Google's 16 Green AI Projects Get Models and Mentors, Not Money

Google DeepMind announced on 7 September that it has selected 16 organisations for the first Asia-Pacific cohort of its accelerator programme, AI for the Planet. The cohort mixes startups, nonprofits and research teams working on biodiversity, sustainable agriculture and carbon solutions.

What participants receive

The programme opened with a hands-on bootcamp in Singapore this week and runs for three months. Participants get access to Google's AI stack, tailored technical support and mentorship from Google experts. The named models are specialised rather than general-purpose: AnthroKrishi, ForestCast, AlphaEarth Foundations, SpeciesNet and Perch — an agriculture model, a forest-loss forecaster, an Earth-observation embedding model, a species classifier and a bioacoustics model.

What the framing gets wrong

Headlines describing Google as "backing" or "funding" 16 projects import a financial commitment that the announcement does not contain. There is no money figure in the post at all — no grant size, no cheque, no equity, no compute credit value. What is being supplied is access and advice, both of which have real worth, and neither of which is a cash investment. For nonprofit and research participants in particular, three months of mentorship is a different proposition from a grant, and the distinction matters to anyone assessing whether these projects can now be built.

Selection is the harder number to read

Google does not disclose how many organisations applied, so the 16 cannot be read as a selectivity signal. Nor does the post say what happens after the three months — whether cohort members keep model access, or whether the specialised models named here become generally available. Both are the questions that determine whether the programme compounds or ends.

The models are the story

The five named models are the most concrete thing in the announcement, and several are not broadly available. AlphaEarth Foundations produces embeddings of satellite imagery that let a small team query global Earth observation without building a pipeline; Perch classifies species from audio, which is how biodiversity monitoring scales beyond human listening; ForestCast forecasts forest loss rather than recording it after the fact. For a conservation nonprofit, access to these is plausibly worth more than a small grant would be — which is exactly why the absent funding figure should be reported as absent rather than assumed.