Photonics vs GPUs: Is Light the Future of Computing?

The AI hardware market is poised for a radical transformation by 2026, where light could become the new computational superstar. This isn't just theoretical or a futuristic research direction; several industry players and research institutions are actively working to shift parts of digital computing from electrical to optical (photonic) processes.

Recently, Jon Peddie Research, a leading analyst in graphics and AI technologies, released a report featuring at least ten companies developing photonic or quantum-photonic processors. These chips use light, rather than traditional electronic transistors, to process data.

This trend is not an isolated experiment. Photonic solutions are backed by industrial investments, startup funding, and major corporate R&D programs. For instance, the American company Neurophos has raised $110 million in Series A funding to bring its photonic AI chips to data center levels, promising performance at the exascale for AI computations.

Traditional processors and GPUs, whether from Intel, AMD, or NVIDIA, use electron movement in transistors for computations. While fast, they face increasing challenges with heat generation, energy efficiency, and scalability, especially when running massive AI models or building large data center infrastructures.

In contrast, photonic chips utilize the wave nature of light to transmit and process information. Light can travel with almost no loss at high bandwidths, enabling:

• Much faster data transfer
• Less heat generation, which is critical for energy consumption
• High parallelism, as multiple wavelengths (colors) can run simultaneously on the same wires

These advantages are particularly appealing for AI-intensive tasks, such as running generative models or inferring large neural networks, where efficiency and speed are crucial.

What signs of change are we seeing?

Photonic technology is spreading beyond research papers to real industrial applications:

• The silicon photonics market is expected to grow dynamically, from about $1.8 billion in 2025 to over $17.8 billion by 2035, with an annual growth rate of around 25%, as more data centers adopt optical data transmission solutions.
• Some photonic chips can already perform AI operations at speeds that, in certain tests, surpass traditional GPUs for specific generative tasks, while significantly lowering energy consumption.
• Meanwhile, major traditional manufacturers are not idle. NVIDIA and its partners are investing billions in photonic and NVLink/optical interconnect developments to integrate light-based data movement and computation in the next generation of AI hardware.

Where are we now, and what's next?

Photonic chips haven't yet replaced classic GPUs across the board; they show significant advantages primarily in specialized, parallel AI tasks and data center network connections. However, experts increasingly talk about a point where these optical solutions could not only complement but also dominate electronic architectures in certain areas.

This doesn't mean GPUs will become obsolete overnight. Instead, we're heading towards a hybrid world where AI hardware combines different physical principles: electronic, photonic, and potentially quantum computing methods in the future.

In the 2026 hardware market, alongside the traditional "stronger chip" competition, a new physical paradigm emerges: where light not only serves for data transmission but becomes a true computational resource. This could be as revolutionary as silicon breaking through the limits of elementary circuits—only now, light carries this legacy forward.

Sources:
nature
jonpeddie
tomshardware
notebookcheck