Data center proposals worth $64 billion have been blocked or delayed by strident civil society organizing across the United States in 2026. But the AI infrastructure boom is showing no signs of slowing down. So if not in the US, where are Big Tech companies taking this vital piece of the AI supply chain? Northeast Brazil has emerged as one lucrative data center construction hotspot. Attractive tax and utility subsidies aside, the region is also a wind and solar energy hub, allowing companies to effectively greenwash their data centers’ socio-environmental impacts. So what does the fight against data center expansion in northeast Brazil look like? 

As part of their work in our community of practice, Madhuri Karak and Cathy Richards are co-leading a series of events exploring the global socio-environmental impact of data centers. The first edition of the series took place in spring 2026, focusing on the environmental impacts of data centers in the US and the European Union, intending to unpack the development of sustainability metrics within the industry.

To continue our event series focusing on the environmental impacts of data centers, how they are being regulated and what possibilities exist for resistance, we are inviting Rhavena Madeira (of Filha do Sol, in Piauí, Brazil) and André Fernandes (of IP.rec, in Pernambuco, Brazil) for a virtual roundtable taking place on July 29, 2026, at 12pm ET/1pm GMT-3/6pm CEST. 

In Brazil, the federal government has been investing in data center expansion, all the while offering little evidence of what the country and its citizens stand to gain from this effort to lure investors. Though the antidemocratic practices of Big Tech companies pushing for data center expansion and the environmental impacts of these large, resource-intensive infrastructures have been pointed out by several civil society organizations, Brazil already has more than 160 data centers and is expected to receive billions in new investments in the sector. 

We are honoured to be joined by Rhavena Madeira, director at Filha do Sol; and André Fernandes, director at IP.rec, for a conversation about their work against data center expansion in northeast Brazil. Filha do Sol is an organization working in Piauí to uplift feminist and collaborative leadership on the frontlines to regenerate tropical nature and restore climate justice; while IP.rec is an independent center engaged with a focus on the social, ethical, and legal impacts of technological development. IP.rec has recently published this report about AI and the environmental impacts of data centers, as well as this paper on the lack of clarity and opacity that often surrounds conversations about the environment and artificial intelligence.

About the speakers:

Rhavena Madeira (she/her) is Co-founder and Strategy Director of Filha do Sol. She is a Brazilian lawyer and project manager with a master’s degree in Human Rights (University of Vienna). For the past ten years, Rhavena has been working to defend socio-environmental rights and a stable climate with the government and civil society. In addition to Filha do Sol, she is also the director of the Center for Climate Crime Analysis Brazil. Rhavena has accumulated experience in managing complex projects and building alliances to defend human rights and climate justice, traditional communities, and local and regional organizations that live and defend their communities in threatened ecosystems, especially in the Amazon and the Brazilian Northeast.

André Fernandes (he/him) is Founder and Director of IP.rec. André is an Attorney and Professor in Graduate Programs in Law and Technology at UFPE, UPE, and CESAR School. He holds a PhD in Law, with a focus on Artificial Intelligence, the History of Legal Concepts, and Innovation, and a Master of Laws from the Federal University of Pernambuco, with a focus on Legal Theory. He specialized in Artificial Intelligence at the Amsterdam Law and Technology Institute (ALTI) & Vrije University Amsterdam, in AI Policy from the Center for AI and Digital Policy (CAIDP), and is certified in Internet Governance by the Brazilian Internet Steering Committee (CGI.br), DiGI (UCU) and SSIG.

Starting October 1 2026, anyone using WhatsApp as a platform to reach their target audience will be charged by Meta for outbound messages. Until that date, it will have been free to do so within a 24-hour window opened by the user’s last message, meaning that as long as your chatbot (or human agent) was engaging and helpful enough, users seeking support for health, livelihoods, education and psycho-social support could chat for as long as they wanted to (and could afford to themselves). All implementers had to pay for was the cost of the WhatsApp Business Service Provider (BSP) platform, such as Turn.io, as well as any re-engagement messages to wake up ‘dormant’ users. In the near future, any message sent in response to users will be charged at the same rate as ‘utility templates’, whose costs vary region by region.This change has huge scale, sustainability, user experience and impact implications for many social impact organisations using, or thinking about using, WhatsApp to reach people across the Global South.

How WhatsApp became the channel of choice

WhatsApp became the digital intervention channel of choice starting in 2018, when its API was released, allowing many of us to seize the opportunity to reach our target audience on the channels they actually use. I was one of very few people designing chatbots back then, experimenting with chatbot copy-writing, design conventions, and figuring out the best way to leverage a fairly restrictive format into a rich experience. Sure, the ‘chatbots’ we created were pretty basic, relying mostly on text, pre-determined decision-trees, and of course, emojis. But they allowed even small NGOs to reach high volumes of users with potentially life-changing information, advice, and support, on an app that was often 0-rated by MNOs, much less prone to connectivity issues than mobile web, and so embedded in users’ lives that it was less likely to be deleted from their phones to preserve memory or battery power. Crucially, users trusted them, loving the human-not-human chat-based format.

The pandemic turbo-charged the use of WhatsApp by development and humanitarian practitioners, including multilaterals like WHO who deployed chatbots on WhatsApp and other messaging apps to millions of users in over 19 languages, with the help of Turn.io and organisations like Reach Digital Health. At a smaller scale, civil society organisations like the International Fact Checking Network, created a chatbot allowing users to submit health misinformation and get it fact checked in minutes. Today, WhatsApp usage represents 69% of all internet usage worldwide, and 83% of monthly users open it every day. WhatsApp chatbots now feel like old news in the development space ( although we’re still trying to figure out how to measure their impact.)

GenAI allows chatbots to talk more; Meta pricing is forcing them to talk less

Whilst launching a chatbot for social impact no longer feels like a novelty, WhatsApp chatbot fever is in no-way diminished. GenAI means that we can finally create *actually* smart chatbots, either run entirely using a Large Language Model, or which include genA-powered components to handle specific conversations requiring more personalisation and flexibility. 

What this means is that at the exact moment that we are being put under pressure to use genAI to make chatbot experiences richer, more personalised, more flexible, and more accurate, Meta’s new pricing model is forcing us to do the opposite. Turn’s advice to customers using their platform is to ‘reduce the amount of back-and-forth and get users to what they’re looking for as quickly as possible.’ They position this as a user experience improvement as well as a cost-reduction method, and in a general sense, they’re not wrong – many chatbots could do better at being succinct and reducing their feature bloat. But it’s also at odds with the parallel push towards AI-powered, conversational experiences – whose underlying agents will now have to be prompted to be ‘less chatty’, when a big reason for their integration was exactly this. 

What this will actually cost implementers

Those with existing WhatsApp services are scrambling to calculate the implications. Informal estimates circulating among implementers range from $500 extra per month, to $4,500 per month to engage with around 10,000 users a month. As one implementer shared: “this is a major constraint to scaling plans and sustainability projections.” For smaller NGOs, what was a cost-effective way of reaching high-volumes of individuals in their target audience with valuable messaging and support, will become prohibitively expensive.

Examples of projected monthly cost increases shared by chatbot implementers

On the other hand, because exact costs vary per country, some implementers are less concerned. Similarly, for chatbots where the service offering is less dense, and less conversational, the impact may be less significant. As one implementer shared: “I am personally unconcerned. This is because in India, our charges are going from 2 US cents to 7 US cents. Our per-participant cost overall will remain nothing for us.” (their emphases). Additionally, where users come from in the first place matters. Anyone who relies on Meta ads which click-to-WhatsApp for the majority of their marketing will still be able to message users for free within a 72 hour window of the first conversation started. Given that average user lifetime is often only a day, those relying on Meta ads may not be affected at all: basically, don’t panic until you’ve done the maths for your particular set-up.

The human cost

Perhaps more seriously, this pricing change may also impact on the role of humans in increasingly automated services. WhatsApp is also used to support human-staffed Helpdesks, with trained operators able to jump in and out of an automated conversation to provide human-led support, including last-mile service provision referrals. These agents are central to safeguarding processes, with users expressing suicidal ideation, survivors of Gender Based Violence, or women going through miscarriages referred as soon as possible to human support. Let’s imagine for a second how it would feel to tentatively disclose spousal abuse, or the loss of a baby, when our interlocutor wants to wrap up the conversation as quickly as possible for cost reasons. 

How we can push back

As we speak, implementers leveraging WhatsApp, and the integration platforms who allow them to do so, are discussing ways to fight back, for example by demanding clarity and seeking exemptions for non-profits from Meta; forming a coalition; engaging with country-level regulators; or convincing donors to supplement budgets to cover the unexpected surplus. Whether or not they prevail remains to be seen – the social sector has sought for years with limited success to get Meta to make changes that would enable us to use WhatsApp more effectively in the contexts in which we work. Our collective bargaining power also ebbs away with every cut to international aid. In the meantime, here are some things you can do, right now.

What you can do – as a donor

• Get to grips with the cost implications for any organisation in your portfolio using WhatsApp currently and do what you can do to enable them to absorb this unforeseen cost without compromising on service quality.
• Communicate up the chain about the severity of this situation and the longer term ramifications. Tap into your relationships with regulators, governments and big tech to see how you can make a difference.
• Make sure anyone proposing to use WhatsApp as an intervention channel is aware of the attendant costs, and can meet them as their intervention scales.
• Stop expecting too much from a single WhatsApp intervention, unless you can provide the budget to back it up. More features = more back & forth = more budget.
• Consider the implications for performance and impact indicators. In an age of endless, free, back and forth, conversation length was a measure of success. Not any more. Evaluators will have to disentangle cost-driven user behaviours from organic ones.

What you can do – as an implementer

If you already have a WhatsApp chatbot:

1. Do the maths and figure out how it will actually impact your budget.
2. As Turn suggests, this is a good opportunity to revisit your User Experience, without compromising on quality or safety. Get help from people with conversation design and UX expertise (see here, here, here and here for staters).
3. Figure out if you really need that shiny new AI integration.
4. Talk with your users. Your budget challenges are, of course, not their problem, but talking openly about the implications on the service you provide them could yield valuable insights and unexpected ways forward.

If you don’t have a WhatsApp chatbot, but were planning one:

1. Revisit Lean Design first principles and think about what Here I Am calls Minimum Viable Value. What is the one thing that would provide immediate, tangible value, fast, for your users? Build that. And only that, for now. (P.S: we should all be doing that anyway).
2. Explore other channels. Anyone remember Freebasics? Tom from mySpace? Nothing lasts for ever. Especially for those working with young people, there is almost certainly a platform out there, like Moya Messenger, which will allow you to reach the same volumes of users without the conversational restrictions. 

*

This isn’t the end for WhatsApp chatbots, but it’s yet another lesson about relying on corporate behemoths to run essential services. Something to think about before you make GPT, Llama, Gemini or Claude central to your entire intervention – but you know that already. Given their substantial profitability challenges to date, I would bet that the same thing will happen with genAI before too long. 

In a recent event hosted by MERL Tech’s Sandbox Working Group, Alyssa Young and Xian Ho of Dimagi walked through several years of work building, testing, and deploying generative AI chatbots for family planning and sexual reproductive health in Kenya and Senegal.

The project, Open Chat Studio started back before the arrival of ChatGPT as a way to build simple, menu-based chatbots. Once GenAI was in the picture, the team shifted to a different approach. They realized their team, as well as implementing partners, needed a way to create their own GenAI chatbots, so they built OCS and then open sourced it as a digital public good with support from the Gates Foundation.

OCS sits in between the user and large language model. It’s agnostic to which LLM is being used, whether it’s Anthropic or OpenAI, other providers or open source models found on Hugging Face. It supports a number of channels, including WhatsApp, Facebook Messenger and Telegram. By sitting in the middle, Dimagi can help orchestrate how the different models are responding to user input.

The session covered some of the important lessons learned during the development of GenAI chatbots in Kenya and Senegal by Dimagi and two organizations working on Adolescent and Youth Sexual and Reproductive Health. 

In Kenya, the Dimagi team worked with Shujaaz Inc, a Nairobi-based youth brand whose comics, radio, and social channels reach a majority of adolescent girls and young women in the country each month to build the Tubonge chatbot which features two older characters with an existing combined following of nearly 5 million users, were selected through co-design practices as the voices to share the messaging.  

In Senegal, they partnered with RAES, a Dakar based NGO behind the edutainment series C’est la vie! which already had around 1 million followers across Facebook, Instagram, YouTube and Tiktok. There the chatbot takes on the personality of Marguerite, a midwife, and Dr. Moulaye, a health specialist.

Key takeaways to consider in chatbot design:

1. Focus on safety more than the conversation

When a user messages one of the Kenya bots, the query doesn’t go straight to a conversational model. First, it gets filtered by a crisis classifier which is built to tell the difference between a joke in slang and a genuine cry for help. If it detects an active crisis, it triggers two events:

• A silent alert to the research team for follow-up
• Routes the users to a referral with urgent hotline numbers and care options

If a message is deemed safe, only then will they be sent to the main router which uses chain-of-thought reasoning using the current message and chat history to decide which of the specialized nodes is best prepared to develop a response. The nodes typically include: an advisor (a neutral “probe-then-solve” conversationalist), a referral node (which pulls verified clinics), a roleplay node, a quiz node, and a Shujaaz brand-logistics node. Underneath both instances is retrieval-augmented generation (RAG) drawing only on validated local source material (national health policies, partner content, curated databases, etc)

2. Listen to users

Roleplay and quiz features weren’t on the original spec. They came directly out of co-design sessions with young people. Dimagi found that roleplaying gave users a safe simulation space to rehearse hard conversations such as those with a strict parent, a difficult partner, or a judgemental nurse. 

Users also made requests for sexual and reproductive health (SRH) and mental health content for boys and men as well as offered feedback on the quality of local languages and ways to interact with the chatbot. 

3. Make sure there is a human in the loop

Deployment of this chatbot has not been a case of ‘set it and forget it.’ Transcripts are reviewed daily, with a prioritization mechanism that flags conversations needing closer follow-up and a defined human-escalation plan. Dimagi maintains chatbot managers, and that role transfers to local partners as the tools are handed over. Because SRH is sensitive and the technology is new, live implementation includes a safety disclaimer and consent flow, with an age requirement (18 and up) at the front door. As the team shared, the goal is to support not replace human referrals and relationships.

4. Building it is easy, proving it’s useful is hard

Dr. Ho’s closing reflection provided the project’s biggest takeaway. Developing an LLM chatbot is simple, but building a genuinely useful one (and keeping it that way) is hard. To mitigate that, the team developed a four part evaluation framework which they use continuously.

Measuring performance of GenAI Chatbots

 It includes:

• Regular user testing which measured acceptability and experience through a validated questionnaire. 
• Expert evaluations with study-team reviewers rating query-and-response pairs on key metrics including accuracy, acceptability, safety, authenticity, empathy, and language quality.
• Automated evaluations using another LLM as a judge to acquire those same metrics more quickly and widely
• Head to head comparisons which compared the custom bots to regular ‘vanilla’ bots in blind tests.

A resource to further explore: 

Open Chat Studio, Dimagi’s open-source platform for building, testing, and deploying LLM-based chatbots with built-in guardrails. Learn more at sites.dimagi.com/open-chat-studio or browse the code at github.com/dimagi/open-chat-studio.

All roads led to Nairobi from the 20th to 22nd of May 2026 for the ICT4D conference, where sessions centred on the theme “Delivering impact through digital transformation, together.” The MERL Tech Initiative (MTI) was busy leading some amazing sessions as a consortium partner. We led panels on evaluating GenAI challenge funds, resourcing African language NLP for MERL, joined a session on Voice AI, and facilitated a workshop on Competencies for Made in Africa AI in MERL. 

The week was packed with rich conversations and impressive technology demonstrations. It was fascinating to see how people are advancing geospatial intelligence and climate early warning systems, while still keeping community farmers at the centre of the work. I had stimulating conversations around how technology can support mental health, with some genuinely thought-provoking ideas emerging from those discussions. The innovations on display during the tech demonstrations were truly standout, each offering a fresh and purposeful approach to solving some of  Africa’s problems. Particularly eye-opening was learning about deeper strategies for safety by design, especially around gendered considerations, and how embedding these principles can actually give consumer-facing technology a meaningful competitive edge.

In this post, I want to share my reflections of the entire experience and some key learning and insights. 

Minimum Viable Intelligence: Africa’s Existing Asset 

A lightning talk by Nasubo Ongoma of Qhala was one of the highlights of the event. She posed what may be the most critical question facing the continent: with Africa accounting for less than 1% of global compute, how can we realistically expect to compete with the Global North? More importantly, where do we even begin to build? 

She reinforced the persistent gap in African language datasets, noting that much is lost in translation when English continues to serve as the default language in AI models. Her argument is that to build AI for Africa, we must first understand and respect the constraints we are operating within, and design around them. If we do this well, we may actually find a way to create and use AI in Africa in a way that is distinctly and powerfully our own.

Central to this is the nuance of how we can meaningfully layer in existing knowledge systems and cultural frameworks as we build. For technology to truly work in Africa, it cannot be imposed from the outside; community practices and indigenous knowledge must be treated as valid knowledge systems in the development process. AI development in Africa must come with caution, to avoid repeating past mistakes and reinforcing the existing inequalities that place Africa in a perpetual adopter position. 

Building for and with “community” requires formative research and responsible digital design 

“Community” was a common and revolving word at this conference. Yet, I kept finding myself in sessions that questioned, reflected on, and interrogated the extent to which communities are actually involved in the technologies built in their name. Practitioners are recognising a critical gap: designing for the sake of technology, rather than designing with the people it is meant to serve. The innovation boom has accelerated this disconnect, prioritising novelty and scalability over meaningful community engagement. 

A fundamental question remains largely unanswered: how do we actually measure community engagement in tech? Without clear metrics, “community” risks becoming performative, a buzzword dressed up as a value. This is precisely where formative research becomes a valuable asset. Formative research and human- and community-centred design processes aim to ensure that community voices are not an afterthought but a foundation that shapes not just what is built, but the why, for whom, and to what end.

Data Colonialism 

While it was encouraging that the organizers elevated this topic to a main agenda item, the execution left much to be desired. I found it deeply problematic for conversations on decolonizing data in the Global Majority to be led by individuals from Silicon Valley and the US. Chief among the issues worth naming explicitly is the positioning of “my experience working in Africa, with Africans, or with African organizations” as a sufficient basis for authority. 

This framing is problematic because it flattens and renders invisible the African perspective, reducing a continent of immense diversity to a backdrop for external expertise. Our data has been extracted for decades, and now even the story of that extraction, i.e. its meaning, its harm, its trajectory; is being told on our behalf.

This is what left a bitter taste at ICT4D for me. It raised an uncomfortable but necessary question: why were leading African scholars and practitioners with lived experience not at the center of this conversation? Local voices and in particular, Black African women, feminist scholars, and human rights organizations have been theorizing, documenting, and challenging these dynamics for years. It is only right, then, that the owners of the story be the ones to tell it.

Building on top of Big Tech models: The need to decentralize power as Africa builds 

The work of developing AI in Africa cannot happen without directly addressing the issue of power dynamics. It was apparent from the conversations that big tech models such as Open AI’s GPT, Anthropic’s Claude, and Google DeepMind remain the foundational base upon which Africa-focused models are built. While the development of more localized base models in African languages is underway, it is important to continue asking: what alternative models exist, and how are they being built? 

Equally critical is the question of data ownership and governance — who is collecting data, who controls it, and is it flowing back to the communities that produced it? Community hesitancy and resistance to data leaving the continent reflects a deeper, legitimate concern about extractive practices that have long characterized Africa’s relationship with outside institutions. 

From the questions raised in various discussions it was a stark reality that vast amounts of institutional knowledge exist in low-resourced African languages, yet many communities lack the technical capacity and resources to transform that knowledge into structured, local-language datasets. 

What is needed, therefore, is a deliberate decentralization of power, away from big tech and toward smaller organizations, community-led initiatives, and sovereign language models. Alongside this decentralization must come a re-centralization of collaboration: among African institutions, across the Global South, and with aligned partners globally who are committed to equitable AI development. This requires intentional cooperation, shared infrastructure, and a collective insistence that the terms of AI development on the continent are set and led by Africans themselves.

The Future of Digital Development: Toward African-Led, Locally Grounded AI

The threads running through ICT4D 2026 converge on a single, uncomfortable truth: localized, Africa led, Africa owned efforts are the only way to reap AI benefits that develop the continent and benefit Africans. Mobile Network Operators (MNOs) offer a rare entry point to this dynamic. They are advancing localization agendas because their commercial viability depends on it. The reach of MNO infrastructure into underserved communities combined with their investment in local language services positions MNOs as consequential actors in the AI localization story. 

The lesson here is not that MNOs are the solution, but that localization follows where incentives are aligned, and MNOs are strategic partners and enablers for the pathway of locally grounded AI. The hard work here will be ensuring that this localization serves communities and not merely markets – a distinction that matters enormously.

What is clear is that no amount of community-centered design, sovereign data governance, or African-language AI will gain sustainable ground without an enabling environment at the policy level. The conversations at ICT4D made clear that the public sector holds significant leverage over the terms on which AI enters African contexts through regulation, procurement, data policy, and infrastructure investment. This makes governments an important partner for scaling digital development that is accountable, equitable, and durable. Yet engaging them is rarely straightforward. High turnover in key positions, limited AI expertise, and stretched institutional capacity mean that knowledge and momentum can be lost quickly. Election cycles also bring shifting priorities, and AI itself is increasingly becoming a political issue. These realities don’t make government engagement optional, they make it more necessary to approach governments with strategy.

Africans are the future of digital development in Africa

ICT4D 2026 ultimately surfaced that the tools, the talent, the cognisance and the urgency exist. From the question of who tells the story of data colonialism, to who governs the datasets, to who sets the terms of AI development on the continent, the answer must increasingly and unapologetically be Africans themselves. The future of digital development in Africa will be determined by the strength of the ecosystems, institutions, and communities who are leading these efforts. And that is precisely why this moment feels like a blueprint.

The AI inevitability narrative would have us believe that the path forward is already written, with Africa’s role is simply to adopt, adapt, and keep up.  What emerged from Nairobi is that the African ecosystem is actively choosing the terms of its own digital future and building blocks of an alternative path, one that refuses extraction. Insistence on indigenous knowledge systems, community-centered design, sovereign data governance, African language models, and African-led policy conversations, are not just good practices, they are acts of resistance. If Africa can hold this line, investing in local ecosystems, demanding accountability from big tech, and keeping communities at the heart of innovation, then yes, Africa’s response is the blueprint.

AI Use Disclosure: The blog utilized Claude Sonnet 4.6 for copyediting, grammar and syntax. The copyedited content was reviewed thoroughly and further edited by the author. The content remains the author’s original ideas and reflects the author’s thoughts and style of writing.

The current conversation around artificial intelligence in the Global Majority has, for too long, been framed as a race; a race centred on the rapidly evolving nature of AI and emerging technologies; a race premised on AI’s inevitability and the urgent need for the Global Majority to catch up. This urgency is largely skewed toward regulation and adoption.

On 25 May 2026, the MERL Tech Initiative’s Community of Practice hosted a conversation on AI autonomy, regionalization, and ethics, led by Chenai Chair, Director at Masakhane African Languages Hub, Wayan Vota, ICT4D strategist and founder of ICTworks, William Tjhi, Deputy Director, AI Products at AI Singapore and Maria Luciano, Tech Policy Analyst. The discussants unpacked what “AI sovereignty”  means for them and who it truly serves. They debated whether it can center people’s rights and well-being, whether it fuels a harmful global AI race, and whether it genuinely challenges Big Tech’s dominance or quietly reinforces it. They also questioned whether locally built AI can serve as a real alternative to giant tech systems, and what sovereignty could offer countries in the Global Majority as a path toward greater independence in knowledge and technology.

This blog captures the key threads of that dialogue and explains why they matter.

Why This Conversation Matters for Global Majority Digital Futures

A narrative circulates across development, policy, and technology spaces: AI adoption is simply the next logical step, and you either get on board or get left behind. Maria pushed back on this framing directly, noting that when people are told there is no choice to be made, the question of accountability disappears. If the AI race is inevitable and adoption is already decided, then the communities most affected never sat at the table where that decision was made.

This conversation emphasized how narratives shape technological realities. The speakers drew a sharp line among narrative-level framing, the legal and regulatory frameworks that narratives produce, and the real-world infrastructure that ultimately enforces or undermines those policies. When the dominant narrative is “AI is inevitable,” resulting policies tend to treat inclusion as a footnote, and the infrastructure built to enforce those policies carries the same blind spots. Accountability does not disappear in a single event; it erodes layer by layer, beginning with how we talk about technology in the first place. (Read more from Maria on this.)

Communities across the Global Majority already experience the downstream effects of AI systems built without their input, trained on data extracted without their consent, and deployed in languages never designed to carry the weight of their realities. As Chenai noted, any technology that affects how people access information or exercise their social and cultural rights must also be able to connect with them on their own terms. The choice of language, the choice of modality, the choice to engage or not, these are not technical preferences but foundational design choices. The central question, therefore, is not whether AI is inevitable, but who determines its terms. That question must be present throughout the Global Majority AI conversation: not only at the implementation stage, but at the levels of narrative and design. 

The Sovereignty Narrative: Local Models Versus Big Tech

The framing of sovereignty must be honest about what is actually achievable and where real leverage lies. The speakers moved the conversation on AI sovereignty critically beyond its most marketable framing: competing with hyperscalers on foundation model development is not realistic. However, governments and communities can build with genuine purpose and genuine control in targeted areas.

Wayan put the numbers plainly. Hyperscalers in the United States are projected to spend approximately $700 billion in capital expenditure in 2026 alone, with projections exceeding $1 trillion over the next three years. By comparison, the European Union has announced a $2 billion AI investment, Canada has committed $2 billion to a sovereign AI strategy, and India has allocated $1.5 billion to its AI mission. Even the most well-resourced governments in the world cannot compete with these hyperscalers at their own game. For governments in the Global Majority, the gap represents an entirely different category of problem.

What made this discussion particularly sharp was the observation that the loudest voices promoting sovereign AI are often those with the most to gain from governments attempting to build it. When Nvidia, the world’s largest chip manufacturer, tours the world’s capitals telling each government they need their own sovereign AI, and that same manufacturer supplies the chips required to build it, the alignment of interests deserves to be interrogated. 

An alternative exists in small language models. Small language models optimized for specific community needs, edge computing solutions that operate on devices without expensive cloud infrastructure, and domain-specific models in sectors such as health, agriculture, or legal aid represent areas where hyperscalers hold no particular advantage and, frankly, limited incentive to invest. These may be precisely where autonomous and sovereign AI can take root.

Realistic sovereignty, then, looks like this: build what you can control, contribute where you can influence, and be honest about where you cannot compete. Governments should direct their energy toward what is within reach, thoughtful procurement policy, domestic data governance, investment in local infrastructure, and support for community-led language initiatives, rather than expensive sovereign compute investments that primarily benefit hardware vendors. The conversation also surfaced what sovereignty discussions too often omit entirely: people. Whether the lens is state security, economic competition, or cultural and linguistic preservation, communities—particularly marginalized ones—are rarely the subjects of these conversations.

Autonomy Is Not Sovereignty: The Practical Distinction

One of the most clarifying moments in the discussion was the framing of autonomy as distinct from sovereignty in application. The distinction is not limited to definition; it lies in what you build, how you resource it, and who you design it for. Sovereignty, as the speakers explored, anchors its ambitions at the level of the state: national compute, national models, national data governance. Autonomy addresses a more immediate and survivable question: if access to an external system disappears tomorrow, can you still function?

William offered a practical framework for thinking about whether governments and institutions should build or buy AI capability. The buy-first-then-build logic was endorsed as a reasonable general approach, map your people’s needs through what already exists, then develop targeted domestic capability for what is not being served. The timing of the shift from buying to building matters enormously and varies by context; for some countries and communities, waiting too long means losing the generation of talent and institutional knowledge needed to build at all. (See the full paper here.)

The goal of building domestic AI capability is not purity of origin but negotiating power. Chenai described the reality of depending on compute infrastructure provided by large technology companies under time-limited agreements. When that agreement ends, the question becomes: what do you have left? If the answer is nothing, then every product built, every community served, and every promise made was contingent on a boardroom decision made without your presence. Autonomy, in this framework, means having enough capability to survive an interruption, even if you choose not to exercise that capability every day. This has direct implications for how organizations think about community guardrails and AI safety in practice. Well-intentioned guardrails tend to protect those closest to the infrastructure. If AI autonomy is to offer meaningful protection for the most marginalized communities, the design of those guardrails must explicitly account for who sits furthest from the system and what reaching them would actually require, in terms of connectivity, cost, language, and trust.

A hybrid approach emerged as the clearest practical recommendation: work with multiple AI providers, maintain core capability in areas of genuine community need, and build collaborative infrastructure with others who share your goals. No single country or community organization can do this alone. The future of autonomous, community-serving AI in the Global Majority will be built through deliberate cooperation, across institutions, across the Global South, and with aligned global partners.

The Outlook: Dreaming Carefully, Building Durably

Several threads converged as the conversation turned to future outlooks.

First, there was a consistent return to the value of focusing on what is within reach. A well-built, community-governed model in a specific language and domain can serve its people with a depth and responsiveness that no general-purpose model will ever prioritize. Hyperscalers will not build a model that captures the slang, dialect, or specificity of a language spoken by smaller populations, but the people who speak that language can.

Second, the call for locally defined, alternative evaluation metrics was emphasized. If Global Majority communities develop assessment techniques for their own AI systems that are systematic, not prohibitively expensive, and rooted in what actually matters locally, they gain a form of influence that does not require competing at the infrastructure level. Communities that can assess what works and what does not, by their own standards, change the terms of the conversation.

Third, the technical complexity of AI is not neutral. Keeping the language of AI governance inaccessible to non-specialists is itself a political choice, a way of ensuring that those most affected by these systems remain outside the room where decisions are made. Education and literacy, not just digital literacy but narrative literacy, the capacity to recognize and contest the stories being told about technology, are as important as compute or data in determining who benefits from AI’s development.

Finally, environmental sustainability cannot keep being deferred. The Global Majority’s natural resources and human labor are fueling the AI boom, often at the risk of further marginalization and for the benefit of Big Tech. The infrastructure required to build and maintain large AI systems demands enormous resources, frequently extracted in ways that replicate the very patterns of exploitation that communities in the Global Majority are working to escape. A vision of AI that is sovereign and community-serving but environmentally catastrophic is not one worth building toward.

Closing Reflection

Collaboration among institutions in the Global Majority, as we continue to interrogate autonomy and sovereignty, is a necessary first step: collectively naming what is at stake and pushing back on narratives that would reduce communities to perpetual adopters of Big Tech’s outputs. Strategizing collectively and charting alternative paths to reclaiming agency in AI is a practical starting point. It may one day produce shared language, shared values, and shared pushback strategies powerful enough to have communities shape how technology is built, countering the mainstream narrative that digital futures depend on Big Tech.

AI Use Disclosure: The blog utilized Claude Sonnet 4.6 for copyediting, grammar and syntax. The copyedited content was reviewed thoroughly and further edited by the author. The content remains the author’s original ideas and reflects the author’s thoughts and style of writing. 

We’re excited to take the conversation about Made in Africa AI in MERL from theory to practice and we want you to help shape what comes next.

On July 14, 2026, at 11am ET / 4pm BST / 5pm CAT / 6pm EAT, the Africa AI Learning Group at the Natural Language Processing Community of Practice (NLP-CoP) will convene a special World Café working session.

Since publishing the Made in Africa AI in MERL Landscape Study in 2025, we’ve been in ongoing conversation with working group members about what it actually means to use AI in evaluation practice in ways that are grounded, ethical, and locally owned. Most recently, we facilitated a workshop at ICT4D on the competencies required to operationalize Made in Africa AI approaches within evaluation practice. Building on this momentum, we’re ready to take the next step and build a framework together.

What is this framework, and why does it matter to MERL practitioners in Africa?

This will be a practical tool and shared reference point that helps practitioners navigate questions like: How do I integrate AI into my MERL work in a way that reflects African values and contexts? What should I be able to do, and what should I expect from the systems and institutions I work with? Currently, those questions don’t have a shared, agreed-upon answer, and this session is where we start working towards building shared agreements.

What to expect

There are no formal speakers. No presentations to sit through. Instead, participants will work in small groups with peers from across the continent, debating, co-creating, and stress-testing ideas together. The session centers on values and principles: what should a Made in Africa AI in MERL framework actually stand for, and what competencies, i.e technical, ethical, governance-related, and contextual, do evaluators need to put those commitments into practice?

By the end of the session, participants will have 

• Contributed and shaped a  first draft of a Made in Africa AI in MERL Framework
• Clarity on the competencies that will keep them competitive and grounded as an evaluator
• New connections with peers who are grappling with the same questions

We invite you to be part of this important conversation. Anyone who wants to be part of shaping how AI is used in African evaluation practice is welcome to be a part of this. Whether you bring deep technical knowledge, years of evaluation experience, or simply a strong conviction that this needs to get right, we would love to have you participate in this co-creation session. 

Sign up via the link below to join us for this interesting conversation.

This 2-day free in-person workshop will equip participants to think critically about AI, evaluate where it genuinely adds value to MERL workflows, and apply it responsibly. We will cover two key aspects: 1) how to apply AI to support tasks along the MERL lifecycle, and 2) how to conduct MERL of AI-enabled efforts (such as chatbots in health or agriculture).

Moving beyond theory, participants will leave with hands-on experience using real AI tools, a sharper understanding of the risks, and a clear ethical framework to guide their decisions. 

Apply here to join us on August 17-18, 2026, in Dar Es Salaam, Tanzania.  Applications close on June 30, 2026.

Who is this course for?

This workshop is designed by The MERL Tech Initiative for professionals who work in monitoring, evaluation, research, and learning (MERL) roles across the international development, humanitarian, and social impact sectors. It is particularly relevant for 

• M&E officers, advisors, and managers seeking to understand AI’s practical relevance to their work
• MERL leads and directors responsible for guiding organisational data and learning strategies 
• programme staff who commission or use evaluation findings 
• researchers and data analysts exploring AI-assisted approaches
• consultants or independent evaluators wanting to stay ahead of sector trends.

Level of AI Expertise: Beginner to Intermediate 

Level of MERL Expertise: Intermediate

What Participants Will Learn

By the end of this workshop, participants will be able to:

• Define key AI concepts and distinguish between different types of AI relevant to MERL
• Critically evaluate AI tools and assess whether they are appropriate for specific tasks and contexts
• Identify practical applications of Generative AI across the MERL cycle 
• Apply prompt engineering techniques to extract reliable, useful outputs from AI systems
• Recognise the risks of AI adoption, including bias, hallucination, data privacy, and dependency, and apply mitigation strategies
• Navigate the ethical dimensions of AI use in development settings, including issues of consent, representation, and accountability
• Make informed decisions about AI adoption (if at all) in their own MERL systems and workflows
• Adopt frameworks and approaches for monitoring and evaluating the use of AI in programs

Course Topics 

• Introduction to key AI terms and concepts 
• Overview of AI uses and AI use in the MERL sector
• Introduction to monitoring and evaluation of AI systems
• GenAI tools and applications for MERL
• Risks, ethics, and best practices
• Hands-on experimentation with AI tools for social impact
• Prompt engineering for MERL practitioners
• Frameworks for MERL of AI Programmes

Space is limited, so if you’re planning to be in Dar Es Salaam, Tanzania then, and you’d like to join, be sure to submit your application for the workshop by June 30, 2026.