Microsoft's AI for Earth Grants $50M to Climate Startups Using Foundation Models

Published 2025-05-30 · Environment

Microsoft announced on May 28, 2025, that its AI for Earth programme has distributed $50 million in grants, cloud computing credits, and technical support to 40 climate-focused startups deploying foundation models for environmental applications. The announcement, made at the company's annual Sustainability Summit in Seattle, represents the largest single round of AI-focused climate funding from a major technology corporation.

The funded companies span five continents and address a range of environmental challenges: direct air capture optimisation, satellite-based deforestation monitoring, soil carbon measurement, ocean plastic detection, wildfire prediction, and precision irrigation. What unites them is the use of large-scale AI models — including fine-tuned versions of GPT-4, open-source large language models, and custom computer vision architectures — applied to problems that have traditionally required expensive physical instrumentation or manual expert analysis.

The Scale of the Programme

Microsoft launched AI for Earth in 2017 with an initial commitment of $50 million over five years. The programme has since evolved from a small grants initiative into what the company describes as a "full-stack AI accelerator for climate innovation." The 2025 cohort received awards ranging from $500,000 to $3.5 million, supplemented by Azure cloud credits valued at up to $5 million per company over three years.

Beyond funding, each startup gains access to Microsoft's AI research division, which provides model architecture consulting, training data curation support, and deployment engineering. Several companies in the cohort are collaborating directly with Microsoft Research labs in Cambridge, Beijing, and Bangalore on joint publications and shared model development.

Dr. Lucas Joppa, Microsoft's chief environmental officer until 2024 and the architect of AI for Earth, has argued that foundation models represent a step-function change in environmental monitoring. "The bottleneck in climate technology has never been a lack of data — satellites produce petabytes of Earth observation data daily. The bottleneck has been turning that data into actionable insight at the speed and scale the problem demands. Foundation models compress what used to require a team of remote sensing specialists and weeks of computation into a single inference call."

Selected Grant Recipients

CarbonSight — Optimising Direct Air Capture

CarbonSight, based in Zurich, uses a fine-tuned vision-language model to analyse real-time sensor data from direct air capture (DAC) facilities and dynamically optimise fan speeds, sorbent regeneration cycles, and energy consumption. The company's pilot plant in Iceland reduced energy use per tonne of CO2 captured by 18% during a six-month trial, a figure that CEO Dr. Hanna Lindqvist attributes to the model's ability to anticipate weather-driven changes in ambient CO2 concentration up to four hours ahead.

Canopy Analytics — Real-Time Deforestation Monitoring

Canopy Analytics, operating from Sao Paulo and Nairobi, deploys a computer vision model trained on Sentinel-2 satellite imagery to detect logging and land-clearing events in near-real-time across the Amazon and Congo basins. The system processes imagery within 90 minutes of satellite acquisition and flags potential deforestation events with a spatial resolution of 10 metres. In a validation study across 2.3 million hectares of Brazilian rainforest, the model detected 94% of deforestation events larger than one hectare, with a median detection latency of 3.2 days — compared with the 30-to-60-day delay typical of manual interpretation workflows.

TerraByte — Soil Carbon Measurement

TerraByte, headquartered in Melbourne, combines satellite multispectral imagery with ground-penetrating radar data processed by a custom transformer model to estimate soil organic carbon stocks at field-level resolution. Traditional soil carbon measurement requires physical core sampling and laboratory analysis at a cost of approximately $25 per sample, with each hectare requiring 5 to 10 samples for reliable estimation. TerraByte's model reduces the number of required physical samples by approximately 70%, lowering measurement costs from $150-250 per hectare to $40-70 — a threshold that could make carbon farming economically viable for smallholder operations.

AquaVision — Ocean Plastic Detection

AquaVision, based in Amsterdam and Accra, operates autonomous surface drones equipped with multispectral cameras and an onboard neural network that classifies floating debris as plastic, organic matter, or other material in real time. The system maps plastic accumulation zones in coastal waters, enabling targeted cleanup operations by partner organisations. In a three-month deployment along the Ghanaian coast, AquaVision's drones identified 847 individual plastic aggregation events and directed cleanup vessels to 312 sites, recovering an estimated 23 metric tonnes of plastic waste.

Environmental Impact of AI Itself

The programme has not been without criticism. Several environmental groups have noted the paradox of using energy-intensive AI models to address climate change. Training a single large language model can emit as much carbon as five automobiles over their entire lifetimes, according to a 2022 study by researchers at the University of Massachusetts Amherst.

Microsoft has addressed this concern by requiring all grant recipients to run their AI workloads on Azure regions powered by renewable energy and to report the energy consumption and carbon footprint of their model training and inference operations. The company has also committed to publishing an annual environmental impact assessment of the AI for Earth programme itself, including the embodied carbon of the hardware used.

Dr. Sasha Luccioni, a climate AI researcher at Hugging Face who was not involved in the programme, offered a nuanced assessment: "The carbon cost of training these models is real and should not be dismissed. But the question is whether the net environmental benefit — the emissions avoided, the forests protected, the carbon sequestered — exceeds the computational cost. In many of these applications, the math is clearly favourable. In others, it's too early to tell. Transparency in accounting is essential."

Foundation Models and the Democratisation of Environmental Science

Perhaps the most significant structural impact of the AI for Earth programme is the way it lowers barriers to entry for environmental AI applications. Several grant recipients reported that the availability of pre-trained foundation models reduced their development timelines from years to months, because they no longer needed to collect and label the massive training datasets that computer vision and natural language processing historically required.

Agrim Shukla, founder of TerraByte, described the shift in practical terms: "Three years ago, building a soil carbon estimation model from scratch would have required 100,000 labelled soil samples and two years of training time. With foundation model fine-tuning, we achieved comparable performance using 12,000 samples and six weeks of compute. That compression is what makes the economics work for a startup."

The trend raises deeper questions about the concentration of AI capability in a small number of technology companies. Microsoft's programme, while generous in scale, effectively creates a dependency relationship between climate startups and a single cloud provider. The company has pledged to ensure that all models developed through the programme are published under open-source licences within two years of commercial deployment, but critics argue that infrastructure dependency persists regardless of software licensing.

For broader coverage of AI applications in environmental monitoring and sustainability, explore our AI for Environment research repository. Related reporting includes DeepMind's flood forecasting partnership in South Asia and our analysis of EU AI Act enforcement implications for environmental monitoring technologies.

← Back to Publications