Nectar provides an on‑premise AI datacenter that consolidates multiple NVIDIA H100 GPUs in an immersion‑cooled chassis, delivering sub‑5 ms inference latency for robotic fleets. Its orchestration software, the Brain, schedules workloads, monitors thermal health, and automates model versioning across nodes, enabling continuous inference and overnight model fine‑tuning while keeping data on‑site.
Funding
Funding not disclosed
Founders
Product
Problem
Robotic fleets on the factory floor require ever‑increasing AI compute for perception, planning, and control, but scaling GPU resources per robot is cost‑inefficient and introduces high network latency. Cloud‑based GPU services add 50‑200 ms round‑trip delays and force proprietary production data off‑site, creating privacy and compliance concerns. As robot numbers grow, the lack of a purpose‑built edge compute layer hampers throughput and operational scalability.
Solution
Nectar delivers an on‑site AI datacenter that consolidates high‑performance GPU compute into a single immersion‑cooled node (“the Box”) and pairs it with fleet‑wide orchestration software (“the Brain”). The Box houses NVIDIA H100 GPUs interconnected via NVLink, providing sub‑5 ms inference latency for all robots on the floor while keeping data within the premises. The Brain schedules workloads, monitors thermal health, and automates predictive maintenance and model versioning across multiple nodes, enabling continuous day‑time inference and night‑time model fine‑tuning on the same hardware. This architecture eliminates the connectivity tax of cloud GPUs, reduces per‑GPU cost from $3.50 /hr to an amortized $0.40 /hr, and scales seamlessly as the robot fleet expands.
Target Audience
The primary customers are manufacturers and system integrators operating large robot fleets in sectors such as automotive assembly, electronics manufacturing, and logistics automation, where low‑latency edge AI and data sovereignty are critical.
Features
- 27 ft³ immersion‑cooled chassis integrating up to multiple NVIDIA H100 GPUs with high‑bandwidth NVLink and redundant power supplies.
- Sub‑5 ms on‑prem inference latency achieved by colocating compute with the robots, eliminating network round‑trip delays.
- Fleet orchestration platform that handles workload distribution, thermal monitoring, predictive maintenance, and automated model versioning across any number of Boxes.
- Dual‑use compute cycle: real‑time inference during production shifts and overnight fine‑tuning of models on collected data, maximizing hardware utilization.
- Zero data egress architecture: all raw sensor streams and model updates remain on‑site, supporting strict data‑privacy and compliance requirements.
- Scalable node management
- additional Boxes can be added without manual reconfiguration; the Brain automatically balances load and provides failover.
- Cost‑effective pricing model with on‑prem amortized cost of $0.40 per GPU‑hour versus $3.50 per GPU‑hour for equivalent cloud resources.
- Secure remote diagnostics and firmware updates via encrypted channels, ensuring continuous uptime with minimal onsite intervention.