100% of parameters per token
Dense execution activates every weight at each step, even when the input only needs a fraction.
Dispersion AI
Inference, redesigned.
Current models run 100% of their parameters for every token, even though most do not contribute to that prediction. Dispersion AI activates only what matters — and we build the compiler and the silicon to make that possible.
The problem
The hidden cost of dense inference.
Large-model inference is expensive because almost all the compute is wasted: most parameters do not contribute to the prediction of each token.
Energy on idle data
Moving weights that do not contribute burns energy and bandwidth without improving quality.
Memory as bottleneck
Reliance on external memory limits efficiency and makes hardware more expensive.
Architecture
Dynamic activation, not Mixture-of-Experts.
Dispersion AI splits model parameters into blocks and activates only the ones relevant to each input. A neuromorphic approach: high total capacity, low activation per step.
01
Input-dependent
Each token activates a different subset of blocks based on its content.
02
A single base model
No parallel experts: capacity lives in one model, not in a mixture.
03
Block-level sparsity
Block granularity enables efficient hardware without sacrificing quality.
04
Neuromorphic by design
The brain does not fire all its neurons for every stimulus. Neither should a model.
Software
A compiler that understands weights.
The internal compiler takes open model checkpoints, analyzes their weights and transforms them into an optimized binary format for sparse execution.
1
Open checkpoint
Public models as the starting point.
2
Weight analysis
The compiler studies the sparsity and relevance structure of each block.
3
Proprietary binary format
An optimized .bin artifact that decouples the model from hardware and protects IP.
4
Inference runtime
Sparse execution with the same golden reference, bit by bit.
Hardware
Silicon designed for sparsity.
Dispersion Processing Units (DPUs) run Dispersion AI natively: less external memory, less wasted energy and a clear path to order-of-magnitude efficiency gains.
Less external memory
Idle weights never move: the silicon only accesses what gets activated.
Less energy
Without useless data traffic, power scales with real activity.
Path to order-of-magnitude
Structural efficiency versus general-purpose hardware.
Comparison
Dense vs MoE vs Dispersion AI.
Three ways to run a large model, only one with real dynamic activation.
| Dense | MoE | Dispersion AI | |
|---|---|---|---|
| Activation | 100% of parameters | Experts per token | Relevant blocks per token |
| Base model | Single | Multiple experts | Single, with internal sparsity |
| Hardware | General-purpose GPU | GPU with routing | Dedicated silicon (DPU) |
| Efficiency | Low per token | Medium, with routing overhead | High, real activation |
FAQ
Frequently asked questions.
Is this just another AI accelerator?
No. Most hardware accelerates dense matrices. Dispersion redesigns the model-machine relationship: the model becomes sparse by design and the silicon is built around that sparsity.
Does it require retraining models?
No. The compiler transforms existing open checkpoints into the internal format without retraining from scratch, preserving model behavior.
What is the current status?
Ongoing research and development: real model validation, compiler-runtime integration and RTL simulation. It is not an available commercial product.
When will physical hardware exist?
The 6-month roadmap targets end-to-end RTL simulation and virtual synthesis before deciding on fabrication or FPGA prototyping.
Want to know more?
Investors, partners and researchers: let us talk.