Quantum AI Report

The convergence of Quantum with AI

Google DeepMind

Google DeepMind is Alphabet's central AI research organisation, formed in April 2023 by merging Google Brain with the London-based DeepMind lab. Its quantum work sits on the software and theory side rather than hardware: it develops machine-learning methods for quantum error correction, most notably the AlphaQubit neural-network decoder for surface codes, and applies neural networks to simulating and controlling quantum systems. It does not build quantum processors itself, instead publishing research and tools intended to run alongside hardware from Google Quantum AI and other vendors. The organisation is a wholly owned Alphabet subsidiary and is not separately listed on any exchange.

AI-written profile · not yet reviewed · 30 September 2026

Headquarters
London, UK
Status
Private

Coverage

Google Quantum AI & DeepMind Demonstrate Modular Hyperbolic Codes for Over 30x QEC Overhead Reduction

Google Quantum AI, Google DeepMind, and MIT researchers have described a quantum error correction architecture based on modular hyperbolic codes. In a new arXiv manuscript, they report that the approach reduces physical qubit overhead by more than 30 times compared with standard 2D surface code constructions. The result is currently a theoretical proposal with no hardware demonstration reported in the abstract.

OutlookPlausible

Within two years, Google could target a fault-tolerant logical qubit experiment on superconducting hardware with roughly 30 times fewer physical qubits than surface-code estimates would imply, if the modular hyperbolic code's connectivity requirements can be met.

error correctionsuperconductingGoogle DeepMindGoogle Quantum AIMIT
arXiv quant-ph

Foundation Neural-Network Quantum States for Molecular Potential Energy Surfaces in Second Quantization

Researchers have introduced a geometry-conditioned neural-network quantum state for molecular electronic structure formulated in second quantization. The method aims to share wavefunction coefficient representations across molecular geometries, so that a single trained network can describe a potential energy surface rather than being retrained for each nuclear configuration.

OutlookPlausible

Within two years, this approach could be benchmarked against established potential energy surface datasets for small molecules, providing accuracy comparable to multi-reference methods at a fraction of the repeated training cost.

algorithms softwareGoogle DeepMindMicrosoft ResearchQunaSys