From two states to three.
How we’re building three-state hardware: a small prototype, reliable memory and compute, higher density, and a native ternary software stack.
Read the transcript
Every digital computer begins with a simple physical idea: control the flow of electricity. A transistor uses an electrical signal to let current flow or hold it back. Circuits turn that behavior into two reliable states: zero and one. From those two states, we build everything from spreadsheets to artificial intelligence.
We're taking the next step: building digital hardware around three reliable states. In balanced ternary, those states represent minus one, zero, and plus one. Each digit, called a trit, carries three possibilities. The opportunity is to make those three states useful throughout the computer, from how it stores information to how it calculates.
The manufacturing approach builds on the semiconductor industry's existing foundations. We're using CMOS-compatible metals and processes, with a path through CMOS-compatible fabs. The central engineering challenge is creating devices and circuits that can write, preserve, distinguish, and use three states reliably, then integrating them into a practical computing architecture.
Our first hardware target is deliberately small: a prototype with fewer than one hundred thousand transistors. At this stage, the priority is learning from real hardware. We want a manageable chip that lets us examine the three-state building blocks, connect them into working circuits, and run useful tests from end to end.
We'll measure how clearly the states remain separated, how accurately we can read and write them, and how well they hold their values over time. We'll test repeated operation and changing temperatures. Error detection and recovery belong in the design from the beginning. Reliable operation is the foundation on which everything else grows.
Once those results are strong, we increase density. More devices fit into less area. Circuits become more integrated, and we move toward smaller manufacturing processes. Each generation builds on measured results from the previous one: preserve reliable states, reduce the space and energy needed, and increase the amount of useful computation on a chip.
We're also changing the relationship between memory and computation. In today's mainstream systems, processors repeatedly fetch data from memory, perform calculations, and send results back. AI makes that traffic especially demanding because models contain enormous collections of numerical weights that must be accessed again and again.
Our direction brings nonvolatile memory and computation into the same physical fabric. Nonvolatile means stored information remains when power is removed. Useful computation happens where the data lives, reducing the need to move it back and forth. Model weights can stay resident, ready for the next calculation. Keeping data local attacks a major source of energy consumption.
Ternary model weights make the connection particularly clear. A weight of plus one means add. Minus one means subtract. Zero means skip. Those three actions map naturally onto three-state hardware. The machine can work directly with this simpler structure, while accumulating the results needed by the model.
Open model weights give developers a foundation to adapt and optimize models for this architecture. Models designed for ternary weights can store those values directly in persistent three-state memory. The hardware, model representation, and computation can then be developed together, with accuracy and useful performance guiding the work.
The software must follow the same direction. We're building toward a fully native ternary stack: firmware that starts the machine, native instructions, compilers, a runtime, and an operating system. Programs will use three-state values and operations throughout. The aim is for three-state hardware to execute that software directly, replacing binary logic at the machine's foundation.
The journey starts with a small prototype and grows through reliability, integration, and density. The destination is head-to-head competition with binary on real workloads: energy, speed, accuracy, and cost. We're upgrading the physical foundation of digital computing from two states to three, and building the computer around it.
