With AI’s energy and compute demands accelerating, Dr Nick New, CEO of Optalysys, argues that photonics could be critical to building a faster, cleaner and more scalable digital backbone.

The combination of aggressive funding and rapid innovation is pushing AI capabilities further and faster than early forecasts anticipated. Since the arrival of ChatGPT in 2022, venture capital investment in AI firms has more than doubled, increasing its share of the total market from 30% in 2022 to over 61% by 2025. Gartner predicts that worldwide AI spending will total $2.5 trillion in 2026.

Despite this financial tailwind, the industry is approaching a physical ‘scaling wall’. Sustaining AI’s growth trajectory will become increasingly difficult as traditional infrastructure comes under pressure from data centre capacity constraints and rising energy demands. Without a fundamental shift in how we process information, these bottlenecks could limit the economic and societal benefits AI can deliver.

Industry leaders are already pivoting to address these limitations. NVIDIA CEO Jensen Huang, known for his calculated, multibillion-dollar strategic moves, has recently turned his attention to silicon photonics. By investing in companies such as Lumentum and Coherent, NVIDIA is betting on a future where data is transmitted via light – photons – rather than electricity.

These strategic investments point to a growing industry view: conventional electronics may lack the energy efficiency and speed required to sustain the next phase of AI growth.

Why the energy wall is arriving faster than expected

Data centres, the infrastructure powering AI growth, already consume nearly 3% of global electricity. As AI models proliferate and require orders of magnitude more compute, global energy demand is set to rise. We’re already seeing the economic fallout of this; notably, OpenAI has cited rising energy bills as a reason for pulling back on its UK expansion plans.

The industry has also hit the physical limits of shrinking silicon – we cannot make computer chips any smaller without compromising reliability. Transistors, which act as the tiny electronic switches that process information inside a chip, are now so small – just a few atoms wide – that it is increasingly difficult to shrink them further and have them operate predictably and precisely. We are left with a phenomenon known as dark silicon: chips packed with billions of transistors that cannot all be switched on at once without overheating or drawing excessive power.

Computing has reached a crossroads. With traditional silicon scaling increasingly constrained, the search for an alternative is now a commercial priority. Jensen Huang’s multibillion-dollar pivot towards photonics serves as a clear signal of where parts of the industry are looking next.

How light could help address scaling limits

Photonics provides a host of benefits for data processing. Light travels faster, carries more information simultaneously and produces less heat than electricity. By integrating photonics directly onto silicon, the scalability of semiconductor manufacturing can be combined with the speed and efficiency of optics.

This results in higher compute density, lower power consumption and improved thermal performance, helping to address the limits imposed by the rise of dark silicon on conventional chips.

Every watt saved at the chip level has a multiplier effect across the world’s digital infrastructure. Using light instead of electrons has the potential to deliver orders-of-magnitude performance gains per watt, reducing power draw and cooling requirements while enabling greater computational throughput. This creates a foundation for more sustainable AI and data processing, one that is faster, cleaner and more scalable.

What a photonics backbone means for AI infrastructure

Training an AI model means coordinating thousands of chips simultaneously, each constantly exchanging data with the others. The bottleneck in large-scale AI is not raw compute, but the energy cost of moving data at the speed and volume modern AI demands. Current computing methods and the available data centre infrastructure are struggling to keep pace with demand. Photonic infrastructure addresses this directly, replacing electrical signals with light across these data pathways. Early industry estimates point to around five times greater power efficiency and ten times higher network resilience compared with conventional electronics.

That said, the benefits of silicon photonics extend beyond sustainability and efficiency gains. The same parallelism that makes photonic systems greener also enables new types of computing once thought impractical, including privacy-preserving computation techniques such as fully homomorphic encryption – processing encrypted data without ever having to decrypt it – which has historically been too resource-intensive to deploy at scale.

Photonics is also uniquely placed to support compute-in-transit: rather than transporting data between memory and processors, transformations are embedded within the data path itself at transceiver line rates.

In photonic systems, information is carried by optical signals rather than electrical currents, enabling high-bandwidth transmission across multiple wavelengths. Optical propagation through a medium supporting interference and diffraction inherently implements linear transformations, allowing operations to be applied directly to signals in transit, without intermediate storage or scheduling.

By processing information optically, we effectively collapse the distinction between data movement and computation. This allows photonic systems to align computational throughput with data bandwidth, bypassing the bottlenecks that currently constrain electronic AI infrastructure.

When compute moves with the data, latency decreases, energy efficiency improves and infrastructure can scale more sustainably.

This has significant implications for sectors such as defence, finance and healthcare, where both performance and data privacy are critical. The same principles also apply to AI acceleration, scientific modelling and edge computing, all of which benefit from high-bandwidth, low-power processing.

The AI industry’s approach to its energy problem has so far focused heavily on incremental improvements, rather than addressing the fundamental infrastructure challenges on which it is built. NVIDIA’s investment is a signal that the hyperscalers recognise this as an infrastructure problem and, crucially, that more of the same may not get them where they need to go.

Photonics presents a new era in which using light instead of electrons could reshape the infrastructure that powers not just AI, but compute more broadly.