The Agent Loop.In Silicon.

Agents think in loops, calling one model after another. Phynomy runs the whole loop on a single chip.

A new kind of AI chip, built for agents.

The problem

Where the energy goes

Today's chips were built to run one big model from start to finish. An agent loop works nothing like that.

The model lives far away

On a normal chip, each model lives far from where the work happens. Using it means hauling it across the chip first.

The loop never rests

An agent runs model after model, dozens of times in a single task — so that hauling barely stops.

Most power just moves models

Step after step, moving the models around burns more power than the thinking ever does.

It's like a genius who re-reads the whole manual before every sentence.

The answer

The whole loop stays on the chip

Three things ordinary silicon can't do.

It thinks where the model lives

The work happens right where a model's knowledge already sits — so there's almost nothing to move.

Every model stays loaded

Planner, coder, checker all live on the chip at once. Switching between them takes microseconds.

Hand-offs never leave the chip

Each step passes its work to the next without anything leaving the silicon.

Nothing reloads. The loop just runs.

The payoff

Far more thinking, for the same energy.

~1×
a flagship GPU
~30×
Phynomy · indicative

useful work per unit of energy

The evidence so far

Early work, grounded in results

The approach is demonstrated by independent labs, the device is patented, and the core has been made in the lab.

Demonstrated independently

The approach behind it has been demonstrated in independent, peer-reviewed studies.

Patented at the core

The device at the heart of it is covered by granted patents.

Made in the lab

The core memory arrays have been fabricated and measured as physical devices.

Modern AI thinks in loops.

The hardware needs to catch up.

THE AGENT LOOP. IN SILICON.

Get in touch