PHASER

PHASER

A neural accelerator that runs at the speed of light.

Designed around parts the telecom and display industries already make by the million. No new fabs, no exotic materials, and no new power plants.

01 / The light

Intelligence runs on electrons.PHASER runs it on light.

200–3,000×1

less energy per step than an equally capable dense recurrent network, modeled at a million optical modes.

  1. LightA steady rain of 650 nm wave packets. The bands are its crests.
  2. GlassLight slows to c/1.5 inside and passes through; the glass absorbs almost none of it.
  3. ElectronTunnels toward the next atom, mostly falls back, and only sometimes gets through. Each hop leaves the lattice shaking: heat. The glass stops it entirely.
  4. AtomA nucleus inside its electron cloud: points sampled from the orbital's density.
~1pJ2per multiply-accumulate on today's AI chips (0.7–2 pJ at chip level), almost all of it ending as heat.
≤0.001pJ2per equivalent multiply in PHASER (0.06–1 fJ), modeled at a million optical modes. Light loses 10–30 % per round trip4 and the gain replaces it every trip; the glass itself isn't where the energy goes.

Pushing charge through a wire costs energy every time. Light crossing glass barely loses any: it interferes with itself on the way through, and that interference is the arithmetic. It still loses 10–30 % per round trip to mirrors and coatings, and PHASER pays to put that back.

02 / The machine

Watch the light compute.

This is not an animation. It is our research simulator running in your browser, modelling red light bouncing between two mirrors through a stack of four etched glass phase plates. Each plate's pixels are the program; every pass transforms the whole light field at once, and the light makes about four billion round trips a second.

The cavity at true scale: 1.28 mm wide, 25 mm between the mirrors. Each wavefront carries the simulated intensity |E|² at its plane.6round trip 000000
Phase plate 2, all 64 × 64 pixels. Graphite: its etched phase program. Red: the light crossing it this round trip, up and back.
0.26ns5

per input step: one round trip up the stack and back, glass included.

10B3

input steps per second, modeled: this stack gets there with about 3 light pulses in flight at once, sharing one optical program (a ring needs 13). 3.9 billion per stream without them.

79%4

of the light kept each round trip, modeled. The gain replaces the rest every trip.

Live simulation in your browser of the research's linear stack of fabricated phase plates (TensaCo/phaser-design, Exps. 30 and 33): a scalar field on a 64 × 64 grid at 20 µm, angular-spectrum propagation between thin elements and glass slabs, gain clamped just above threshold, one round trip per input. The browser port matches the research simulator to the last bit over 4,000 round trips.

03 / The world

AI is running out of electricity.

128weeks7

to get a large power transformer.

5years+8

typical wait to connect a new power plant to the US grid.

$400B+9

spent on AI infrastructure by five companies in 2025 alone.

Latest developments

From the lab.

  1. Phaser: a hyperparallel quantum photon computing systemRead →
All posts

Get updates as we build it.

New results, the first bench prototype, and when there is something you can run.