When you look at the direction computing is going, the numbers are almost too large to process. Leading AI labs projecting a 1,000x increase in effective compute demand over the next five years. Roughly $100 trillion in infrastructure investment required to meet it with conventional silicon. Around 1,000 GW of additional electrical capacity, enough that hyperscalers are now, in all seriousness, exploring the option of building nuclear power stations next to their data centers.

I presented at the Optica Global Advanced Manufacturing Alliance (GAMA) 2026 in Brussels last week to an audience of photonics and manufacturing leaders. The point I kept returning to was this: the energy constraint cannot be solved by doing more of the same. The compute paradigm itself needs to change.

The paradox at the heart of the problem

We are in the midst of a textbook case of Jevons Paradox. The demand for AI tokens is increasing rapidly, as businesses and the general public get more acquainted with the capabilities on offer. To cope with demand, the compute HW is becoming more efficient, but there is a rate mismatch between the demand growth and the efficiency gains, so overall energy consumption increases despite increasingly better HW. We need to sidestep the curve, to find a technology direction where power consumption per token reduces faster than the increase in token demand. The power reductions offered by better silicon solutions simply cannot scale fast enough.

If silicon is hitting a wall, the corollary is that the architecture of the solution matters as much as the efficiency of individual components. Incremental improvement on the current curve is not enough; a whole new holistic approach is required.

Why now for optical computing

Optical computing has been a research topic for decades and the question I get asked most often when talking about Lumai is: why now?

The answer is urgency and scale. Nuclear power stations next to data centres are not the solution to the problem of energy demand by AI compute. In other words, the gap between what silicon can deliver, what AI workloads require, and what power is available can no longer be bridged by incremental means. A fundamental rethink that simply was not necessary five years ago has now become an imperative.

At Lumai, our answer is the Lumai Iris Server: a fully integrated optical AI accelerator that uses light rather than electricity to perform the matrix multiplications that dominate inference workloads. The key property of optical scaling is that it is qualitatively different from electronic scaling - as matrix size increases, compute grows quadratically while energy grows at most linearly.  

The manufacturing opportunity

Speaking to a room of photonics and manufacturing leaders in Brussels, the message I most wanted to land was the scale of the opportunity. Free-space optics opens a genuinely new direction for how compute is built, and it puts the photonics manufacturing ecosystem at the center of the next era of AI infrastructure.

A concrete example: fiber attachment with active alignment can take two to three orders of magnitude longer than wirebonding or flip-chip bonding. The industry has been able to cope with these timescales, but they are not compatible with the expectations and needs of compute products. However, this is not unique to optical compute: it is one of the central challenges the co-packaged optics community is already tackling, as optics replaces copper for interconnect. The investment and effort being deployed across the industry to resolve this issue (one of the many being addressed) are already significant – automated high-density fiber attach, passive-alignment techniques and standardized optical connectors are all advancing quickly. Optical compute provides an additional impetus to the demand, providing opportunities to speed up the advancements that the packaging and assembly ecosystem requires for optical interconnect. Optical compute is a driver for these advances, while benefiting from the outcomes.

This is the way I framed it for the GAMA audience: an invitation. Optical compute is on course to be one of the largest growth areas in AI infrastructure, and there are real, practical ways for photonics manufacturers to build it alongside us. The companies that advance photonics manufacturing will be foundational to what comes next.

Prefill: the right place to focus

For those less familiar with how large language models run in production, inference is typically split into two phases: prefill, which processes the input prompt and is massively compute-intensive, and decode, which generates the output token by token and is more memory-bound.

Lumai Iris is purpose-built for disaggregated prefill: handling the compute-heavy phase optically, freeing conventional hardware to handle decode. Because prefill carries disproportionate computational weight in the overall inference pipeline, improving its efficiency has an outsized effect on the power consumption of the entire system.

Iris Nova, the first generation of the Lumai Iris Server, is built, validated on Llama 3, and available for evaluation today.

A 20-year roadmap and why I believe it

One of the slides that generated the most discussion was the claim that our architecture is designed to scale for more than 20 years of AI. The healthy skepticism of this statement is welcome, and deserves a more in-depth analysis – here is the reasoning.

As I mentioned, this is not a small inflection on Moore's Law – it is a completely different improvement curve. We are in a similar position that the electronics industry was in roughly 60 years ago at the very beginning of a new development path, not near the end of an existing one. That is what gives us confidence in the long-term trajectory.

The roadmap reflects this progression: Iris Nova uses discrete photonic components; Iris Aura, targeted within approximately two years, moves toward integrated photonic devices; Iris Tetra follows with a more comprehensive and complete solution. Each generation accelerates the trajectory of the underlying technology platform, rather than simply pushing harder on the same one.

Brussels, energy policy, and the European opportunity

Presenting in Brussels was not incidental. Europe's regulatory environment around AI and energy is active, and the business case for sustainable AI infrastructure resonates differently here compared to other markets.

This is a commercial opportunity as much as a compliance story. A 10x reduction in energy per inference versus GPU-only equivalents, deployable in existing air-cooled data center racks without liquid cooling, aligns naturally with the direction of European regulation. That alignment translates into concrete advantages in procurement conversations across the continent and gives us a real edge against competition that is not thinking about this dimension.

Millions of beams of light, performing AI

The image I want to leave people with is this: optical computing, millions of beams of light performing AI operations in parallel, in three dimensions does not just improve on what came before – it jumps over the conventional Moore's Law curve altogether.

The parallelism of light and the geometry of 3D optics make this simple in concept, and massively powerful in practice. We are at the beginning of a new development path. The opportunity this opens is necessary, and substantial.

Iris Nova is available for evaluation now at lumai.ai/eval

‍