GPAI India Summit 2023: AI for the Global South

I spent most of this past week at the GPAI Summit in New Delhi. The Global Partnership on AI, which India chaired this year, convened its annual summit here for the first time, and the framing of the host country shaped the agenda more than I expected. The theme that ran through most of the substantive sessions was AI for the Global South, and specifically the question of whether the current trajectory of AI development was building anything useful for the four billion people who live outside the OECD core.

This is a trip report from someone in the agriculture corner of that conversation. I want to put down what I heard, what I disagreed with, and where I think the line of argument is heading. The official summit outputs will be published through the GPAI channels in the next few weeks; this is the unofficial parallel.

The argument in the rooms

The framing the host country put forward, with substantial support from the African and Southeast Asian delegations, ran roughly like this. AI as currently being built is a luxury good. The data centres, the model training, the application development, the deployment infrastructure: all are concentrated in a small number of high-income economies. The benefits of this AI flow primarily to consumers in those same economies. And the people whose lives could most be improved by AI (smallholder farmers, primary-care patients, students in under-resourced schools, workers in the informal economy) are the people the current AI ecosystem is largely not building for.

This is not a new thesis. What was new at GPAI 2023 was that it was being articulated at a multilateral institutional level with the host country’s political weight behind it. The session attendance, the panel composition, and the floor questions all reflected that this is not a fringe position anymore. It is becoming the framing that the rest of the field will have to answer to.

Agriculture was not a side track

Agriculture got more dedicated session time than I have seen at any comparable summit. The framing was that agriculture is the test case for whether AI can deliver in the Global South, because the agricultural workforce in lower-income economies is enormous, the existing infrastructure is uneven, and the gap between what an applied AI could do and what current practice delivers is wide.

A few specific threads from the agricultural sessions.

The data gap came up first and kept coming up. Multiple speakers from African and South American delegations made the case that agricultural AI built on Western corpora does not transfer to their contexts. The crops are different, the pest profiles are different, the language structures are different, the operational realities are different. The work of building local corpora (in the major languages of the region, with regional taxonomy, with provenance) was identified as the precondition for everything else, and the bottleneck.

On deployment, the speakers were remarkably aligned: the binding constraint is not the model but the deployment infrastructure, the local-language voice layer, the integration with the cooperative or distributor or extension network that already reaches farmers. The teams who get into production are the ones who have invested in this layer, and most of the AI-pilot-graveyard stories that came up in side conversations were teams who underinvested in it.

And then there was compute, which took up more oxygen than I think it deserved, and which is where I found myself disagreeing with much of the room. GPU access in the Global South is genuinely painful (roughly an order of magnitude more expensive than in the US and Europe at the same notional pricing, once you account for availability, contracting friction, and the latency cost of routing inference across continents), and several sessions sketched what a Global-South-friendly compute layer would look like, a thing nobody has built yet. But a compute-first framing imports the frontier lab’s worldview into a context where it does not hold. We trained and shipped a working agricultural model for India this year, and at no point was compute the thing keeping me up at night. What kept me up was the corpus, the voice layer, and the last kilometre of deployment through people who already hold the farmer’s trust. A summit that budgeted its sessions by actual constraint would have given the compute slot to extension networks.

My read after three days

The Global South framing is going to reshape AI policy more than people expect. The G20 had already moved in this direction. GPAI just amplified it. Over the next two years, the international AI safety, evaluation, and access conversations are going to be substantially harder for the frontier-model labs to navigate if they cannot show meaningful work outside their home markets. The pressure is going to be on the frontier labs to either build out, or partner with, the Global South ecosystem in a more sustained way than they have to date.

The Global-South-friendly outcomes are likely to happen in domain-specific AI first. General-purpose AI access is gated on infrastructure that the Global South does not yet have, while domain-specific applied AI in agriculture, healthcare, education, and financial inclusion can ship value with much less infrastructure overhead. The per-query value is high and the deployment shape is targeted. The teams who get this right in agriculture this year will be the case studies that other domains build on.

And the data and corpus work remains the most underfunded part of the stack. Foundation models get funded. Application UI gets funded. The middle layer, the curated, region-aware, language-correct corpus that makes the model useful in context, is everybody’s problem and nobody’s headline. The institutions that step up to fund this layer in a sustained way will shape what is possible.

Dhenu at the summit

We were at GPAI as part of the Microsoft for Startups pavilion. We were also there as collaborators with Sarvam AI and NimbleBox.ai, who helped with the Dhenu 1.0 training and corpus pipeline. Dhenu 1.0 was launched at the summit; Lokesh’s technical post on the release went up two days ago.

Being able to ship a launch into a context where the audience was already primed for the Global-South framing was useful. The conversations at the booth were structurally different from the conversations at, say, a US-centric AI conference. People wanted to talk about deployment, about cost-per-query, about the realities of farmer access, and far less about model architecture. Which is the right set of questions to be asking.

Some thanks are owed. To the GPAI host country team, for shaping a summit that took the Global South framing seriously instead of treating it as a side track. To Sarvam and NimbleBox, for the work on Dhenu 1.0 that we were able to demonstrate this week. And to the agricultural delegations from across the Global South who shared what they are building; the international agricultural AI ecosystem is more connected after this week in Delhi than it was before.

The next stop for us is turning the launch into deployments, through partners who already reach farmers. Delhi in December turned out to be a good place to start those conversations.