Explore the idea
The Third Possibility.
Why binary became the standard, what the early Setun computer showed, and how Trinary.ai is approaching computing with three values.
AI is evolving.
So should its chips.
We’re developing a new kind of chip to run AI with less energy. Some AI models simplify many of their learned numbers to just three values: −1, 0 and +1. We’re designing hardware around those values to bring more capable AI to everyday computers and devices.
The research behind the idea.
Meet Trinary.ai
Ternary AI in 60 seconds.
What three values have to do with AI, why moving data takes energy, and how we’re approaching chip design differently.
Read the transcript
AI is moving into everyday devices, from glasses that understand what we see to cameras that recognize what matters, and assistants that respond wherever we are.
Each response involves inference. That means running a trained model on new information to produce an answer.
Models rely on weights, numbers that guide their calculations. Memory stores those weights. Compute uses them. Storing and moving those numbers takes energy.
Today’s digital chips use binary, zeros and ones. Some AI models use just three weight values: minus one, zero, and plus one.
That simplifies multiplication. Flip the input’s sign, contribute nothing, or keep the input unchanged.
Here’s why three states matter. Eight binary digits give two to the eighth power: two hundred fifty-six combinations. Five ternary digits give three to the fifth: two hundred forty-three. Nearly the same capacity, with fewer digits. We’re building native hardware to turn that information advantage into lower-energy AI.
Trinary. Ternary AI. Native silicon.
01 · The problem
Smarter AI needs more than smarter software.
Every time AI answers a question or recognizes a sound, it runs a trained model. That’s called inference. Doing it means storing data, moving it around and making calculations. All three use energy.
Moving data costs energy.
Memory holds the model’s data. A processor uses it to do calculations. Sending data back and forth takes energy—even before the calculations happen.
Simpler models. Different needs.
Today’s digital chips store data using two values: 0 and 1. Some AI models use just three values for their learned numbers. Existing chips can run them; we’re designing a chip specifically for them.
Batteries have limits.
AI has to share a device’s battery with its cameras, screen and wireless connection. Using less energy for AI could mean longer battery life, less heat or more useful features.
02 · Why ternary
Three values. Simpler calculations.
AI models learn numbers called weights during training. Those numbers guide their answers. In a ternary model, many of those weights use just three values: −1, 0 and +1.
Instead of multiplying by a complicated number, the chip can reverse a number’s sign, produce zero or leave it unchanged. These simple operations are the starting point for our design.
More choices in every digit.
A binary digit has two possible values. A ternary digit has three. That extra choice adds up: five ternary digits can represent almost as many combinations as eight binary digits.
- Three values instead of two
- Nearly as many combinations, using fewer digits
- Our goal: turn that simplicity into energy savings
Fewer digits alone don’t guarantee lower power. The opportunity is to design the chip around them.
03 · What we’re building
A chip built for the way ternary AI works.
We’re developing an AI accelerator: a chip dedicated to running AI models. It will store information using three physical states and do calculations with those stored values. That’s what we mean by native ternary memory and compute.
Our first milestone is a working chip that connects to an existing computer and completes a useful AI task, such as recognizing a sound or an image.
We’ll compare it with an efficient conventional system doing the same job: how accurate are the answers, how quickly do they arrive, and how much energy does the whole system use?
- Less energy to run AI
- Longer battery life
- More AI running on your device
Give each value a home.
Build memory with three distinct states, so it can store the model’s −1, 0 and +1 values directly.
Do the work where the data is.
Use those stored values in calculations, reducing the need to move data between separate memory and processing hardware.
Make it useful.
Build the software that lets existing computers use our chip to run AI. Measure the benefit across the whole system.
In developmentWe’ll share performance results as we build and test the hardware.
04 · Use cases
What could this make possible?
First, run useful AI on an existing computer. Then build toward more capable assistants and smarter devices that can do more on each charge.
01 / First focus
Run AI on your computer
Connect our chip to a computer and run a useful AI task there. Doing more on the device can reduce the need to send data to a cloud service.
02 / Early examples
Recognize sounds and images
Identify a sound, recognize an object or spot an unusual machine vibration. These focused tasks are candidates for our first demonstration.
03 / As we scale
Bring AI assistants closer
Build toward larger language models that can answer questions and respond to voice on your own device, using less energy to do the work.
04 / Longer-term opportunity
Smarter devices. Longer between charges.
Glasses that interpret what you see. Cameras that recognize important events. Drones that respond to their surroundings. Lower-energy AI could help these devices do more with the battery they have.
The battery-life benefit depends on how much power AI uses compared with the device’s other jobs, such as powering a screen or keeping a drone in the air.
Better hardware should make a practical difference: useful answers, quick responses and less energy to deliver them.
Build with us
05 · Company
Built in Austin. Thinking beyond binary.
Trinary.ai is an Austin-based chip company co-founded by Vinny Lingham and Isaac Castro. We’re developing hardware designed around the simpler calculations of ternary AI.
Our mission is to make capable AI practical in more places, by rethinking how chips store information and use it.
A clear mission. Deep engineering.
Vinny Lingham is a recognized thought leader in crypto and digital identity who brings decades of experience as an entrepreneur, investor, and blockchain pioneer. He founded Civic and Gyft (acquired by First Data), was a General Partner at Multicoin Capital, and a seed investor in leading projects such as Solana, Filecoin, and Render. He also founded digital media powerhouse, incuBeta, acquired by The Carlyle Group. He is currently also the Chairman and Co-founder of Praxos Capital, a delta neutral, liquid hedge fund.
Isaac Castro is a serial deep-tech entrepreneur, engineer, and inventor with nearly two decades taking breakthrough hardware from lab to market. He co-founded Emerge (acquired by a semiconductor company), which pioneered a proprietary MEMS ultrasonic chip fabricated at top foundries. Earlier, he co-founded Adaptable and licensed his cancer radiation therapy technology to a medical device company. He is a former technology leader at Siemens. Named MIT Innovator Under 35 and World Economic Forum Technology Pioneer.
Research Partners
Working together on materials and AI hardware.
06 · Contact
Help shape what AI runs on.
For researchers, builders and potential partners.
hello@trinary.ai