LGND scales geospatial AI from zero to 160 GPUs with Union.ai

LGND is on a mission to make Earth data as searchable as the web. The company turns satellite and aerial imagery into geo-embeddings (compact numerical representations that let you scan millions of square kilometers in milliseconds) powering spatial search, change detection, and analytics for industries from insurance to logistics. Their team includes co-founders from NASA, Google, the World Bank, and Mapbox, and they're backed by a $9M seed round led by Javelin Venture Partners.

To expedite time to market and not divert resources from their core offering, LGND sought an orchestration workflow tool. Their technical requirements included:

• Distributed model training at scale. LGND's flagship workflow trains a model that translates geo-embeddings to text-embeddings across tens of millions of samples. This required scaling GPU training from single-node experiments to dozens of nodes — a leap that typically demands dedicated infrastructure engineers.
• Zero latency tolerance. LGND highly values the fast, seamless experience they’ve created for their users. They sought to minimize the latency created by Kubernetes pod cold-starts.
• ETL pipelines that couldn't keep up with growth. As new enterprise customers onboarded, LGND wanted automated data pipelines moving data across S3, Snowflake, and Postgres. 
• Full infrastructure ownership on a startup budget. LGND preferred to run everything on their own AWS account — not absorb another managed compute bill with opaque pricing.

The team wanted a platform that could handle Python-native workflow authoring, eliminate container overhead for interactive workloads, scale GPU training across dozens of nodes, and run entirely within their own cloud.

LGND deployed Union.ai, the enterprise Flyte platform, into their existing AWS environment. This gave the team full infrastructure ownership and cost transparency while immediately unlocking Union's orchestration layer across every workload type: training, inference, and ETL.

1. Reusable Containers for Interactive Workloads

This was the feature that convinced LGND to adopt Union.ai. Reusable Containers (Actors) keep pods warm between job invocations, eliminating Kubernetes cold-start latency entirely.

2. Ray Distributed Training at 160-GPU Scale

LGND's core ML workload is fine-tuning a model that translates geo-embeddings to text-embeddings, enabling natural-language search over satellite imagery. Using Union's PyTorch Jobs integration, LGND runs Ray distributed training clusters directly from their Python workflow code – no infrastructure rewriting required. The team scaled from early single-node experiments to a 40-node, 160-GPU cluster running on L40 GPUs (g6e.12xlarge instances), with Union.ai handling node provisioning automatically.

3. ETL & Batch Pipelines

As new customers onboard onto LGND's platform, Union.ai triggers automated ETL workflows that move data across S3, Snowflake, and Postgres. This is critical as LGND scales from pilot projects to enterprise deployments across insurance, carbon, maritime, and energy verticals.

4. Large-Scale Satellite Imagery Inference (In Progress)

LGND is now scaling into large-scale inference, processing tens to hundreds of terabytes of satellite imagery using high-memory CPU nodes. Union's Flyte 2 data model enables parallel tiling jobs at scale, supporting the global embedding database that powers the LGND API's ability to search satellite imagery the way you'd search the web.

5. Python-Native Authoring

Scientists and ML researchers on the LGND team author and run workflows directly in Python. There's no YAML, no engineering translation layer, and no bottleneck between the science team and production infrastructure.

LGND signed with Union.ai on a 30-day trial. Within weeks, the team had validated Union.ai across ETL pipelines, their labeling UI, and early distributed training experiments and chose to become a customer. In the months that followed:

• Compute usage grew 111.5% and workflow executions grew 63.6%, driven by active scaling of distributed GPU training for their core geo-embedding model.
• GPU footprint expanded 4× from a ceiling of 10 nodes to 40 nodes running 160 L40 GPUs, without any changes to their underlying infrastructure or workflow code.
• Full cost transparency preserved. Union's BYOC deployment meant every dollar of compute ran directly on LGND's own AWS account, giving leadership full visibility into infrastructure spend.
• True partnership to unblock complex issues. When LGND encountered issues with Ray (including a Karpenter node scheduling bug, Ray version incompatibilities, and memory quota configuration) Union's support team worked through each issue alongside the LGND engineering team, earning a 5/5 support satisfaction score.

Launching the LGND API on Union.ai Infrastructure

In March 2026, LGND launched its API into public beta, a product that lets developers search satellite and aerial imagery using natural language queries or coordinates, with no model training required. The API's three core capabilities — creating embeddings with geo-foundation models like Clay, running similarity search against a global embedding database, and fetching source imagery — all leverage the distributed training and inference infrastructure that Union.ai orchestrates.

Looking Ahead

LGND plans to significantly expand its large-scale inference workloads as enterprise customers scale, processing ever-larger volumes of satellite imagery through Union's orchestration layer. For a seed-stage geospatial AI startup, Union.ai provided what's otherwise hard to find: enterprise-grade workflow orchestration that scientists can use directly, running entirely on their own cloud, with a support team invested in their success.