A new inference architecture, the compiler that feeds it, and the silicon built to run it. Less data movement, more efficiency, and models that behave more like the brain.
Integrated Stack
Dispersion AI is not just an algorithm. It is a full stack, from the internal representation of models to the chip design that runs them.
01 / Algorithm
Instead of running 100% of parameters per token, Dispersion AI splits them into blocks and activates only the ones relevant to each input. A neuromorphic approach: high total capacity, low activation per step, without relying on Mixture-of-Experts.
02 / Software
An internal compiler that takes open model checkpoints, analyzes their weights and transforms them into an optimized binary format. Accelerating inference without retraining the model or breaking its behavior.
03 / Hardware
DPUs are the silicon designed to run Dispersion AI natively: less external memory, less wasted energy and a clear path to order-of-magnitude efficiency gains.
Why it differs
Most AI hardware is optimized for dense matrices. That forces models to adapt to the machine. Dispersion flips the relationship: we design the machine around the real sparsity of inference.
Not MoE. Activation is dynamic and input-dependent, but it keeps a single base model and requires no parallel experts.
Proprietary format. The compiler generates its own binary artifact that decouples the model from hardware, protecting IP and enabling software updates.
Built for the edge. Lower power, lower bandwidth and less reliance on external memory: aimed at devices, autonomous agents and energy-constrained data centers.
Everything presented describes ongoing research and development work. It does not represent an available commercial product, nor does it guarantee final performance, power or time-to-market metrics.
Horizon · 6 Months
The priority for the next semester is closing the validation loop: converting a real model, measuring its quality and proving the path to silicon.
Convert open checkpoints to the internal format and compare quality against dense execution using perplexity on standard corpora.
The compiler and the inference runtime share the same golden reference, bit by bit, layer by layer.
Simulate a full token inference in an RTL simulator using the binary artifact generated by the compiler.
First area, frequency and power numbers with open-source tools before deciding on silicon fabrication.
Take small model conversions to 3B, 7B and larger architectures, validating that the sparsity pattern holds.
Validate the DPU datapath on reconfigurable hardware before committing a design to fixed silicon.
Company
A DeepTech startup building proprietary AI and intellectual property from Latin America. All of our research energy is focused on Dispersion AI and the Dispersion Processing Units.
Led by Dylan Orzatty, the company bets on proprietary, independent technology for the Global South.
“We are not optimizing the last generation. We are asking whether the next one should look different.”
Our network
Orzatty provides technology and development to companies that share our vision.
Team
A small, focused team building proprietary technology from scratch.
Founder & CEO
Legal President & COO
CFO
More team members will be announced soon.
Press
Coverage, announcements and material for the media.
Announcements and coverage — coming soon
Contact
Investors, partners and media: reach out directly.
📍 HQ: Puerto Ordaz, Bolívar State, Venezuela. Global operations via OrzattyCloud Edge.
