- Category
- Photonic compute hardware
- Audience
- Foundry partners, photonics engineers, and early workloads
- Stage
- Pre-first-light
A processor with no clock
Exciton Labs is building an AI processor that computes by interference rather than switching. Data is written onto a beam of visible light, and the answer forms while that beam crosses the chip.
Conventional silicon computes by switching transistors, and almost everything expensive about a modern AI datacentre follows from that: the energy each switch costs, the heat it produces, and the plant required to move that heat away. Exciton's argument is that the heavy mathematics of inference does not need to be switched at all. Where waves add and cancel, the result already exists — it is not computed so much as it arrives.
The architecture is described publicly as five layers over one beam. Encode writes data onto the light. Interfere crosses beams so the mathematics resolves in the pattern. Decide is where light meets matter and one beam blocks another, a decision made without ever converting to electricity. Route lets bright beams bend the medium and carve their own channels. Remember freezes the finished pattern into a crystal that holds its state with the power off.
The commercial claim follows from the passive step: a same-workload comparison of $1,260,000 a year on conventional silicon against $11,300 on the Exciton engine — the company's own figure rather than an independent benchmark — on the reasoning that passive computation leaves almost nothing to cool. The company is equally plain about where it is. First light has not happened, every fabrication step the design needs already runs in commercial foundries, and what remains is yield.
What the site had to do
Explain a machine nobody can be shown
There is no chip to photograph and no benchmark to publish. The architecture is disclosed under NDA, so the public page has to convey a genuinely unfamiliar computing model using only what can be said openly.
Make interference legible to a non-physicist
'The answer forms while the beam crosses the chip' is the whole idea and also the hardest sentence in the pitch. A reader has to leave understanding why that is different from a faster transistor.
Carry a hard cost claim without overclaiming
A hundredfold running-cost figure is the most persuasive thing on the page and the easiest thing to disbelieve. It only lands if the mechanism behind it is stated in the same breath.
Be honest about pre-first-light
Frontier hardware sites routinely read as though silicon already exists. Exciton needed to attract foundries and photonics engineers, which means saying what has not happened yet.
How we built it
Lead with the absence, not the feature
'No clock. Just light.' states what the machine does without, which is the fastest way to signal that this is a different category rather than a faster version of the existing one.
Make the hero the mechanism
A live interference field renders behind the headline and responds to pointer movement, so the reader steers a beam and watches a pattern resolve before reading a word about how the chip works.
Structure the architecture as five named layers
Encode, Interfere, Decide, Route, Remember. One verb each, one paragraph each, in pipeline order — enough for a reader to hold the whole machine without any detail that sits behind the NDA.
State the mechanism next to the number
The cost comparison is placed immediately beside the reason for it: the heavy math is passive, so there is almost nothing to cool, and the only wearing part is a field-replaceable front-of-rack module.
Close with the ask, not a waitlist
The final section names exactly what the company is looking for — foundry partners, photonics engineers, and a first workload — against a single contact address and a request line for the technical brief.
The architecture, as disclosed
Encode
Incoming data is written onto the beam itself. From this layer forward, nothing in the pipeline is electronic.
Interfere
Beams cross, and where waves add and cancel the result already exists. Nothing is computed — it arrives.
Decide
Light meets matter and becomes something that can push back, so one beam blocks another. A decision made without ever converting to electricity.
Route
Bright beams bend the medium they travel through and carve their own channels, so the chip reshapes itself around whatever it is currently doing.
Remember
The finished pattern is frozen into a crystal and read back indefinitely — memory that holds its state with the power off.
Passive heavy math
The expensive part of the computation costs almost no power, which is the mechanism behind the running-cost claim.
Parallel by medium
Many streams share one physical path, so parallelism is a property of the optics rather than something a scheduler has to arrange.
Self-routing
The chip rewires itself around the workload it is running, rather than the workload being mapped onto a fixed fabric.
The visual language
Exciton is dressed as instrumentation rather than as a product. A near-black void carries a live interference field, one wavelength-red accent marks every active state, and monospace labels sit in the corners the way readouts sit on the edge of a screen — so the page reads as a machine being observed rather than a thing being sold.
Colors
Void
#08060A
Base field, dark enough for the interference pattern to be the only light source
Bone
#ECE5DF
Display type and body copy, warm rather than pure white
Nanometre Red
#FF3B15
The single accent — the beam itself, and every active state
Graphite
#7A707A
Secondary copy, layer numbers, and status text
Hairline
#ECE5DF24
Section rules at 14% bone, dividing layers without drawing a box
Typography
Archivo
Sets the headline in heavy condensed caps. The type is large enough to sit over the interference field without a scrim, so the pattern reads through the negative space.
IBM Plex Sans
Carries the five layer descriptions and the argument copy, at short measure so each layer is one readable paragraph.
IBM Plex Mono
Layer numbers, property labels, corner status strips, and the NDA line. Monospace marks every piece of text that is an instrument reading rather than prose.
Principles
- 01
Show the mechanism before explaining it
The hero is a live interference field the reader steers with the pointer. Watching a pattern resolve does more to explain computing-by-interference than the paragraph underneath it.
- 02
One accent, because there is one beam
Every active state uses the same wavelength red. A second accent would imply a second signal path, which is exactly the thing this architecture does not have.
- 03
Monospace means measured
Prose is set in the sans; readouts, layer indices, and status lines are monospace, so a claim and an instrument reading never share a typeface.
- 04
Say the stage out loud
The footer carries 'architecture disclosed under NDA' and the closing section says first light has not happened. Withholding detail is fine; implying a finished chip is not.
Questions
An AI processor that computes with visible light. Instead of switching transistors, it computes by interference — the answer forms while the beam crosses the chip, with no clock driving the pipeline.
A conventional processor spends energy switching transistors to produce a result. In an interference architecture, the result exists wherever waves add and cancel. The heavy mathematics is passive, so it is not a faster version of the same operation — it is a different operation.
Encode writes data onto the beam. Interfere crosses beams so the mathematics resolves in the pattern. Decide uses light meeting matter so one beam can block another. Route lets bright beams carve their own channels through the medium. Remember freezes the finished pattern into a crystal that holds state with the power off.
Exciton publishes a same-workload comparison of $1,260,000 a year on conventional silicon against $11,300 a year on its engine. It is the company's own figure rather than an independent benchmark, and the stated reason is that passive computation leaves almost nothing to cool.
No. The company's own framing is that first light is a scheduling problem rather than a miracle: every fabrication step the design requires already runs in commercial foundries today, and what remains is yield.
Foundry partners, photonics engineers, and the first workload worth putting on the machine. The architecture itself is disclosed under NDA, with a technical brief available on request at hello@exciton.org.
Interested in Exciton Labs?
Tell us how you would like to be involved — investing, building, advising, or simply following along — and we will be in touch.
Register Your Interest
Work with Vital Ventures
We study, support, and build the foundational systems that shape how societies function. If you are working on something in that direction, we would like to hear about it.
