Photo of the rackable optical accelerator prototype developed at Microsoft Research Cambridge. Image credit: Francesca Parmigiani.
The rapid expansion of artificial intelligence is putting unprecedented pressure on the technology that makes it possible. Behind every AI query lies an increasingly complex network of processors, data connections and enormous datacentres, all of which require significant computing power and energy. But what happens when the hardware underpinning this growth can no longer keep pace?
This challenge was explored by Dr Francesca Parmigiani of Microsoft Research Cambridge in a QAMSS Advanced Technology Lecture, which examined some of the emerging technologies that could shape the future of AI infrastructure.
Beyond conventional computing technologies
One of the central themes of the lecture was the potential of light to overcome some of the limitations of conventional electronic computing. As AI models become larger and more demanding, moving information quickly and efficiently between processors is becoming just as important as the ability to process it.
Optical technologies could offer a way forward. By using light to transmit information, optical interconnects could help connect the vast numbers of processors needed for increasingly powerful AI systems, while reducing the energy costs associated with moving data. The research also explores optical coprocessors, which could take on specific computational tasks and accelerate AI inference without relying entirely on conventional electronic processors.
Beyond the processing hardware itself, Dr Parmigiani also considered how the datacentres of the future might operate. Robotics and physical AI could enable more automated facilities, potentially allowing equipment to be managed and maintained with less human intervention while making better use of available space.
Asked where she sees the greatest opportunities for fundamentally different approaches to computing, she highlighted the need to look beyond established technologies:
"As conventional silicon scaling reaches physical and economic limits, the biggest opportunities lie in rethinking the full computing stack. Optics is particularly promising: it could enable high-bandwidth, energy-efficient communication between AI accelerators and accelerate compute-intensive operations."
From promising ideas to practical technologies
Developing new hardware for AI is not simply a matter of making individual components faster. A technology that performs well in isolation must also work as part of a much larger system, meeting the practical demands of real-world applications.
This is particularly important when hardware and software are evolving together. Advances in AI algorithms can create new requirements for computing infrastructure, while new hardware capabilities can influence how those algorithms are designed and deployed.
Dr Parmigiani explained that Microsoft Research focuses on identifying opportunities where new approaches could deliver meaningful improvements rather than marginal performance gains.
"We start with the bottlenecks of real AI workloads and pursue ideas that could create durable competitive differentiation through significant improvements in performance, energy efficiency, cost or scalability – not simply incremental gains."
She stressed that understanding how technologies work together is essential to making research applicable beyond the laboratory.
"Crucially, we consider the end-to-end system rather than optimizing individual components in isolation. This systems-level approach helps us turn promising proofs of concept into reliable, manufacturable technologies that can operate at datacentre scale."
Ultimately, meeting the demands of future AI will require advances across multiple fields, from photonics and materials to computer engineering, machine learning and robotics. The lecture demonstrated why bringing these disciplines together will be vital to developing computing infrastructure that is not only more powerful, but also more efficient and sustainable.
About the interviewee
Dr Francesca Parmigiani, Microsoft Research Cambridge
Dr Francesca Parmigiani, FOSA, is a principal research manager at Microsoft Research Cambridge, where she leads research in optical computing within the Future AI Infrastructure team. Her work explores how photonic technologies can deliver faster, more energy-efficient AI systems and cloud infrastructure. Before joining Microsoft, Francesca was an academic at the University of Southampton’s Optoelectronics Research Centre, conducting pioneering research in nonlinear optics, optical signal processing and communications. She has published around 300 peer-reviewed papers, holds eight patents and is a Fellow of Optica. Francesca is also an active mentor and advocate for greater equality, diversity and inclusion across photonics and engineering.
By Dr Karen Steward