Photonics Is Gaining Traction — Pushed By AI Infrastructure and Other Factors

It is important to understand where the new interest in photonics comes from. Quite clearly, this is telecom again — but on a different level. While the core topic of the 2000 era was long-haul transmission, the 2025 surge is driven by a much more localized and urgent bottleneck: the I/O wall. And that is mostly due to the enormous expansion of data centers as needed for Artificial Intelligence (AI).

In modern AI clusters, the challenge isn't moving data across continents; it is moving data across a meter of rack space. As we scale to trillion-parameter models, traditional copper-based electrical interconnects are hitting severe physical limits. It is not only bandwidth limitations of copper wires but also the increasing power consumption; we are hitting a “power wall”. At speeds of 200G (200 Gbit/s) per lane and beyond, the energy consumption caused by pushing electrons through copper traces becomes unsustainable. It is interesting to realize that not only computation as such requires power, but also high-speed data transport. There are various physical reasons for that, including the skin effect, dielectric losses and the required digital signal processors (DSPs) and equalizers. Some picojoules per bit may sound like nothing relevant, but for so many bits it really bites.

A passive copper cable does not consume so much for transmission; the problem is the required electronics behind. Fast active copper cables consume on the order of 10 pJ/bit or more due to DSPs and the like. In AI systems, this is similar to the consumption of high-performance GPUs, which are optimized more and more, getting more efficient.

Current pluggable telecom transceivers are actually not better than active copper; they tend to consume even a bit more. This is mostly because the long electrical path from the ASIC on again consumes power. To get significantly better, co-packaged optics (CPO) are a useful approach: Here, the optical engine sits directly on a substrate, eliminating long copper PCB traces. One gets to the region of 5 pJ/bit, for example — a substantial improvement. But the next step brings even more: optical I/O on the chiplet level. Direct-drive photonics may reduce or partly eliminate the need for power-hungry DSP stages and can thus get down to a few pJ/bit or even less than 1 pJ/bit. This can be a game-changer.

The impact of these developments on the photonic markets is not simply demanding more of existing devices. Instead, there is a system-level transition taking place, involving a fundamental re-architecture of the data center, for which new hardware is required. Photonics becomes a core architectural requirement. This has triggered significant shifts in the communications markets:

• Scale-out (between racks): Pluggable 800G and 1.6T telecom transceivers remain the workhorses. However, the market is rapidly consolidating as hyperscalers move toward “Open ecosystems” (like the Open CPX MSA announced in March 2026) to reduce costs and prevent vendor lock-in.
• Scale-up (inside the rack): This is where the real disruption is happening. As copper reach at 224G drops below one meter, photonics is “scaling in”. There are first pilot production deployments of co-packaged optics (CPO) and optical circuit switching (OCS), allowing GPU clusters to behave more like tightly coupled computing systems.

In the 2026 market, power per token has become an essential metric. Companies like Nvidia, Broadcom, and Marvell are no longer just competing on FLOPs; they are competing on electrical/thermal efficiency. Adoption of photonics, substantially reducing the heat load on the switch ASIC, is becoming a prerequisite for the high-density “AI factories” currently being built.

Geopolitics also becomes increasingly relevant. Governments have recognized photonic integrated circuits (PICs) as critical technology where dependencies on unreliable supply chains must be critically examined and worked on. With initiatives like Europe’s PhotonDelta and other “photonic valleys”, we are seeing a “re-shoring” of the supply chain as an important trend in industrial policy. For the first time, market trends are being substantially influenced by national security and supply chain resilience rather than just the lowest cost per bit only.

If photonics is already winning in data transmission, using it for computation as well would potentially eliminate the O-E-O (optical-electronic-optical) conversion bottleneck entirely.

Photonic computing provides essential advantages:

• Massive parallelism: Photonic circuits can process multiple signals of different wavelengths (wavelength division multiplexing) and spatial modes (space division multiplexing). This allows for “Parallel Matrix Multiplication”, the backbone of AI, to happen with extremely high bandwidth and parallelism.
• Passive addition: For example, intensities of optical signals at different wavelengths can add up on a fast photodetector with potentially far lower incremental energy cost than digital electronic accumulation.

Only, photonic computation not exactly easy to implement, considering factors like crosstalk, noise floors and limited dynamic range, apart from fabrication challenges. Right now, this is at the research state, and it is not clear yet which architectures will in the end work best for practical applications. However, I think that sooner or later we will see massive improvements in this field, leading to photonic-based AI with hugely better energy efficiency than current technology.

Consider also biological systems, e.g. insects navigating reasonably well in complicated surroundings. Their miniature brains must perform operations which are not trivial with modern technology — and they work with minimum power. Instead of digital/Boolean logic, they use analog/neuromorphic processing. Biological systems rely on highly distributed, noisy, low-precision analog signaling. Photonics may be well suited for similar approaches.

Biological systems are amazing, but still probably don't constitute the most powerful data processing solution which is physically possible. So maybe in the longer term we can even get better than those. It is hard to tell how that technology will look, but for now, photonics appears to be most promising.

The current upward trend for photonics should not be naively extrapolated. There are actually plenty of warnings signs, indicating a possible market crash in the AI area. At the core, the challenge is that while required hardware investments are extremely costly, it is not clear yet how to monetize the coming results such that these huge investments can be amortized. Competition holds prices down — which is nice for people like me, benefiting a lot from AI which costs a tiny fraction of what it gives us, but puts the AI industry into a difficult position. I find it quite conceivable that investors will at some point become too nervous and trigger a big bust by pulling out. You find some more thoughts in that direction in the last section.

A natural conclusion from this is that photonics companies need to be aware of a possible crash, which could be very destructive for them. It is certainly wise to not only observe developments closely, but also to consider other market opportunities to which one's technology may be oriented as needed.

Quantum computers are transitioning from specialized labs to photonic quantum processors that can operate at room temperature. They might eventually supersede superconducting qubits, which need extreme cooling, and exploit the opportunity of sending qubits through telecom fibers. Another interesting aspect is the potential for compact setups with many qubits. But this technology still has a long way to go.

Massive government investment in quantum communications, involving e.g. quantum key distribution, has already created a secondary spike in photonics interest, as nations race to secure their networks against future decryption threats. Quantum computation may later have even substantially stronger effects.

LIDAR is recognized as essential for autonomous vehicles, and this application area is surely far more important than some safety technologies introduced into traditional cars. Although the technical challenges of autonomous driving are formidable, the economic drivers for that are so strong that these obstacles will sooner or later be overcome with massive investments. And that will have major implications on our transport sector and others, creating opportunities for solving various currently hard problems in transportation. That will be a topic for a future blog article.

Modern agriculture faces severe challenges, e.g. related to pesticides. While serious health effects are discovered (e.g. dementia caused by pesticide exposure), more and more organisms develop resistance to popular pesticides. A promising way out appears to be precision agriculture where e.g. hyperspectral imaging in combination with AI allows accurate problem assessments which can either aid more targeted use of pesticides or enable entirely new methods, such as robotic mechanical measures implemented on small autonomous vehicles. Photonics is again essential in various ways, in particular involving spectroscopy and LIDAR. And the potential market size is huge.

Photonic integrated circuits technology is increasingly being explored for compact medical diagnostics and wearable sensing applications. Spectroscopy-based approaches for noninvasive monitoring remain an active research and development area which may lead to significant market volumes.