The problem
A modern factory streams sensor data from every machine, yet downtime still surprises, new materials take years of lab trial and error, and robots freeze when the real world differs from the training set.
How it works
Preferred Networks, founded in March 2014, builds the whole AI stack itself: MN-Core AI chips, its Preferred Computing Platform cloud, generative-AI foundation models, and applications for manufacturing, materials and chemicals, life sciences, retail, and the public sector. Joint R&D with Toyota spans autonomous driving and service robots, and group company Matlantis sells a universal atomistic simulator for materials discovery.
Pain points
Costly unplanned stoppages, decade-long materials development cycles, brittle automation, and simulation workloads too expensive to run on rented GPUs.
Business model
Paid enterprise AI products and solutions per industry, compute sold on its own AI cloud, chip-accelerated services, and product revenue from group companies such as Matlantis.
Challenges
Designing custom silicon burns cash ahead of revenue; every new vertical demands its own domain expertise and sales motion; and at each layer — chips, cloud, models — sit trillion-dollar incumbents.
Funding
- Raised: $95.4M equity investment from Toyota for mobility-AI joint R&D (total disclosed ~$116.4M at the time, per JnewsTech).
- Valuation: MISSING.
Latest — September 2026
On 18 September 2026 PFN and Preferred Computing Infrastructure announced MN-Core 2 compute for the Matlantis simulator within 2026 via the PFCP cloud — up to 4x faster than GPUs on selected tasks — the full vertical stack (chip, cloud, model, app) running inside one group.