses it with a decoder that reacts in 67 microseconds, 32 times inside the limit its Helios machine demands. A recent review in Nature Communications confirms the pattern: AI now touches the whole quantum stack, from qubit calibration to circuit and experiment design. This direction of the convergence is not hype: it is published, measured engineering.

WHAT QUANTUM PROMISES AI (AND HAS NOT DELIVERED). 'Quantum machine learning' (QML) — using quantum circuits to learn from data — currently hits three walls. First, 'barren plateaus': as circuits grow, the training signal fades and the model stops learning. Second, a devastating result from Los Alamos: the known QML models that avoid those plateaus turn out to be simulable on a classical computer. In other words: precisely where training works, the quantum advantage vanishes. Third, the data-loading cost: converting classical data (text, images) into quantum states is so expensive that it often eats any downstream gain. The honest conclusion: there is no demonstrated practical quantum advantage in machine learning today. Where there is well-founded hope is in data that is born quantum — molecules, materials — not in replacing today's models.

THE HARDWARE, WITHOUT THE MARKETING. Google's Willow chip (105 physical qubits) demonstrated in 2024 what the field had chased since 1995: making the logical qubit bigger drives the error rate down rather than up — the milestone known as 'below threshold'. In October 2025, Google published in Nature the first 'verifiable quantum advantage': its Quantum Echoes algorithm ran 13,000 times faster than the best classical method on a leading supercomputer, with a result that can be repeated and checked — exactly the criticism that toppled earlier records. IBM keeps the industry's most concrete roadmap: the 120-qubit Nighthawk processor already unveiled, the Kookaburra module with error-corrected memory in 2026, and Starling in 2029 with roughly 200 logical qubits on ~10,000 physical ones, able to run 100 million gates, according to the company. Microsoft is the case to watch closely: its 'topological' Majorana chips promise intrinsically stable qubits, but much of the physics community — as reported by Science and Science News — does not consider it proven that a functional topological qubit even exists, and the company has previously had to withdraw a key Majorana result. Translation: we are still in the NISQ era (noisy, intermediate-scale machines, useful mostly for experiments), but the transition to the fault-tolerant era now has credible dates. And AI is part of the reason.

THE MONEY HAS ALREADY VOTED. Venture capital poured $4.9 billion into quantum in 2025, more than double the previous record, according to PitchBook. In 2026 private funding is cooling — about $1.2 billion so far this year, per Crunchbase — but public markets are taking over: $5.7 billion in liquidity in the first quarter alone, IonQ's $1.08 billion purchase of Oxford Ionics, PsiQuantum valued at $10.5 billion after raising $1.5 billion in May, and Quantinuum heading to market near a $13 billion valuation. The nuance matters: all of this is a tiny fraction of what AI moves, where a single compute contract can exceed a whole year of quantum investment. Quantum is not 'the new AI': it is a decade-long bet. Anyone entering for 18-month returns will be funding the next quantum winter.

OUR READ AND THE HORIZON. The convergence is real, but its engine is the plumbing, not the magic. The pattern we saw with AI agents repeats: value migrates to whoever integrates, and NVIDIA positioning itself as the classical layer of every quantum computer is the most revealing move of the year. Today's asymmetry is also good news in disguise: if AI accelerates error correction, it shortens the path to the fault-tolerant machine — the only quantum computer that truly matters. Short term, we keep our guard up: contested claims that must be verified, oversold QML, and valuations pricing in a perfection that does not exist. Long term, the prize justifies the patience: an AI-corrected quantum computer that simulates nitrogenase — the nitrogen-fixing enzyme that fertilizers depend on — new superconductors, or drugs designed molecule by molecule. That is the full loop: AI builds quantum, quantum generates data about nature that AI cannot obtain any other way, and both push toward the horizon we defend here — eradicating disease and expanding abundance. The flywheel is already spinning. You just have to read which phase we are in.