The cost of running AI is inherited.

When you run an AI model today, you’re barely paying for the math. You’re paying for the mismatch between models and the hardware they inherit. It shows up in how the hardware computes, moves data and spends energy.

The industry treats that cost as a law of physics. It isn’t. It’s a design choice.

At Exstill, we design the silicon and the model together, so that neither is compromised to fit the other. Only the math remains.

By evaluating more than ten million combinations of silicon and model, we arrived at our first result: the frontier model’s co-processor. It runs the small models around frontier AI 110× faster than today’s GPUs and uses 690× less energy.

Fast enough to change how people work with AI.
Efficient enough to change where AI can go.

Our approach isn’t tied to one result. It starts from the work and where it has to run (a data center, a robot, a device at the edge) and designs the chip that does that work best.

We are just getting started.

Meet the founders

Founders

Between us, we have built both halves of the problem: the silicon, and the models that run on it.

Valavan Manohararajah

Co-founder, CEO and chief hardware architect

Twenty-five years of hardware and software co-design at Altera, Intel and Cerebras, ten of them on AI chips; 65 patents. Most recently Chief Product Architect at Cerebras: shaped three generations of wafer-scale silicon from tile architecture to ASIC, rearchitected the compiler twice and led its low-bitwidth numerics.

Vithu Thangarasa

Co-founder and chief AI architect

Previously Principal Research Scientist and Inference ML Lead at Cerebras. Led AI research and productization: speculative decoding, distillation, sparsity, compression and model-architecture co-design. Made Cerebras’ production inference 4x faster. 20 papers at NeurIPS, ICML, ICLR, MLSys and more; 10 AI patents.

Careers

Join us

We design silicon and models as one team. The hardware people argue about models, the machine learning people about silicon, the software people about the path between them, and the design is better for it.

Hardware design

You’ll design silicon where every picojoule and nanosecond of a token is accounted for, and every circuit earns its place. If you count picojoules for fun, you’ll fit right in.

Machine learning

You’ll push the limits of small models, from better speculative drafts and more capable agents to entirely new workloads. You’ll explore new architectures and training methods, co-designed with the silicon they run on. If you’ve ever looked at a small model and thought, “Surely we can get more out of this,” we should talk.

A bit of software

You’ll write the code that ties it together: the search that weighs millions of designs to find the best one, and the software that programs the chip, so each new model loads as weights.

Sound like your kind of work? Tell us what you’d build.Write to careers@exstill.ai