Last week, I attended the 2nd annual ESA-NASA Workshop on AI Foundation Models for Earth Observation. The tone of the event was noticeably different from similar conversations even a year ago. In 2025, much of the industry discussion revolved around the possibility. In 2026, the conversation has shifted toward operational reality. The downstream applications and tools are used in production environments.  

The geospatial community is no longer asking whether foundation models can work for Earth Observation. The focus is now on reliability, reproducibility, scalability, and production deployment.

That transition matters because operational systems have very different requirements than research demonstrations. Accuracy and reproducibility are no longer the defining metrics. Stability, latency, throughput and governance are becoming equally important.

This shift aligns closely with how we have been approaching GeoAI at Sparkgeo through initiatives such as Prescient and our broader work around geospatial intelligence systems. It also reflects themes we have been discussing publicly for some time: geospatial AI only becomes valuable when it moves beyond experimentation and starts supporting real operational decision-making.

Reliable Agents Over Fragile Demos

One of the clearest themes of the workshop was the growing consensus that agentic systems should prioritize reliability over novelty.

Several speakers emphasized that Large Language Models should not function as authoritative reasoning engines for scientific workflows. Instead, LLMs should primarily orchestrate deterministic systems underneath them.

Leo Thomas from Development Seed summarized the philosophy:

“Orchestrate tools, don’t generate answers.”

The workshop repeatedly reinforced four operational principles:

• Use LLMs for routing and workflow coordination rather than factual generation
• Ground outputs in deterministic systems such as SQL, APIs, and Python workflows
• Keep prompts simple while embedding complexity inside tools
• Store generated analysis code to ensure reproducibility end-to-end  

This direction mirrors broader trends emerging across GeoAI research and operational systems. Increasingly, the industry recognizes that explainability, reproducibility, and governance matter as much as raw model capability.  

For geospatial workflows specifically, reliability becomes even more important because Earth Observation systems are inherently stateful and data-intensive. 

Satellite imagery, DEMs, SAR products, vector systems, and temporal archives cannot realistically fit inside an LLM context window. Passing large spatial datasets directly through generative systems creates silent transformation risks and unpredictable behaviour. Managing state and data access is extremely important when working with geospatial data and LLMs.

This architectural pattern closely aligns with the direction we are pursuing with Prescient.

The GeoAI Stack Is Beginning to Standardize

Another significant takeaway was the maturation and consolidation of the GeoAI tooling ecosystem.

Historically, geospatial machine learning workflows have fragmented across custom preprocessing pipelines, inconsistent model interfaces, and disconnected evaluation approaches. That fragmentation has slowed operational adoption.

The workshop demonstrated growing alignment around a more standardized GeoAI stack centered on TorchGeo. Both TerraTorch and TerraKit recently joined the TorchGeo ecosystem to present a more unified approach to geospatial foundation models.

Together, these frameworks are helping centralize geospatial model training, evaluation, and data preparation workflows.  

This standardization is important because operational systems require consistency more than experimentation. Standardized tooling reduces implementation variance, improves reproducibility, and makes deployment architectures more maintainable across organizations.

This is where operational GeoAI increasingly looks less like model experimentation and more like systems engineering.

Embeddings Are Becoming Infrastructure

Embeddings were another dominant topic throughout the workshop.

The industry increasingly views embeddings not simply as performance optimizations but as reusable semantic infrastructure layers for geospatial reasoning. Instead of repeatedly training task-specific models, embeddings provide generalized representations of imagery, SAR, DEMs, and multimodal Earth systems that can support downstream tasks more efficiently.  

Google’s AlphaEarth Foundation (AEF) embeddings were frequently discussed as an emerging proxy for ground truth in some research applications because they incorporate multimodal datasets, including SAR and elevation data. Researchers are beginning to use these embeddings to corroborate environmental phenomena.  

A hot topic at the workshop was Terrabytes BetaEarth product. In their experiment, they recovered 97% of downstream task utility from AEF embeddings. 

What exactly is in that latent space? 

Benchmarking

At the same time, benchmarking remains unresolved. This tension reflects a broader reality that surfaced repeatedly during the workshop:

There is no universally “best” foundation model. The community is increasingly converging on the idea that organizations must develop bespoke benchmarks aligned with their operational goals and defend those claims using grounded scientific methodology rather than relying solely on leaderboard metrics.  

From Accuracy to Latency and Throughput

One of the more important operational discussions centred on how success metrics change when systems move from research into production.

Research environments prioritize benchmark accuracy. Production systems prioritize stability, latency, throughput, cost efficiency, and scalability.

Those priorities are beginning to reshape deployment architectures across GeoAI.

The workshop featured several discussions around serving geospatial foundation models operationally using:

• vLLM
• Kubernetes-native inference pipelines
• Distributed serving architectures
• Cost-optimized model orchestration systems  

Operational deployment is no longer hypothetical. The tooling ecosystem supporting scalable GeoAI inference is maturing quickly.

The Geospatial Knowledge Gap Still Exists

Despite rapid advances in AI systems, a recurring theme throughout the workshop was that generalized LLMs still struggle with geospatial reasoning.

Coordinate systems, topology, spatial relationships, temporal consistency, and Earth Observation semantics remain weak areas for standard language models.  

This reinforces something the geospatial industry has known for years: location is not just another datatype.

Spatial systems require specialized semantic layers, coordinate-aware architectures, and domain-specific reasoning frameworks. As GeoAI continues to evolve, the challenge will not simply be building larger models. It will be building systems capable of understanding how the Earth actually behaves.

GeoAI Has Entered Its Operational Era

The most important takeaway from the workshop was not tied to any individual model or framework.

It was the realization that GeoAI is transitioning from exploratory research into operational engineering.

The industry is beginning to converge around:

• Standardized tooling ecosystems
• Reproducible workflows
• Deterministic orchestration architectures
• Kubernetes-native deployment systems
• Science-driven evaluation methodologies
• Reliable agentic workflows grounded in tools rather than generated facts

That transition represents an important maturation point for the industry.

The experimentation phase is not over, nor should it be. But the conversation is now shifting toward operational trust, scalable delivery, and decision-grade Earth intelligence.

That may ultimately be the strongest signal coming out of this year’s workshop.