Moving numbers is the costly part
Today's AI chips spend much of their energy carrying numbers back and forth between memory and the maths, not on the maths itself.
Corvexsa Research · Lantern
A computer made of light.
Lantern is our long bet: a new kind of AI, and the machine that runs it, built to think with light. Light goes in. Answers come out. What happens in between stays ours.
- Research, not a product
- Started 17 October 2026
- Real-laser test: early November
Why light?
Light does maths in passing
Light can multiply a whole table of numbers at once, just by travelling through the right material. Nothing has to be fetched.
Built for messy physics
Light is never perfectly exact. So we design the AI to expect noise and work with it, instead of fighting it.
How does it work? That stays inside the box.
We don't publish Lantern's design: not the parts, not the materials, not how it learns. This page shows what we can share without giving it away: what it's for, the results so far, and the dates.
The goal
This is the finish line. None of it is reached yet.
- 300×faster learningthan one NVIDIA H200 data-centre GPU.
- 5,000tokens a secondwhen it writes. A token is about ¾ of a word.
- Top tierintelligenceAs capable as today's best AI, above all at code, taste and vision.
- AED 10kfor the whole machineEverything included, about US$2,700.
- AED 200a month to runAt full speed, day and night.
Where we stand: no known design reaches all five today. Physics doesn't forbid it: the energy a light machine needs is tiny. What's missing is at least one breakthrough, and that's what we're hunting.
What we've tested
We write down what counts as a win, and what kills the idea, before a test runs, and we keep the failures. These are the results we can share without giving the design away.
Real light · 17 October 2026
Light did the maths, at home.
Our first test with real light, done at home with no lab. It read handwritten digits it had never seen before.
Toy task
Long chains of reasoning
100%our design
~10%a transformer
Answers that depend on a long chain of earlier steps. Ours got every one right, doing about 60 times less work.
Tiny models · 3 runs
Real English
5.15our design
5.25–5.28a transformer, ⅓ bigger
Loss on real English text, where lower is better. Ours won all three runs. Short training on tiny models, where transformers often catch up later: promising, not proof.
Early model
Surviving damage
95.8%our early model
11.6%a normal network
Handwritten digits read right after 30% of the memory was destroyed on purpose. 11.6% is about the same as guessing.
What didn't work
- Making an AI built from many experts use all of them at once, instead of its best few: 2 to almost 4 times worse.
- Letting a transformer re-read and fix its own draft 12 times instead of 4: no better on long chains.
- Our first goal: a 6-trillion-weight AI, trained in a month, for under AED 5,000. We checked more than 350 ideas and none closed the gap, so the goal changed from how big to how smart.
What's next
- 17 Oct 2026
Day one: the idea, and a first test with real light at home.
- 18–20 Oct
Three rounds of design search: more than 350 ideas checked against real numbers and prices. The goal moved from how big to how smart.
- 21 Oct
The current design takes shape.
- Now
Testing the current design in software, one written-down test at a time.
- Early November
First bench test with a real laser.
- After that
A first version of the machine on the bench, inside the AED 10,000 ceiling.
Questions, or want to help?
Write to us at hello@corvexsa.com.