AI from the Margins: The Namibian Hydroponics Helper

by Leo Gaggl • WOL Notes

The dominant narrative surrounding artificial intelligence is one of extreme concentration. We are told that AI is a race that can only be run by a handful of multi-billion dollar companies in California or Seattle, using massive server farms that consume megawatts of power and millions of litres of water. In this frame, the rest of the world is reduced to mere consumers of foreign models, source material for scraping pipelines, or low-paid data labelling labour.

But this concentration is a political choice, not a technical requirement.

From the margins of the global digital economy, communities are demonstrating that technology can be built differently. The clearest example I’ve found is Namibia, where university researchers, and a national AI policy still being written, are placing local knowledge and community governance at the heart of machine learning.

It is an approach that flips the script on technological development. It shows that AI can serve community dignity rather than corporate extraction.

I pay attention to Namibia’s strategy for practical reasons, not just admiration from a distance. I’ve written before about turning a bare, compacted Adelaide backyard into something that grows bananas, and the lesson underneath that applies here at a different scale: closely observed local conditions, water, soil, microclimate, beat generic advice every time, whether the observer is one person with a handful of dirt or a research team building Ham-AI and Xaru-AI alongside the farmers who’ll use them. The data commons work I am involved in runs on the same basic premise at a larger scale: that a community generating environmental or agronomic data should govern that data itself, rather than handing it to whichever platform got there first. Namibia is starting to ask that question at the scale of a nation, and its researchers are answering it at the scale of a grow tent. Anyone working at the scale of a farm, a valley, or a backyard can learn from both.

Namibia’s Responsible AI Push

Namibia doesn’t have a finished national AI strategy yet, and it’s open about that. In August 2025 it launched an AI Readiness Assessment, built with UNESCO, that found no dedicated AI strategy or regulatory framework, low public awareness of AI, limited broadband and devices in rural areas, and a shortage of AI training data in Indigenous languages. A national strategy, drafted with input from 215 local experts, has since reached national validation and still has to go to Cabinet. In September 2026, the government opened a National AI Institute in Windhoek with responsible AI as its guiding principle.

For a country of three million people and around 30 languages, starting with an honest audit of the gaps, rather than importing ready-made models and governance templates from the West, is a choice worth noticing.

The more interesting work is happening well away from the policy drafts.

At the Namibia University of Science and Technology, researchers around its UNESCO Chair in Digital Technology Design with Indigenous People are working side by side with ovaHimba and San communities. They are not treating these communities as subjects to be studied or sources to be scraped. Instead, they are co-designing digital tools that combine traditional ecological knowledge with machine learning.

A basket weaver at work in Ongula, northern Namibia, by World Bank Photo Collection, on Flickr
"Basket weaver in Ongula" by Philip Schuler / World Bank Photo Collection (CC BY-NC-ND 2.0)

Ham-AI and Xaru-AI: Co-Designed Hydroponics

One practical result is a community-managed hydroponics project with the San community in Donkerbos, in the Omaheke region. In a dry country like Namibia, managing water and soil health is a matter of survival.

Rather than dropping in a commercial agtech platform that sells subscription-based advice, the researchers took a sensor-monitored hydroponic grow tent first built and tested with farmers in South Africa and co-designed its interface with the community. The dashboard was rebuilt around colours and icons the community chose themselves, covering light, temperature, humidity, water level, nutrients and compost, so the interface reads as something the community made, not something handed to them.

The community named the resulting AI assistants themselves: Ham-AI (meaning Sun AI) and Xaru-AI (meaning Food AI), chatbot advisors built to explain why the system is recommending something, not just issue the recommendation.

What makes this more than a co-design workshop that produced a nice-looking interface is what happened after the researchers left. In Donkerbos, Shorty Kandjengo, acting chief of the !Khuisi Traditional Community, took over one of the grow tents with no prior farming background. He did not just follow the AI’s recommendations. The tent draws on borehole water so mineral-rich that the nutrient readings start out too high for the algorithm’s defaults, so he overrode its dosing and topped up the nutrient solution by hand until the reading was right. He also reset the light schedule to spare an ageing solar system. That is the real test of “governed by the farmers themselves”: not whether the system launched well, but whether the person running it after the researchers have gone home feels entitled to argue with it, and wins.

The researchers are upfront about where it falls short. The two chat assistants run on OpenAI’s GPT-4o mini through its cloud API, chosen so the team could prototype quickly with the community, and small offline models are on the list for later. Beyond the assistants’ names, the app doesn’t yet speak the community’s language. Community-designed is not yet community-hosted, and that gap is exactly the argument for public compute further down.

The CARE Principles of Indigenous Data Sovereignty

The work in Namibia aligns with a global movement that rejects the extractivist logic of data scraping. The core framework for this movement is the CARE principles for Indigenous data sovereignty:

  • Collective benefit: Data ecosystems must be designed to enable communities to derive value from their own data.
  • Authority to control: Communities must have the right to determine how their data is collected, stored, and shared.
  • Responsibility: Those working with data have a duty to build relationships with the communities they represent.
  • Ethics: Data collection must prioritise the rights, wellbeing, and dignity of the people involved.

These principles challenge the assumption of individual consent that underpins Western privacy laws. They insist that data about a collective (a language, a soil system, or a shared watershed) must be governed by that collective. This is not a new insight invented for the data economy. Ecologist Fikret Berkes has spent decades documenting how traditional ecological knowledge is relational and situated by nature, bound up with the culture and practice that produced it, and stripped of much of its meaning when it gets flattened into a generic technical schema (Berkes, 2012). Any system trying to hold land, water, or community knowledge in a ledger runs into the same wall CARE describes for data.

When a technology company scrapes Indigenous cultural stories or agricultural records without consent to train a commercial model, it violates these principles. It encloses the knowledge commons. When a community co-designs a system like Ham-AI, they are maintaining the commons.

The Pattern Holds Elsewhere

Namibia is not the only place this shows up, and the second example matters precisely because it comes with a different kind of proof. In Bihar and elsewhere in India, Digital Green makes short instructional videos starring local farmers, on practices like the System of Rice Intensification, and screens them in village self-help group meetings where neighbours watch and discuss them together. The advice comes from farmers the audience knows, not from a distant platform.

The difference from the Namibian case is that Digital Green’s results have been tested by randomised trials. Researchers evaluated the video approach across 400 villages in Bihar, as summarised by J-PAL, and found yields up 12 to 18% and estimated profits up 9 to 24% after one year compared with conventional extension, with the effect smaller in the second year. (Measured adoption of SRI itself did not rise, a reminder that even a good trial leaves questions open.) IDinsight, working with Digital Green, ran a randomised trial of its newer FarmStack platform with more than 2,000 cashew farmers in Andhra Pradesh and found gains in knowledge and reported adoption of new practices. This is what it looks like when “farmer-to-farmer advice beats generic top-down advice” gets tested properly, rather than resting on a conference paper written by the same team that built the tool.

The Necessity of Public Compute

Those two GPT-4o mini assistants in Donkerbos point to the bottleneck facing any community-led technology effort: compute.

While software designs can be shared globally and data can be governed locally, the actual processing power required to run and refine machine learning models remains concentrated in the hands of a few cloud hyperscalers. To break this dependency, we need public compute infrastructure.

A public compute commons is the missing foundation. It means treating computational power as a public utility (like roads or water systems) funded by the public and governed for the public good. Without it, local initiatives remain dependent on the APIs of foreign monopolies, and the talent developed in the margins keeps draining to the financial centres of the global north.

None of this should be read as a solved problem. Namibia’s own readiness assessment names real, unresolved hurdles: limited broadband and devices outside the cities, low public awareness, no AI regulatory framework yet, and a shortage of training data in Indigenous languages. The draft strategy itself was built with support from UNESCO, UNDP and Germany’s Konrad-Adenauer-Stiftung. The hydroponics project shows the same pattern at ground level: when the tent in Donkerbos broke down, fixing it meant the research team talking the community leader through the repair remotely, and other farmers in the study took longer than Shorty to trust the system and asked for more control built into it. Community governance of a tool is not the same thing as community independence from the people who built it, not yet. A community assistant that calls a Californian model, and a national AI effort leaning on outside partners, have not escaped dependency. They have moved it around. Closing that second dependency is the harder, slower half of the work, and it is not finished.

The lesson from Namibia is that we do not have to accept technology as a force of enclosure. Even with limited resources, a community can shape tools around its own symbols, its own judgement, and its own soil. The future of technology is not a single corporate monoculture. It is a network of diverse, locally governed commons, co-designed by the people who use them.

Sources

←GrapheneOS in Practice: What Worked, What Didn't and What's Still Missing

Comments

Be the first to comment! Reply to this post from your Mastodon/Fediverse or Bluesky account, or mention this post's URL in your reply. Your comment will appear here automatically via webmention.

Follow this blog on Mastodon at @gaggl.com@web.brid.gy or on Bluesky at @gaggl.com