Lightly.ai Sustainable AI Performance Powered by NVIDIA DGX™
Operational Benefits at Production Scale
About the client
Pioneers in AI Technology
AMBER and Lightly: Transforming industries through advanced computing and vision AI
Lightly is the computer vision suite for ML teams building real-world vision systems. Founded in Zurich by ETH and Harvard alumni, Igor Susmelj and Matthias Heller, Lightly provides an end-to-end platform for managing datasets, preparing training data, and training better AI models.
Lightly’s flagship products, LightlyStudio and LightlyTrain, cover the full lifecycle of computer vision development. LightlyStudio supports dataset curation, labeling, quality assurance, visualization, and dataset management in a single workflow. LightlyTrain enables self-supervised pretraining and fine-tuning to improve model performance while significantly reducing labeling effort.
Lightly is a key partner for teams across industries such as agriculture, manufacturing, robotics, retail, and beyond that rely on high-quality computer vision systems in production.
To continue scaling vision workloads sustainably, Lightly sought an infrastructure model that combined high performance, cost control, and a significantly lower carbon footprint.
Challenges
Why Lightly Needed a Cloud Alternative
Balancing performance, cost, and sustainability for large-scale vision AI
Unsustainable Cloud Expenses
As demand for training and experimentation grew, Lightly’s GPU cloud spend climbed into the $17,000–$70,000 per month range, with peak prices reaching $16.10 per GPU-hour. At this scale, variable per-hour billing became a major business risk and limited the scope of experiments the team could justify.
Unpredictable Cloud Performance
Shared cloud instances introduced noisy-neighbor effects, virtualization overhead, and fluctuating throughput. For production-grade computer vision workloads, this meant training jobs could slow down or behave inconsistently, breaking continuous-training pipelines and making it harder to plan releases.
Limited Hardware Access
Gaining access to the latest, high-performance GPUs for long-running workloads was increasingly difficult and expensive. This constrained Lightly’s ability to train larger models, process more complex datasets, and keep up with the performance expectations of enterprise customers in safety-critical domains like autonomous driving.
Co-located, renewable-powered, and enterprise-supported
AMBER`s Solution
From Cloud Variance to Predictable Throughput on NVIDIA DGX™
Co-located, renewable-powered, and enterprise-supported
AMBER worked with Lightly to move from variable cloud instances to a dedicated NVIDIA DGX™ B200 cluster hosted in a Norwegian co-location facility powered by 100% renewable energy. The goal: stable performance, linear costs, and a clear sustainability story.
Each DGX B200 system provides eight Blackwell GPUs with 192 GB of HBM3e per GPU, enabling larger batch sizes and more complex models without running into memory ceilings. High-speed NVLink and NVSwitch interconnects keep multi-GPU throughput high for both computer vision training and future LLM workloads.
By shifting to self-hosted DGX capacity, Lightly replaces volatile per-hour cloud spend with a predictable operational run rate. AMBER and NVIDIA experts helped design, deploy, and tune the environment so that jobs can be scheduled efficiently and infrastructure stays optimized over time.
The cluster is tightly integrated with Lightly’s existing tooling and data pipelines, and – crucially – it operates in a renewable-powered data center, reducing the carbon footprint of compute-intensive AI workloads while maintaining enterprise-grade reliability.
Reasons for NVIDIA DGX™
- End-to-end AI platform
- Blackwell B200 performance
- Optimized for vision and LLM workloads
- Renewable-powered co-location
- Predictable economics
- Enterprise-grade support
- Data sovereignty and security
Benefits
Operational Benefits at Production Scale
Lower OpEx, faster training cycles, predictable throughput, sustainable power, and data control.
Predictable
Spend
With self-hosted DGX capacity, Lightly replaces fluctuating cloud pricing with a consistent operational model. Under continuous use, the effective cost drops to around $0.51 per GPU-hour, compared with $2.95–$16.10 per GPU-hour in the cloud, aligning spend directly with utilization instead of spot-market pricing.
Faster Training
Cycles
Jobs start immediately and finish reliably, enabling more frequent retrains and evaluations that line up with product release cycles. Queue-free access on a dedicated DGX fabric removes multi-tenant interference and reduces idle time between experiments.
Data
Control
Training data and production datasets remain within Lightly’s own environment, simplifying compliance and privacy management. Integration with the DGX software stack keeps data close to compute while maintaining enterprise-grade access controls.
Predictable
Throughput
As experiments and user load increase, the DGX B200 cluster sustains high throughput and stable latency. High-speed interconnects and low-precision optimization paths validated in NVIDIA InferenceMAX benchmarks provide a roadmap for scaling future inference workloads efficiently.
Sustainable
Performance
Running in a renewable-powered Norwegian data center dramatically reduces the carbon footprint associated with large-scale training, while dedicated hardware eliminates virtualization overhead. Lightly can maintain top-tier performance without compromising on sustainability goals.
Sharper Data
Efficiency
The additional performance headroom allows Lightly to run larger, more sophisticated experiments on curated datasets – further amplifying the gains of their core data-selection technology and unlocking new levels of accuracy and efficiency for customers.
Testimonials
Voices from Lightly & AMBER
Insights from Lightly & AMBER on the transformative impact of NVIDIA DGX™ technology.
