The Qubit Question

Three challenges quantum computing poses for AI — before we get into the detail. First, cryptographic exposure arrives well before useful quantum computing does. Encrypted data — including the training data, model weights and inference logs that sit inside AI pipelines — can be harvested today and decrypted later once quantum machines are capable enough, making AI infrastructure security a live migration problem now rather than a future one. Second, the computational upside for AI is real but narrower than the marketing suggests. Genuine near-term value sits in specific niches — hyperparameter tuning, feature selection, and quantum optimisation subroutines bolted onto classical pipelines — not in a wholesale replacement of classical machine learning. Third, quantum will not be the only challenger: neuromorphic and photonic chips are already shipping commercially, aimed squarely at the energy and bandwidth limits of today's GPU-driven AI stack.

What Quantum Computing Actually Is

Classical computers — including the ones running today's AI models — store and process information as bits that are either 0 or 1. Quantum computers use qubits, which exploit two quantum mechanical properties, superposition and entanglement, to represent and manipulate combinations of states at once. This does not make quantum computers faster at everything. It makes them potentially far faster at a narrow class of problems: searching or optimising across enormous combinatorial spaces, and simulating quantum mechanical systems such as molecules and materials, that classical computers struggle to model efficiently.

IBM's Bet on Quantum as the Next Computing Era

IBM has made quantum computing a headline strategic priority. In June 2026 the company committed more than US$10 billion over five years to the programme, with a stated goal of delivering IBM Quantum Starling — a large-scale, fault-tolerant quantum computer — by 2029, and a successor platform, Blue Jay, targeted for the early 2030s. More than 340 organisations across financial services, healthcare, materials science and government are already using IBM's current-generation quantum systems. The headline point is simply that IBM regards quantum as a genuine, well-funded bet on the next era of computing — not a side project.

Our View: Two Different Machines, Doing Two Different Jobs

Having weighed the evidence, our view is that quantum computing will not replace the classical computing systems that run AI, over any timeframe currently on IBM's roadmap or anyone else's. The two are built for structurally different problems, and the reasons are architectural rather than a matter of maturity that time alone resolves.

Today's AI — large language models, diffusion models, recommendation engines — is built almost entirely on matrix multiplication run at enormous scale across GPUs and TPUs, hardware purpose-built for exactly that kind of parallel arithmetic. Quantum computers are not naturally suited to this task, for two compounding reasons. First, classical data has to be loaded into a quantum system before any calculation can begin — a process called quantum state preparation — and that loading step is itself computationally expensive enough that it can erase the very speed advantage the quantum computer was meant to provide. Second, even where quantum algorithms offer a theoretical edge on the underlying mathematics, that edge is overwhelmed in practice by how slowly each individual quantum operation runs today. One recent academic survey of quantum algorithms against deep learning workloads concluded plainly that a further quantum leap in gate speed — beyond anything Moore's Law-equivalent trends currently promise — would be needed before quantum computing meaningfully affects deep learning over the next decade or two.

That is precisely why the two systems can operate independently rather than in competition. Quantum computing's genuine strengths — combinatorial optimisation, high-dimensional sampling, and simulating molecules and materials at the quantum level — sit in a different part of the problem space to training and running neural networks. IBM's own 2029 target, some 200 logical qubits running 100 million operations, is also simply the wrong kind of scale to run a modern AI model, which routinely involves hundreds of billions of parameters trained on trillions of tokens. There is no credible path on the current roadmap from that milestone to anything resembling a quantum replacement for a GPU cluster.

So What Would AI Look Like Once IBM Delivers on 2030?

On the facts available, our answer is: recognisably similar to today, but with a useful hybrid layer bolted on. The mainstream AI stack — training and running the models that power chatbots, search and enterprise copilots — will still run on classical GPUs and TPUs, because that is what the underlying mathematics requires and because the scale gap between AI workloads and IBM's own qubit targets will not have closed. What changes is at the edges. Specific, well-bounded subproblems inside the AI pipeline — hyperparameter search, certain optimisation and sampling steps, and the scientific simulation work that increasingly sits alongside AI research, such as materials and drug discovery — will begin drawing on quantum co-processors as a specialised accelerator, called on for narrow tasks the way a GPU is called on today for graphics rendering. Today's AI world is not replaced by a quantum one. It gains a quantum wing.

Other Technology Challengers to AI's Current Computing Base

There are other challengers — and arguably these are moving faster than quantum. Neuromorphic computing, chips that mimic the brain's sparse, event-driven signalling rather than the brute-force parallel arithmetic of a GPU, has moved from research to shipping product far more quickly than quantum has. Intel's Loihi 3 and IBM's own NorthPole architecture both reached commercial or full-scale production status in early 2026, reported to be up to 1,000 times more power-efficient than conventional GPUs for real-time robotics and sensory processing tasks — precisely because they sidestep the energy-hungry, brute-force approach that defines today's AI data centres.

Photonic computing is the other credible challenger, using light rather than electrons to perform the matrix multiplication at the heart of neural networks. Photonic processors are reported to run AI inference at meaningfully lower energy cost than a conventional GPU, and a wave of capital has followed that promise — including a landmark acquisition in the space valued at over US$3 billion in the past year. Unlike quantum, photonic and neuromorphic approaches are not chasing a different problem class: they are aimed squarely at today's AI workloads, competing directly with GPUs on efficiency. That makes them, in some respects, a more direct near-term challenge to incumbent AI hardware economics than quantum is.

None of this points to classical computing being displaced wholesale. GPUs and TPUs remain the default for training large models, the software ecosystem built around them over the past decade is a genuine moat, and neuromorphic and photonic hardware today are strongest in narrow, real-time, energy-constrained use cases rather than general-purpose model training. The pattern across all three challengers — quantum, neuromorphic and photonic — is the same: coexistence and specialisation rather than replacement, at least on any timeframe currently visible.