---
type: analysis
slug: olmoearth-s-cost-advantage-is-reusable-inference-not-cheap-pretraining
title: "OlmoEarth’s cost advantage is reusable inference, not cheap pretraining"
description: "AI2’s report says inference accounted for 98% of mapping cost in OlmoEarth v1.2. Precomputed embeddings shift that expense into a shared feature layer, which can make later mapping tasks cheaper—though the release does not establish the savings for any particular application."
published_at: 2026-10-24T09:00:00.000Z
topics: ["workflow"]
source_urls: ["https://huggingface.co/blog/allenai/olmoearth-v1-3","https://allenai.org/blog/olmoearth-v1-3","https://allenai.org/blog/planteur","https://deepmind.google/blog/beyond-tree-cover-mapping-the-worlds-forest-types","https://allenai.org/blog/olmoearth-infrastructure"]
author: "Andrey Reshetnikov"
author_profile: https://newruntime.com/owner-profile.md
---
# OlmoEarth’s cost advantage is reusable inference, not cheap pretraining

AI2’s report says inference accounted for 98% of mapping cost in OlmoEarth v1.2. Precomputed embeddings shift that expense into a shared feature layer, which can make later mapping tasks cheaper—though the release does not establish the savings for any particular application.

## Summary

AI2 released OlmoEarth v1.3 with model weights, pretraining data, code, precomputed annual satellite embeddings and a browser-based Explorer. The 10-meter embeddings summarize yearly Sentinel-1, Sentinel-2 and Landsat observations; the Explorer supports inspection, point labeling, small-classifier training and GeoTIFF export. The defensible cost story is reuse: according to the technical report, inference made up 98% of mapping cost in v1.2, so using a prepared representation across tasks may amortize that work. The release does not demonstrate end-to-end savings or application accuracy for a specific region.

## What happened

OlmoEarth v1.3 produces 128-dimensional vectors for each 10-meter pixel from yearly radar and multispectral time series; its first 64 dimensions can also be used independently. AI2 reports that its embeddings ranked first on 20 of 27 evaluations at 128 dimensions and 18 at 64, compared with AlphaEarth Foundations and Tessera v2. The precomputed archive covers 2023–2025 globally and selected countries from 2017–2022, and AI2 says it is about 150 TB per year with source-scene provenance for each tile. The account-free Explorer can compare years, find similar locations, train a small classifier from labeled points and export GeoTIFFs. AI2 estimates the final pretraining run at roughly $2,720 in rented H100 time. That figure is a training estimate, not the cost of serving, storing or validating the archive.

## What the archive adds

The archive’s earlier OlmoEarth infrastructure note describes separating CPU work such as preparation and map assembly from GPU inference. More directly, the dossier’s technical-report finding says inference accounted for 98% of mapping cost in v1.2. That is a report-attributed figure, not an independent cost audit. The dossier also describes a pretraining corpus of 285,288 global samples covering 2.56-kilometer squares over 360-day periods. Together, these details support a specific mechanism: pay to create a reusable representation, then avoid repeating that expensive inference step for every downstream mapping task. They do not establish the savings a team will achieve in its own pipeline.

## Trend signal

Geospatial AI is moving toward shared feature layers that let teams build task-specific maps on top of processed satellite observations. This changes where repeated work may occur: instead of deriving representations from raw imagery for each task, a user can start from prepared embeddings and focus on labels and downstream modeling. The economics depend on reuse; if many tasks use the same embeddings, the inference cost can be spread across them. If a project needs a different input, coverage period or processing workflow, that advantage may be smaller. Open artifacts and an Explorer lower the barrier to experimentation, but they do not make a global feature layer equivalent to a validated local map.

## Why now

According to the technical report finding in the dossier, inference accounted for 98% of mapping cost in OlmoEarth v1.2, giving AI2 a reason to make representations reusable across downstream tasks. The dossier’s pretraining-dataset finding describes a corpus of 285,288 global samples with satellite time series, supporting broad representation learning. The precomputed archive and no-account Explorer make those representations easier to try without first building a local inference pipeline.

## Connections to other events

The key second-order effect is a potential change in the marginal cost of asking another mapping question. According to the OlmoEarth v1.3 technical report cited in the dossier, inference represented 98% of mapping cost in v1.2; reusable embeddings are designed to amortize that step across downstream tasks. A team could therefore test multiple classifiers or questions without rerunning the same representation pipeline each time. That is an inference about the cost mechanism, not a verified end-to-end saving: the evidence gives no task-level cost comparison, and the archive itself is large. The advantage will depend on whether the prepared representation fits the task and whether reuse outweighs data-handling and validation work.

## Why it matters

Builders should evaluate OlmoEarth as a shared input layer, not as a ready-made answer to every mapping problem. A team with several related tasks may have more opportunity to amortize the representation than one-off users. Buyers should compare total workflow costs—including archive access, storage, local labels and evaluation—rather than treating the reported pretraining estimate as the price of a map. The open weights, data and code offer room to inspect and adapt the system; the Explorer makes initial experiments accessible without local GPU setup. The practical question is whether the reusable representation reduces repeated inference enough to matter in the buyer’s own workload.

## What happens next

- **A public third-party evaluation will test OlmoEarth on a named downstream mapping task and report results for a specified geography.** (6 months, 60%). Watch for: A reproducible public report or repository provides task and region-specific evaluation results.
- **A public user report will document a practical archive-access or data-handling constraint when using the precomputed embeddings.** (3 months, 55%). Watch for: A user issue, tutorial or technical post describes a concrete download, storage, tiling or processing bottleneck.

## Limits and uncertainty

The 98% inference-cost share is reported in the technical report for v1.2; it is not an independent audit or a direct measurement of v1.3 savings. AI2’s benchmark rankings are reported by AI2, and the evidence does not establish independent replication or accuracy for a particular task and geography. The roughly $2,720 estimate is for the final pretraining run, not the annual archive’s production, storage, access or downstream validation. Coverage is global only for 2023–2025, with selected-country coverage for earlier years. The evidence also does not settle whether embeddings transfer well to a team’s local task; if local data preparation and validation dominate its costs, the proposed inference savings may have limited practical effect.

## Skill to build now

Practice spatially separated evaluation: this week, label examples across several areas in the Explorer, hold out one distinct area from training, and report results on that holdout before deciding whether the embeddings save work for your task.

## Reader takeaway

This week, choose one mapping task and run a spatial holdout test using the Explorer: label points in several areas, train the small classifier, and reserve a geographically distinct area that is never used for training. Record performance on that held-out area and compare it with your current baseline. The result will test whether the embeddings are useful for your geography while keeping the cost question concrete: measure the work avoided, not just the advertised pretraining estimate.

## Open question

Across multiple downstream tasks, do reusable embeddings actually reduce total mapping cost once archive handling, local labels and spatial validation are included?

## Sources

- [OlmoEarth v1.3: State-of-the-art embeddings for the whole planet](https://huggingface.co/blog/allenai/olmoearth-v1-3)

- [OlmoEarth v1.3: State-of-the-art embeddings for the whole planet](https://allenai.org/blog/olmoearth-v1-3)

- [A crop-mapping benchmark for the tropics | Ai2](https://allenai.org/blog/planteur)

- [Not All Forests Are Equal: Mapping the World’s Forests — Google DeepMind](https://deepmind.google/blog/beyond-tree-cover-mapping-the-worlds-forest-types)

- [allenai.org](https://allenai.org/blog/olmoearth-infrastructure)

- [OlmoEarth v1.3 Embeddings (technical report)](https://allenai.org/papers/olmoearth-v1-3)

- [OlmoEarth Pre-training Dataset](https://github.com/allenai/olmoearth_pretrain/blob/main/docs/Pretraining-Dataset.md)

- [OlmoEarth v1.3 Embeddings — Source Cooperative](https://source.coop/olmoearth/olmoearth-v1-3-embeddings)

- [TESSERA v2: Scaling Pixel-wise Earth Foundation Models — arXiv](https://arxiv.org/abs/2607.03949)

- [OlmoEarth: Stable Latent Image Modeling for Multimodal Earth Observation](https://openaccess.thecvf.com/content/CVPR2026/papers/Herzog_OlmoEarth_Stable_Latent_Image_Modeling_for_Multimodal_Earth_Observation_CVPR_2026_paper.pdf)

Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Workflow](https://newruntime.com/topics/workflow.md): Explore the workflow topic hub.
- collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)