Beyond silicon

A meandering conversation with an AI chatbot feels like an effortless interaction with a listener whose patience appears inexhaustible. For the hardware running in the background, “just one more prompt” means more context held in memory and another round of calculations before replying. You might just be letting off some steam, but somewhere a cooling system is dealing with the heat.

TechUK’s recent survey notes that the annual water consumption for cooling data centres in England is smaller than a day’s leakage from water networks. However, today’s modest AI water demand offers little comfort because present water-resource plans simply don't account for tomorrow’s AI infrastructures, and overlook the reliance of closed-loop cooling systems on adiabatic misting during summer heatwaves.

Keeping chips cool is only part of the battle: interconnect delays and memory bandwidth limits mean that there is only so much we can squeeze out of smaller silicon transistors. While the processors of the future could ease computing bottlenecks, Jevons paradox suggests we will spend the savings on yet more computation!

Constraints on conventional silicon have catalysed frontier computing, which is less a succession contest than a division of labour. Quantum, photonics, and neuromorphic approaches, and new chip architectures, each suit different workloads. The opportunity lies in matching physics to purpose: quantum computing for simulating nature, photonics for optical communication, and neuromorphic computing for event-driven tasks. The prospect is that of integrated hybrid systems in which specialised processors work alongside silicon to address different problems, rather than one chip to solve them all.

While these technologies are progressing at different rates, their shared challenges in characterisation, fabrication and adoption make a case for coordinating their development and interfaces early. Working in silos may well produce excellent components but risks an equally impressive integration challenge later on.

Hybrid computing systems. Source: Government Office for Science (2025)

Smarter chips

Before jumping to exotic hardware, it helps to understand where AI outgrows existing machinery. AI model training is compute-bound, as is prompt processing during language model inference. The decoding side of inference is instead memory-bound, with generating successive tokens being limited by the memory bandwidth available to access model weights and cached attention data. Other neural networks, including robotic vision models face the same tension between doing the sums and getting the numbers.

Dedicated compute units such as NVIDIA’s Tensor Cores increase calculation throughput. However, a faster calculator cannot rescue numbers stuck in transit. High bandwidth memory approaches place stacked memory chips near processors, and transfer data in parallel through wide interfaces. Fractile takes things a step further by physically interleaving memory and compute, thereby efficiently moving data between them.

AI is also lending a hand to the hardware it keeps so busy. Google’s AlphaChip uses reinforcement learning to optimise chip floorplans, while DeepMind’s GNoME has predicted millions of crystal structures, opening avenues for new chip materials. Established materials are also finding new AI uses. For example, gallium nitride can reduce transistor switching losses in power converters supplying AI hardware.

These improvements converge at rack scale, where integration with networking, software, power and cooling can amplify a component’s advantage or swallow it.

Quantum for AI, AI for Quantum

This year, the UK government announced £1 billion to procure commercial scale quantum computers by the early 2030s. Turning that investment into useful capability requires more than a generous supply of noisy qubits. Fault tolerant quantum systems for reliable AI computation require advancements in qubit control, error correction and cryogenic engineering. Meanwhile, classical computers are not retiring; they will still handle compilation, scheduling and measurement results. Quantum testbeds must therefore assess complete hybrid workflows. Where quantum data feeds AI models, clearer classifications are needed to build usable datasets. To this end, the Quantum Standards Network intends to make quantum technologies more comparable and integrable.

While quantum for AI is not quite at maturity, AI for quantum has taken a head start. Google Quantum AI and DeepMind’s AlphaQubit decode quantum errors, while IBM’s AI-powered Qiskit tools optimise quantum circuits. In quantum sensing, AI-assistance has moved tangibly beyond the computer into healthcare, navigation and geophysics. SandboxAQ’s AQNav combines quantum magnetic sensors with AI signal processing for navigation without GPS, while its CardiAQ system enables portable quantum magnetocardiography.

Computers of light

Photonics targets one of AI’s most repetitive demands: matrix multiplications, which optical circuits perform in parallel by encoding values in light and combining them through modulation and interference. Light can handle the arithmetic but needs the electronics for supplying data, controlling modulators and processing results from photodetectors. Converting signals back and forth results in optical losses which can erode the arithmetic gains.

Optical networking connects AI processors at high bandwidth for distributed training and inference, but even AI must accommodate the speed of light: propagation delays matter. Microsoft is deploying hollow core fibres in its Azure Network, guiding light through air rather than glass to reduce latency for processors spread across sites.

Learning from the brain

For AI handling sparse sensor inputs, neuromorphic computing offers a useful economy: not every calculation needs repeating. Spiking artificial neurons are able to retain states and communicate through discrete events, reducing unnecessary computation in robotic perception and control, where event timing is of importance. Intel’s Loihi 2 processors underpin Hala Point, a research system exploring neuromorphic approaches to AI. At implementation level, fabrication differences make compensation learning essential to neuromorphic computing.

Biochips such as Cortical Lab’s CL1 take inspiration from the brain very literally, using cultured neurons as computational substrates, but variability between cultures and maintaining stable activity remain cumbersome challenges.

Convergence for the UK

AI’s convergence with frontier technologies renders interfaces as important as individual capabilities. There is a strong case for considering integration at the outset and throughout scaling. Connecting the AI Hardware Plan, Compute Roadmap, National Quantum Strategy, NeuroSYNC, and the Semiconductor Catapult, offers a route to such coordination. By design, integrated systems cannot develop in isolation, making international collaboration equally necessary. In this context, AI sovereignty is a question of choosing which layers require domestic control while securing resilient supply chains and dependable external access. In this picture, the UK is well placed to capitalise on the convergence opportunity and position itself as a global leader in integrated systems for AI.