Can light solve AI’s energy bottleneck?

Artificial intelligence (AI) infrastructure investment has gone through the roof since the arrival of generative AI language models such as ChatGPT in 2022. With hyperscalers driving investment in AI infrastructure to record highs in 2025, International Data Corp. tallies global investment at a staggering $318B, and it appears poised to continue rising year after year.

Against a backdrop of skyrocketing capital injections, the industry is approaching a physical “scaling wall.” Traditional infrastructure is beginning to creak under the load of data center capacity constraints and surging energy demands.

The increasing energy requirements to power AI is unsustainable, with concerns also growing about the environmental damage that supplying such energy could cause.

We are now at an inflection point. Traditional electrical data transfer methods are reaching their limits, with NVIDIA showing its hand by recently investing $4 billion in two photonics companies, Coherent Corp. and Lumentum. NVIDIA is betting on a future where data is transmitted via light (photons) rather than electricity.

AI energy consumption

The Electric Power Research Institute estimates that data centers could consume up to 9% of U.S. electricity generation annually by 2030, up from 4% in 2023. As AI models face ever-increasing consumer demand and require more compute, we’re going to see global energy demand rise. This poses an acute problem for the costs of scaling AI services, given recent volatility in energy prices. We’re already seeing this problem materialize, with OpenAI citing rising energy bills as a reason to roll back its U.K. expansion plans.

Processors are approaching physical limits. Transistors, the electronic switches that form the basis of electronic circuits, are now just a few atoms wide—a size at which quantum effects and heat become significant limitations.

Light at the end of the tunnel

Beyond the challenge of the amounts of energy being used to process and move data, the physical distance between the processing and memory elements, both on-chip and at system level, now limits the speed at which AI models can be run and trained. Building data centers on a photonic foundation is the next logical step.

Soon, computation within the optical data path will be possible and it offers the opportunity to reduce latency and scale infrastructure without proportional increases in power consumption.

Photonics can be integrated directly onto silicon chips to enable scalability and efficiency improvements over electricity. The crux of photonics’ efficiency gain is simple: Light travels faster and carries more information, while producing less heat than electrons. This results in dramatically higher compute density, lower power consumption, and superior thermal performance to overcome the limits imposed by the rise of dark silicon on conventional chips.

The benefits of improving efficiency at the chip level are evident in how quickly energy savings compound. One watt of energy saved by powering a chip also reduces the energy wasted on power draw and cooling. Photonics opens up a future of developing AI infrastructure, centered on a foundation that is faster, cleaner, and fundamentally scalable.

Photonics data center implementation

The fundamental bottleneck in large-scale AI is no longer raw compute but the staggering energy cost of moving data at the speed and volume modern AI workloads demand. The rapid evolution of frontier models means systems are under constant strain from coordinating thousands of chips simultaneously. Traditional data center infrastructure simply can’t keep up with demand for constant, highly intensive data exchange.

Photonics presents an opportunity to address this problem at a strategic level, rather than merely mitigating the spiraling thermal demands of an increasingly stretched electrical architecture. Early industry estimates indicate using light to transfer data delivers roughly 5x the power efficiency and 10x the network resilience of conventional electronics.

The benefits of silicon photonics extend beyond immediate efficiency and sustainability gains. By eliminating significant data-transfer bottlenecks, photonics also unlocks types of computing that were previously deemed impractical due to energy costs, such as fully homomorphic encryption (processing encrypted data without ever decrypting it).

Through eliminating the limits of traditional compute architectures, photonics has far-reaching implications for sectors where the performance and data privacy required by the defense, finance, and healthcare industries are non-negotiable.

To date, the AI industry’s response to its escalating energy needs has been slow and is failing to address the structural flaws of traditional silicon architectures. NVIDIA’s recent multibillion-dollar investments serve as a clear signal that hyperscalers now recognize this is fundamentally an infrastructure problem.

We must now ask ourselves whether we do more of the same—continue ploughing eye-watering amounts into data center buildout and cooling infrastructure, or invest in innovative solutions such as photonics, which can solve key limitations at source.

Photonics represents a new horizon of opportunity. Rather than replacing existing systems, photonics augments modern compute architectures by unlocking new compute capacity within the network itself. Photonics is driving the chip industry toward its biggest architectural shift since the advent of the von Neumann architecture, and it presents an opportunity to unlock limitless compute.