Image courtesy 123rf
Finchetto’s Mark Rushworth discusses how, as AI’s insatiable appetite for compute power outpaces infrastructure upgrades, photonics, particularly fully optical network switching, may just move the needle.
The AI industry is at a crossroads. On one side, breakthroughs like DeepSeek’s R1 model, which matched ChatGPT’s performance, at a fraction of the cost, promise a future where AI scales efficiently. On the other, DeepSeek’s chain-of-thought reasoning may offset gains by demanding 87% more energy per query than conventional models. Furthermore, Sam Altman of OpenAI recently admitted that the demand for the company’s ChatGPT 4o image generation tools were ‘melting’ their GPUs, forcing the company to introduce usage limits on the model.
The IEA predicts that data centre electricity consumption will double by 2030, with AI accounting for the lion’s share of the growth. Their April 2025 report paints a stark picture: AI is now the primary driver of electricity demand growth in advanced economies, with data centres set to account for nearly half of demand growth in the U.S. by 2030. McKinsey’s latest projections show that AI workloads alone could require $5.2 trillion in capital investment by 2030, with data centre demand set to nearly triple in that time.
This paradox defines the moment – even as algorithms (seemingly) grow leaner, AI’s insatiable appetite for compute is outpacing infrastructure upgrades.
Where does this leave the industry?
The energy issue won’t be solved by the normal incremental improvements in the energy efficiency of compute devices or by increased model efficiency. An answer to the question of our age lies inside the infrastructure that has made up data centres for decades.
Have you read:
DeepSeek: What tumbling energy stocks say about AI’s power consumption
Rethinking Resilience: Why the energy transition needs smarter, distributed infrastructure
Efficiency alone won’t save us
There is a prevailing myth in AI circles that smarter software will ease the pressure on our energy systems. However, efficiency gains continue to be outpaced by the growth in use cases and model complexity.
Even as compute efficiency improves, it often leads to more usage – a dynamic economists call “Jevons Paradox”. In AI terms, this means lower per-query costs can prompt increased adoption, larger models, and more real-time inferencing. The outcome? The demand curve continues to climb, and so does the electricity bill.
The question of how we make AI more efficient doesn’t cut to the root of the challenge. The more pertinent question is – where in the infrastructure powering AI can we find those efficiency gains?
The network infrastructure bottleneck
While much of the conversation surrounding AI’s rising energy use has been on compute – accelerators, GPUs, and training costs – one of the most significant sources of energy consumption, and a growing bottleneck, is the network switch.
To move data between compute nodes, data centre networks rely on switches which act like traffic controllers, deciding where each piece of data (or ‘packet’) goes and how it gets there. This networking accounts for 10% of a state-of-the-art Nvidia data centre’s total power use, and that’s before factoring in the additional power needed for cooling required to manage heat from high-power switches, which can then account for a further 10% of total energy spent.
AI workloads are highly dynamic, data-heavy, and often unpredictable. Inference, the stage where a model responds to new data by making predictions or generating content, requires flexible resource allocation and extremely low latency to perform effectively. As model sizes and inference complexity increase, so does the bandwidth demand on networks. Latency becomes not just a performance issue, but an energy one. When network delays idle processors, even briefly, they consume power without doing work.
Many of today’s networks rely on electro-optical switches, which convert signals from light to electricity, perform switching in the electronic domain, then convert the signal back to light for onward transmission. This optoelectronic process incurs significant energy and latency penalties. At scale, these delays multiply across thousands of nodes and billions of data packets.
With this consideration, it’s evident that the primary challenge to address is the elimination of the necessity for optoelectronic conversion.
How photonics can move the needle
One of the most promising breakthroughs in solving these challenges is through photonics, particularly in fully optical network switching. This approach routes data entirely in the optical domain, eliminating the need for optoelectronic conversion. The long-standing challenge, however, has been finding a way to direct packets of light to the intended destination without relying on electronics.
An answer lies in the properties of nonlinear photonic crystals – materials already familiar to the quantum computing world. These materials can be ‘tuned’ to interact with specific wavelengths of light. By encoding packet routing information onto a parallel control beam of light with a unique wavelength and combining it with the data signal inside a nonlinear crystal, it’s possible to imprint the destination address directly onto the data stream.
This process allows switching to take place entirely within the optical domain – light controlling light. The result is a passive, energy-efficient switch that operates at the speed of the transceivers and eliminates the major performance bottlenecks associated with today’s network architectures. This kind of switch also opens new possibilities for distributed AI computation. High-radix, low-diameter topologies like Intel’s PolarFly – which reduce the number of hops between nodes – become not only possible, but optimal.
These performance benefits unlock substantial returns by increasing the utilisation of compute resources. In today’s data centres, a significant portion of high-performance processors sit idle, waiting for data to traverse congested, latency-prone networks. Thanks to fully optical networking, latency can be cut down to nanoseconds, significantly reducing this idle time. The result means the complex computational tasks demanded by AI – from natural language processing to complex data analysis – can be completed much faster, and at high throughput rates.
With more compute being utilised, more efficiently, it also means fewer servers are needed to deliver the same performance, cutting both capital and operational costs. In large-scale deployments, this combination of improved efficiency, reduced overprovisioning, and lower cooling and power demands adds up to a meaningful ROI.
The optical options
We’ve seen the early-stage adoption of optical circuit switches in some hyperscale and research environments, namely from Google. This technology is offering the benefits of optical networking: lower latency, more efficient data transmission and compute utilisation. However, despite the circuit switching process being best suited for direct and dedicated end-to-end data transmission, including synchronous model training and bulk data transfers, it is not as effective in dealing with other types of AI workloads.
AI workloads for inference services and cloud clusters demand vast amounts of data to be moved between nodes quickly, and this is where packet switching plays a crucial role. Packet switching breaks data into small packets that travel independently across the network, dynamically finding the most efficient routes to their destinations, where it is then reassembled. This makes it ideal for the most pressing demands coming from AI.
We are now seeing fully optical switches emerge that support both circuit and packet switching, enabling data to move at the speed of the surrounding transceivers – the speed of light – and marking a breakthrough in networking architecture. These switches are no longer constrained by the trade-offs of earlier technologies: the processor-limited routing speeds of electronic switches or the millisecond-scale reconfiguration delays of optical circuit switches, which cause latency and restrict them to static or long-lived connections. Operating entirely within the optical domain, this new class of switch operates without switching delay, delivering ultra-low latency, high throughput, and massive bandwidth, while offering the flexibility required to support diverse, dynamic AI workloads without compromising performance.
Powering AI from first principles
If we want to ask more of our AI systems and strive for the goals promised by its evangelists, we must ask first of ourselves: where does that energy come from, and how efficiently are we using it?
Any increase in grid capacity must be twinned with making data centres more energy-efficient and sustainable. Where technologies already deliver equivalent or superior performance at a fraction of the power draw, like fully optical network switches, they must be championed by industry and underpinned by government policy.
Hyperscalers and data centre operators should take a phased approach to adopting more sustainable technologies, starting with R&D assessments, progressing through incremental adoption, and designing hybrid systems with long-term scalability in mind. The aim is to ensure infrastructure can handle the rising scale and intensity of AI workloads. By connecting innovation teams with emerging technology providers, operators can explore real-world performance and compatibility via proof-of-concept trials.
These early-stage collaborations could help shape future infrastructure strategies, where energy and space efficiency can be embedded from the ground up. Photonics presents that opportunity: to dramatically reduce energy use in AI infrastructure and strengthen data centres’ ability to handle the demands of AI.
About the author
Mark Rushworth is the CEO and co-founder of photonics startup, Finchetto. Mark holds certificates in integrated photonics chip design and is a member of the UK Government’s Optical Communications and Photonics Expert Working Group.
