Lumai Finally Has an Optical AI Box You Can Evaluate. The Big Efficiency Story Is Still on the 2029 Slide.

British startup Lumai, an Oxford spinout that said in 2025 that it had raised more than $10 million, finally has a claim outsiders can test: the company says its Iris Nova server is available now for evaluation and can run Meta's Llama 8B and 70B AI models in real time. In optical computing, the idea that light can do part of AI's math faster and with less power than conventional chips, "available for evaluation" is more interesting than another giant efficiency number. It means customers can at least ask to see the box work.

That does not make Lumai's flashiest numbers real. The same company materials say Nova can cut energy use by up to 90 percent versus conventional GPU systems, and EE Times reported that a later Lumai machine called Tetra is targeted at 100 TOPS per watt and 1 exaOPS within a 10 kilowatt power budget in 2029. Nova is the hardware-now claim. Tetra is the roadmap, even if Lumai's own product page is where the biggest efficiency targets are splashed most clearly.

That split is the whole story. In the new EE Times report, founder Timothee Burr said Lumai encodes inputs into 1,024 lasers, uses lenses to copy that light through a three-dimensional optical path, and handles 2048 by 2048 matrix operations in single operations. Matrix multiplication is the heavy lifting behind large AI models. Lumai's bet is that doing more of it with light, rather than electrons moving through conventional chips, can cut the power bill for inference.

The technical pitch is unusual mostly because it is less exotic than optical-computing startups often sound. EE Times reported in 2024 that Lumai avoids integrated photonics, the chip-scale approach other photonics companies have pursued, and instead uses free-space optics built from standard components. If that holds up in real deployments, it would matter for a dull but decisive reason: standard parts are easier to manufacture than a supply chain full of custom photonic hardware.

The company has also moved its own story forward. In that same December 2024 EE Times profile, Lumai's first planned product was an 8,000 TOPS INT8 PCIe card drawing 500 watts, or about 16 TOPS per watt, with volume production still a couple of years away. Now Lumai is talking about evaluation servers, real-time Llama inference, and test clusters by the end of 2026, according to EE Times. That is real progress. It is also a reminder that startup hardware roadmaps tend to mutate when they meet physics.

The missing piece is the boring one that decides whether a new compute architecture is real. Lumai says more than 90 percent of Llama's workload runs in the optical domain, according to EE Times. It says Nova is available for evaluation now. What it has not published, at least in open materials, is the full benchmark detail customers would need to compare Nova against the GPU servers they already know how to buy: workload setup, latency, number format, baseline hardware, and named users willing to stand behind the results.

For readers trying to keep score, the useful update is not that optical computing will save AI from its power bill. We have heard that sermon for years. The useful update is that Lumai finally says it has an eval box. In this category, that is the line between a physics demo and a product conversation, even if the louder efficiency story is still waiting on proof.