Generation-focused inference silicon
Built for
the next token.
Markov Silicon is developing a chip for the generation phase of AI inference, bringing recurrence, attention and mixture of experts together.
Your GPUs process the prompt. Our silicon is designed to pick up the conversation and keep generating.
In development · Private demonstrations by invitation
A focused role in the serving stack
Prepare the prompt. Hand over the memory. Generate.
Processing a prompt and generating an answer are different jobs. We are designing the chip around generation: the repeated work of producing one token after another.
A customer’s GPUs prepare each request. The resulting conversation memory transfers to our card once, and generation continues from there. That handoff is part of the architecture.
- 01 / PREPARE
Your GPUs
Process the prompt
Read the input. Build the conversation memory. - One handoff
per request - 02 / GENERATE
Markov Silicon
Continue the conversation
Generate the next tokens across the hybrid model.
One architecture, multiple layer types
Designed for the hybrid model.
Hybrid models combine different ways of remembering, selecting and computing. Our architecture brings these layer types onto a shared execution path, with high-bandwidth memory for the model and its working state.
Recurrence
Carry a compact, evolving state from one token to the next.
Attention
Look back through the conversation’s key/value history.
Mixture of experts
Route each token to the experts selected by the model.
Where we are
Building the execution path.
The project is in active development, with bounded RTL simulation and ongoing compiler, runtime and hardware work.
The private demonstration presents selected simulation evidence and its scope. We have not yet qualified an FPGA implementation or fabricated an ASIC.
Request a walkthroughLet’s talk about what comes next.
We welcome conversations with investors, model developers and infrastructure teams exploring the next generation of inference hardware.