Google DeepMind Selects 16 AI-for-the-Planet Projects Across APAC: Models, Climate, Agriculture, and Carbon

Google DeepMind has selected 16 organizations from eight Asia-Pacific markets for the inaugural AI for the Planet accelerator, moving the program from an earlier target of 10–15 teams to a final cohort of 16.

The portfolio spans biodiversity monitoring, climate resilience, agriculture, urban energy, carbon removal and nature-based credit verification. Its technical base is not a single foundation model: the program exposes teams to a stack that includes AlphaEarth Foundations, ForestCast, SpeciesNet, Perch and agricultural intelligence models, alongside Gemini and Gemma.

The four-day Singapore bootcamp runs September 7–11, 2026; tailored support continues through December, when participants present at Demo Day. Participation is equity-free, while access to cloud credits or TPUs remains subject to eligibility and approval.

The important question is therefore not whether the projects sound environmentally useful, but whether the accelerator can convert strong model capabilities into calibrated measurements, operational decisions and auditable outcomes under real-world constraints.

⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣

THE 16-PROJECT COHORT.

Six projects target nature and climate resilience, five focus on sustainable agriculture, and five address climate or carbon systems.

........

ORGANIZATION	MARKET	TECHNICAL OBJECTIVE
800 Trust	New Zealand	Continuous bioacoustic monitoring for environmental threats.
Kumi Analytics	Singapore	Remote sensing and deep-learning baselines for conservation.
Listening Lab	New Zealand	Data-efficient bioacoustic monitoring across sparse field datasets.
TelePIX	South Korea	Satellite intelligence for mangrove monitoring and protection.
Wildlife.ai	New Zealand	Open-source AI cameras for scalable wildlife observation.
Yayasan Ekosistem Lestari	Indonesia	Prediction of environmental degradation and disaster risk.
Edufarmers	Indonesia	Near-real-time pest, disease and weather guidance through messaging channels.
Living Roots	Thailand	AI-assisted design of biological fertilizers.
SIGMA	Singapore	Satellite-based crop-yield estimation and climate-resilience analysis.
Terrastack	India	Plot-level land intelligence for agricultural decisions.
X-Centric	Australia	Portable AI-enabled X-ray soil geochemistry.
Archeda	Japan	Satellite measurement for nature-based carbon credits.
City Syntax Lab	Singapore	Agentic AI for district-scale energy and carbon optimization.
Climitra Carbon	India	Geo-AI verification for invasive-species removal and biochar projects.
Farmers for Forests	India	AI drones for agroforestry carbon and biodiversity measurement.
Varaha Climate	India	Remote-sensing verification of regenerative agriculture and carbon removal.

........

The geographic distribution is concentrated rather than uniform: India contributes four teams; New Zealand and Singapore three each; Indonesia two; and Australia, Japan, South Korea and Thailand one each.

That mix matters because the cohort combines sensing-heavy projects with decision and verification systems. Some participants generate primary observations, while others infer future risk, optimize interventions or support carbon claims.

··········

⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁣⁤

HOW GOOGLE'S MODEL STACK MAPS TO THE WORK.

The portfolio joins Earth observation, bioacoustics, camera-trap classification and agricultural intelligence rather than applying one general model to every problem.

Google has named a common model portfolio, but it has not published a one-to-one assignment between every model and participant. The mapping below describes the most plausible technical fit from the disclosed capabilities and project objectives, not a confirmed deployment plan.

• AlphaEarth Foundations compresses multiple Earth-observation sources into 10-by-10-meter land and water embeddings. That representation is a direct fit for TelePIX, SIGMA, Terrastack, Archeda, Climitra Carbon, Farmers for Forests, Varaha Climate and Kumi Analytics, where change detection and spatial comparability matter.

• ForestCast combines satellite observations with historical land-change patterns to forecast deforestation risk. It is most relevant to forest, mangrove and landscape-resilience projects that need to prioritize field intervention before loss is visible in retrospective reporting.

• SpeciesNet recognizes 2,498 animal categories and was trained with more than 65 million labeled images. Wildlife.ai is the clearest fit, but the larger engineering issue is calibration across camera types, habitats, seasons and species prevalence.

• Perch provides reusable bioacoustic representations across animal groups and sensor environments. It aligns closely with 800 Trust and Listening Lab, where useful performance depends on detecting rare signals while controlling false alarms in long, noisy recordings.

• AnthroKrishi, Agricultural Landscape Understanding and Agricultural Monitoring and Event Detection provide field boundaries, landscape structure and event layers. They can support Edufarmers, SIGMA and Terrastack when satellite inference must become timely plot-level advice.

Gemini and Gemma can sit above those domain models as interface, reasoning and workflow layers. They add value only when their outputs remain grounded in measured data, expose uncertainty, preserve the source observation and permit human override.

··········

PROGRAM STRUCTURE, SUPPORT, AND UNRESOLVED TERMS.

The accelerator is concrete about timing, model access and equity, but several operational terms remain outside the public program description.

........

PROGRAM SIGNAL	PUBLIC DETAIL	OPERATIONAL IMPLICATION
Cohort	16 organizations across eight Asia-Pacific markets.	The final intake is larger than the original planning range of 10–15 organizations.
Bootcamp	Singapore, September 7–11, 2026.	A four-day technical and business sprint establishes baselines and project plans.
Tailored support	September through December 2026.	Teams have roughly three months to convert model access into demonstrable progress.
Model access	Google AI and Google DeepMind models are available to participants.	Public materials name the stack but do not disclose a fixed model assignment for each team.
Program terms	Participation is equity-free.	Google does not take ownership merely in exchange for accelerator participation.
Compute support	Cloud credits or TPUs may be available subject to eligibility and approval.	Compute is conditional, so project economics cannot assume unlimited subsidized capacity.
Demo Day	December 2026.	Participants are expected to show technical progress and a credible path to deployment.
Terms not specified publicly	No common cash grant, API quota, project KPI, commercial agreement, or detailed IP and data-rights framework is stated.	These items should be treated as undisclosed rather than assumed to be absent.

........

The final selection of 16 organizations exceeds the originally stated range by one. That indicates discretion in cohort construction; it does not by itself establish that the program's resources, technical staffing or compute envelope expanded proportionally.

The September-to-December window is sufficient for architecture work, data-pipeline repair, benchmark design and early deployment tests. It is not sufficient to validate seasonal agricultural effects, long-horizon ecosystem recovery, carbon permanence or broad causal environmental impact.

··········

WHAT THE COHORT WILL HAVE TO PROVE.

The decisive evidence will be task-specific measurement under deployment conditions, not model access or accelerator participation.

Biodiversity systems should report precision, recall and calibration by site, species and season; quantify false-negative risk; document sensor uptime; and disclose how much expert labeling is required to maintain performance after distribution shift.

Agricultural systems should separate map accuracy from decision value. Plot-level boundary quality, detection latency, advisory uptake, agronomic outcomes and failure handling for rare pest or weather events matter more than a single aggregate model score.

Carbon and nature-credit systems face a higher verification bar. Measurement, reporting and verification must address additionality, leakage, permanence, uncertainty propagation and auditable evidence trails. A high-resolution embedding or risk score is an input to verification, not proof of a climate claim.

Across the cohort, deployment readiness also depends on data rights, sensor reliability, offline behavior, model-update controls, unit economics, vendor concentration and documented fallback procedures. Those controls determine whether a promising prototype can become dependable environmental infrastructure.

Until those results are available, the cohort establishes a credible pipeline of applied AI projects, not verified environmental impact.

··········

FOLLOW US FOR MORE.

DATA STUDIOS

datastudios.org