Optical compute now coming to a data center

Optical computing has long been billed as ten years away. This timeline shifted in April when we launched Iris Nova. Now the question is whether optical compute can hold up in a third-party data center, under independent testing.

Today, we begin to answer this question. ARIA, the UK’s Advanced Research and Invention Agency, will deploy Lumai Iris servers in its Scaling Inference Lab, run by CommonAI CIC. Lumai was selected through a competitive technical evaluation as the lab’s first optical inference platform. An eight-node system will run production-scale models in a commercial data center environment.

What matters most to us is who will run the system and who will be doing the testing. This isn't a demo we stage and control. Once the system is operational, external teams will be able to bring their own workloads and see for themselves what optical compute can do.

Optical computing has been tried before. Here is what's different

Optical computing is not a new idea, however earlier efforts attempted to replace the electronic computer outright; building optical memory, optical logic and optical nonlinearity. Those projects were also trying to race against silicon at a time when its efficiency doubled every couple of years.

Three things have changed:

1. AI inference computation is dominated by one operation. Large matrix multiplication accounts for around 90% of inference compute, and optical systems perform it with exceptional efficiency.

2. Silicon efficiency gains have slowed. Power has become the binding constraint on data center growth.

3. The optical supply chain has matured. Optical components are now manufactured at high volume for data center communications.

We aren’t building an all-optical computer. Lumai Iris is hybrid: light performs the matrix multiplications, and established electronics handles memory, nonlinear functions and software.  

We compute in three dimensions using free-space optics, rather than on a two-dimensional photonic chip. On a chip large matrix computation need more components than will fit or cascaded stages that accumulate noise. Lumai Iris uses a large matrix operation which completes in a single optical cycle.  

Power is now the limit to AI deployment

Power now limits how fast AI can be deployed. If the compute ambition of even one leading AI lab were met with conventional accelerators alone, estimates put the infrastructure bill at around $100 trillion within five years, with more than 1,000 GW of additional power. Building faster won't close this gap - it requires a change in the physics of computation.

Optics offers a solution. As matrix size grows, optical compute performance grows quadratically while energy usage grows, at most, linearly. This efficiency relationship is governed by physics, not by process node, so the advantage doesn't fade as GPUs improve. And because Iris is hybrid, every gain in silicon improves its electronic components too.

Built for prefill, alongside the GPU

We are intentional about where light delivers the biggest gain. Large-model inference runs in two phases. Decode generates one token at a time and is limited by memory bandwidth, so faster arithmetic helps little. Prefill processes the whole prompt in one parallel pass against fixed model weights, reusing each weight across every token. It is compute-bound, and that is exactly the profile our optical system is built for.

Serving frameworks such as vLLM, SGLang and llm-d already route prefill and decode to separate hardware pools. Iris aligns with that architecture: optical compute for prefill, alongside a silicon-based accelerator solution for decode. Paired this way, Iris will deliver at least 16x better energy efficiency per inference than silicon-only systems. Our roadmap targets up to 90% lower power and around one-tenth the total cost of ownership of conventional GPU architectures.

For developers, none of this requires a new programming model. Iris integrates through PyTorch and behaves like a conventional processor, offloading matrix multiplications to its optical core.

Why a third-party data center is the real test

Each new compute architecture faces the same question: does it work outside the lab that built it? We will answer this question by running a multi-node system in a commercial data center. That means real workloads, standard power and cooling, real rack integration, and independent engineers who didn't design the system measuring what it can do.

In April we launched Iris Nova, the world’s only optical AI accelerator engineered for data center-scale inference. It runs production-scale LLMs today in standard air-cooled racks. The Scaling Inference Lab is the next step: a multi-node deployment, operated by a third party and open to outside teams. Evidence earned outside our own lab is the proof optical compute has been waiting for.

If you want to test your own workloads on Iris, we’d love to hear from you. The Scaling Inference Lab will open access once the system is operational.

Frequently Asked Questions

Why didn’t earlier optical computing efforts succeed?

Earlier projects tried to replace the electronic computer outright, building optical memory, logic and nonlinearity, while racing silicon at a time when its efficiency doubled every couple of years. Lumai Iris takes a hybrid approach instead: light performs the large matrix multiplications that make up around 90% of AI inference compute, and mature electronics handle memory, nonlinear functions and software.

Why does Lumai use 3D free-space optics instead of a photonic chip?

On a two-dimensional photonic chip, large matrices need more components than fit on the chip, or cascaded stages that accumulate noise. Lumai computes in three dimensions using free-space optics, so a large matrix operation completes in a single optical cycle. As matrix size grows, optical compute performance grows quadratically while energy grows at most linearly.

Why is Lumai Iris built for prefill rather than decode?

Prefill processes the whole prompt in one parallel pass against fixed model weights, making it compute-bound, while decode is limited by memory bandwidth. Lumai Iris brings optical compute to prefill and pairs with silicon accelerators for decode, matching how serving frameworks such as vLLM, SGLang and llm-d already split the two phases. Paired this way, Iris will deliver at least 16x better energy efficiency per inference than silicon-only systems.

Do developers need a new programming model to use Lumai Iris?

No. Lumai Iris integrates through PyTorch and behaves like a conventional processor, offloading matrix multiplications to its optical core.

Why does testing in a third-party data center matter?

Every new compute architecture has to prove it works outside the lab that built it. At ARIA’s Scaling Inference Lab, a multi-node Lumai Iris system runs real workloads with standard power, cooling and rack integration, measured by engineers who didn’t design it. Once operational, the lab will open access to outside teams.

Want to test your own workloads on Lumai Iris? Contact Lumai.

‍