Naveen Rao leaves Databricks to build next-gen AI compute hardware; Databricks plans to invest
Sep 12, 2025
Key Points
- Naveen Rao departs Databricks to found Unconventional AI, building specialized compute hardware to reduce inference costs for large language models.
- Databricks, valued at $100 billion with $4 billion in annual revenue, plans to invest in Rao's venture and secure a customer relationship.
- Rao's hardware targets memory bandwidth and interconnect latency bottlenecks that make token generation expensive, addressing a core constraint in AI infrastructure.
Summary
Naveen Rao is leaving Databricks to found a new company building purpose-built AI compute hardware designed to reduce inference costs. Databricks, valued at $100 billion with $4 billion in ARR, plans to invest in the venture.
Large language models are bottlenecked by memory bandwidth, interconnect latency, and power consumption, making each token prohibitively expensive to produce. A specialized compute architecture could address this by moving memory closer to compute and wiring chips with faster interconnects.
The specifics of the hardware remain unclear from available information, though there is speculation it could compete with projects like George Hotz's TinyBox. Databricks' investment signals confidence in the thesis and locks in a potential customer relationship for a company already generating $4 billion in annual revenue.
The departure reflects both the competitive intensity around reducing inference costs and the market's willingness to fund infrastructure plays that promise to unlock cheaper model serving at scale.