WS #462 Bridging the Compute Divide a Global Alliance for AI

Summary

This panel discussion focused on “Bridging the Compute Divide” and exploring the need for a global alliance to ensure equitable access to AI computational resources. The conversation was moderated by Fabio Steibel from Brazil’s ITS Rio Institute, who proposed creating a “GAVI for AI” – modeled after the successful Global Alliance for Vaccines and Immunization that has facilitated vaccine access worldwide since 2000.

The panelists included Jason Slater from UNIDO, Elena Estavillo Flores from Centro AI para la Sociedad del Futuro, Ivy Lau-Schindewolf from OpenAI, and Alisson O’Beirne from the Canadian government. They identified several key barriers to equitable compute access, including the concentration of computational power in roughly 30 nations (primarily the US and China), infrastructure gaps creating “compute deserts,” skills shortages, and the compounding nature of digital divides that leave regions further behind over time.

The discussion revealed that the challenge extends beyond inequitable distribution to an overall supply-demand gap affecting even developed nations. Panelists emphasized that solutions require multi-stakeholder collaboration involving governments, private sector, academia, and civil society. They highlighted successful examples like UNIDO’s AI for Manufacturing alliance and OpenAI’s Stargate infrastructure project, which demonstrate how technical, financial, and political partners can work together effectively.

Key lessons from GAVI included the importance of inclusive governance models, corrective mechanisms for historical inequalities, and sustainable financing structures. The panelists stressed that addressing compute access must be coupled with investments in local talent, skills development, and ensuring AI tools are designed responsively for diverse communities. The discussion concluded with calls for multiple collaborative alliances tailored to different communities and contexts, emphasizing that collective action requires genuine listening, compromise, and openness to different perspectives and needs.

Keypoints

## Major Discussion Points:

– **Global Compute Divide and Access Inequality**: The panel highlighted the stark disparity in computational power access between the Global North and South, with Brazil having only 1% of global data centers and 0.2% of computational power. This creates barriers for AI development and perpetuates existing digital divides.

– **Need for Multi-stakeholder Collaboration**: Drawing lessons from GAVI (Global Alliance for Vaccines and Immunization), panelists emphasized the necessity of bringing together governments, private sector, academia, and civil society to address compute access challenges through coordinated purchasing power and resource sharing.

– **Infrastructure vs. Benefits Access**: The discussion explored whether countries need local compute infrastructure or if remote access to AI benefits could suffice. Examples included partnerships between countries for supercomputing resources and making AI tools accessible through platforms like WhatsApp in low-connectivity areas.

– **Sustainable Local Capacity Building**: Panelists stressed the importance of investing in people, local talent, and startup ecosystems rather than just hardware, highlighting how Global South innovation often emerges from creative solutions with limited resources.

– **Governance and Equitable Distribution**: The conversation addressed how to ensure fair distribution of compute resources through inclusive governance models that consider not just technical efficiency but social fairness, involving local institutions in decision-making processes.

## Overall Purpose:

The discussion aimed to explore the feasibility and framework for creating global alliances to address the computational power divide in AI development, specifically focusing on how to ensure equitable access for Global South nations. The panel sought to identify barriers, share lessons from successful global initiatives like GAVI, and propose collaborative solutions for bridging the compute gap.

## Overall Tone:

The discussion maintained a constructive and collaborative tone throughout, with participants building on each other’s ideas rather than debating opposing viewpoints. The tone was solution-oriented and pragmatic, moving from problem identification to concrete examples and actionable proposals. There was a sense of urgency about addressing inequities while remaining optimistic about the potential for international cooperation. The conversation became increasingly focused on practical implementation strategies as it progressed, with panelists sharing specific initiatives and calling for concrete collaborative action.

Speakers

– **Fabro Steibel** – Executive Director of ITS Real Institute for Technology and Society (civil society organization), Panel Moderator

– **Jason Slater** – Chief AI Digital Innovation Officer at UNIDO (United Nations Industrial Development Organization)

– **Elena Estavillo Flores** – Founder and leader of Centro AI para la Sociedad del Futuro (think tank), Former telecommunications regulator in Mexico, Economist

– **Ivy Lau-Schindewolf** – International policy and partnerships at OpenAI, Global affairs team member coordinating work in growth markets (Africa, Latin America, APEC) and multilateral engagements

– **Alisson O’Beirne** – Director of International Telecommunications and Internet Policy, Department of Innovation, Science and Economic Development, Government of Canada

Additional speakers:

None identified beyond the provided speakers names list.

# Bridging the Compute Divide: A Comprehensive Panel Discussion Report

## Executive Summary

This panel discussion, moderated by Fabro Steibel from Brazil’s ITS Rio Institute, explored the critical challenge of creating equitable access to artificial intelligence computational resources through global collaboration. The conversation centered on the concept of establishing a “GAVI for AI” – a global alliance modeled after the successful Global Alliance for Vaccines and Immunization.

The distinguished panel brought together perspectives from multilateral organizations, government, private sector, and civil society. Jason Slater, Chief AI Digital Innovation Officer at UNIDO, represented the industrial development perspective; Elena Estavillo Flores provided insights from her experience as both a former telecommunications regulator in Mexico and current think tank leader; Ivy Lau-Schindewolf offered the private sector perspective from OpenAI’s global affairs team; and Alisson O’Beirne contributed the Canadian government’s policy viewpoint on international telecommunications and internet governance.

The discussion revealed that the compute divide represents both a supply shortage affecting all nations and an access inequality that disproportionately impacts the Global South. Panelists identified multiple interconnected barriers while emphasizing that solutions require unprecedented multi-stakeholder collaboration involving governments, private companies, academic institutions, and civil society organizations.

## The Global Compute Divide: Scope and Scale

### Quantifying the Inequality

Fabro Steibel opened the discussion with stark statistics illustrating the magnitude of the global compute divide. According to “EAA numbers released this week,” Brazil possesses only 1% of the world’s data centers (representing half of all Latin America) and a mere 0.2% of worldwide computational power. This disparity demonstrates how even major economies in the Global South face significant barriers to AI development and deployment.

Jason Slater expanded on this theme, noting that nearly 3 billion people remain unconnected globally, with Africa showing only a 25% adoption rate for AI and digital tools. He identified the concentration of computational power in approximately 30 nations, primarily the United States and China, creating what he termed “compute deserts” – regions with minimal connectivity and substantial skills gaps.

### The Compounding Nature of Digital Divides

Elena Estavillo Flores provided crucial insight into how the compute divide creates self-reinforcing cycles of disadvantage. She explained that the barriers are “not only complex, but they’re also compounding, they’re self-perpetuating.” This creates a situation where regions already lacking computational resources fall further behind as global demand increases and investment flows to areas that already possess compute capacity.

Alisson O’Beirne reinforced this analysis, noting that “as folks are left behind and as there’s a lack in compute capacity, those that are already behind the game are going to be left further and further behind.” This temporal dimension suggests that without proactive intervention, market forces alone will continue to exacerbate existing disparities.

### Universal Supply Constraints

Ivy Lau-Schindewolf introduced a crucial reframing by highlighting that “the problem isn’t just inequitable access. The problem is everyone needs more.” She noted that ChatGPT reached 100 million users in one month and now has 500 million weekly active users, illustrating the explosive global demand for AI capabilities that exceeds current supply capacity.

This universal supply constraint means that solutions cannot rely solely on redistributing existing computational resources from developed to developing nations. Instead, addressing the compute divide requires both expanding overall global capacity and ensuring more equitable distribution of resources.

## Barriers to Equitable Access

### Infrastructure and Investment Challenges

Elena Estavillo Flores highlighted particular challenges facing Latin America, where “there’s not enough private investment” and “governments don’t have enough resources for the investments that are needed.” This creates a funding gap where neither private markets nor public resources alone can provide the massive capital investments required for modern AI infrastructure.

Geographic cost differences and market concentration create additional barriers. The concentration of computational resources in specific regions leads to cost advantages that become self-perpetuating, making it increasingly difficult for emerging economies to compete or develop local alternatives.

### Skills and Capacity Gaps

Beyond physical infrastructure, Jason Slater emphasized that compute deserts are characterized not only by lack of connectivity but also by “significant skills gaps.” These capacity constraints mean that even when computational resources become available, many regions lack the technical expertise necessary to utilize them effectively.

Ivy Lau-Schindewolf reinforced this point, noting that “cultivating vibrant startup ecosystems and investing in people through education programmes are essential beyond just hardware infrastructure.”

## Learning from GAVI and Existing Models

### Multi-Stakeholder Collaboration

Jason Slater highlighted how successful global alliances require bringing together diverse stakeholders as trusted conveners. He pointed to UNIDO’s AI for Manufacturing Global Alliance, which includes 140 members from over 40 countries, as an example of effective multi-stakeholder collaboration.

Alisson O’Beirne emphasized how “critical mass through collective action gives countries greater negotiating power than individual efforts.” This insight suggests that coordinated demand from multiple countries can influence market dynamics and pricing in ways that individual national efforts cannot achieve.

### Concrete Examples of Collaboration

Jason Slater presented UNIDO’s Ethiopia coffee project as a practical example of multi-stakeholder collaboration. The project brings together Italy, Ethiopia, Google, NGS, and the International Coffee Organization to address the EU deforestation directive while building local AI capacity.

Ivy Lau-Schindewolf highlighted OpenAI’s Stargate infrastructure project as an example of private sector leadership mobilizing diverse partners for large-scale compute infrastructure development. She also mentioned OpenAI’s Academy training, which has reached 1.4 million people globally, and partnerships with platforms like WhatsApp to provide AI access in low-connectivity environments.

Alisson O’Beirne announced Canada’s collaboration with the UK’s Foreign Commonwealth Development Office through a $10 million commitment from Canada’s IDRC to develop an “equal compute network.”

## Alternative Approaches and Innovation

### Remote Access vs. Local Infrastructure

The discussion explored whether countries require local computational infrastructure or whether remote access could suffice for many applications. Ivy Lau-Schindewolf argued that creative solutions, such as integrating AI capabilities through widely-used platforms, could provide immediate benefits while longer-term infrastructure development proceeds.

However, other panelists emphasized the importance of local infrastructure for enabling indigenous innovation and ensuring technological sovereignty. Elena Estavillo Flores noted that Global South regions have “managed to develop ingenuity and contextual intelligence to find solutions with very limited resources,” suggesting that “if this ingenuity is met with more infrastructure, then there is an opportunity.”

### Building on Local Innovation

Fabro Steibel mentioned that Brazil’s recently released AI national plan includes examples like “Favela GPT and Amazon GPT,” demonstrating local innovation that could be amplified through better infrastructure access.

Alisson O’Beirne expanded the discussion to include “equitability of design and equitability of use,” noting that “if we don’t have AI tools that are designed responsibly and that respond to the needs of local communities, access is not going to be sufficient.”

## Governance and Implementation Challenges

### Balancing Efficiency and Equity

Elena Estavillo Flores emphasized the need for “inclusive governance models with meaningful civil society participation” to ensure that fairness considerations are not subordinated to pure technical optimization. She highlighted that “credible governance models require trust-building mechanisms and fair benefit-sharing to maintain long-term participation and investment.”

### The Importance of Listening and Compromise

Alisson O’Beirne provided perhaps the most crucial insight for implementation, emphasizing that successful collaboration requires “a spirit of listening and openness and a spirit of compromise.” This recognition that technical and financial solutions alone are insufficient without genuine commitment to understanding diverse perspectives provides a foundation for moving from discussion to implementation.

## Concrete Next Steps and Commitments

### Immediate Opportunities

Jason Slater issued a direct call for participants to join UNIDO’s existing AI for Manufacturing Global Alliance, providing an immediate platform for multi-stakeholder collaboration. He also mentioned that the Global Digital Compact provides “a clear framework for action through multi-stakeholder approaches linking digital economy and AI objectives.”

### Committed Resources

Several concrete commitments emerged from the discussion:

– Canada’s IDRC and the UK’s Foreign Commonwealth Development Office have committed $10 million to develop an “equal compute network”

– OpenAI committed to expanding their “OpenAI for countries” programme and Academy training

– UNIDO committed to continuing development of AI lighthouse solutions beyond the Ethiopia coffee project

### Scaling Successful Models

Participants agreed to explore replicating successful consortium models for international compute infrastructure projects, indicating willingness to adapt proven approaches for broader international cooperation.

## Conclusion

The panel discussion revealed both the complexity of addressing global compute access challenges and the potential for meaningful international cooperation. The strong consensus on the need for multi-stakeholder collaboration, combined with concrete examples of successful initiatives and committed resources, suggests viable pathways for implementing global alliance models.

The most significant insight may be the recognition that successful collaboration requires not just technical and financial solutions, but genuine commitment to listening, compromise, and understanding diverse perspectives and needs. The path forward likely requires multiple collaborative approaches tailored to different communities and contexts, coordinated through existing international frameworks, and supported by a combination of public and private resources.

The urgency of action is clear, given the self-perpetuating nature of compute disadvantages and the rapid pace of AI development. However, the discussion suggests that the foundations for effective global cooperation exist, requiring primarily the political will and institutional commitment necessary to translate shared understanding into coordinated action.

Fabro Steibel: So, hello everyone, welcome to the panel. If you are online, welcome. If you are here in front of us, welcome. This is the panel Bridging the Compute Divide, a global alliance for AI. And if you ask if the global alliance or a global alliance, I think it might be a global alliance and even many global alliances as long as we have this idea of global alliance. We’ll introduce shortly the panel and then I’ll pass the words for three rounds of questions and then we open for your comments. So let me explain why we believe we need a global alliance for AI. My name is Fabio Steibel. I’m Executive Director of ITS Real Institute for Technology and Society and it’s a civil society. Last year on the topic of techno diversity, we came with the problem challenge that compute power will be very limited. And this is a different problem for society of information. Brazil has 1% of all data centers of the world. That’s half of Latin America. According to EAA numbers released this week, Brazil has 0.2% of the computational power. And it’s not to say that Brazil or any other country needs to have its own capacity, its own compute power, but the idea is to access. There is a big challenge for access if you look for… the Global North and the Global South bridge or other bridges you can do. This is why we suggest a global alliance for AI, what we call Gavi for AI. So if you’re not familiar with Gavi, it’s the global alliance for vaccination. In 2000, they started to put countries and foundations together to purchase something that was very limited in the market. Same problem we see here with compute power. Might be energy, might be compute parts, but it’s a limited supply of these elements to buy. So together, they could have three groups working. One, that gets the money and make sure everything is accountable. Two, that makes a technical definition of what they should buy. And three, a group that decides on how to share, how to distribute whatever they’re doing. Today, they are responsible for more than a third of the vaccines purchased in the world yearly. They were able to breach the limited supply and were able to increase access to vaccines. Is it the same for compute power? Some yes, some no. What are the lessons to be learned? What are the different approaches we have? So this is the intro. I will pass the word to the speakers. We have very different stakeholders here, which is the best approach to see the problem from different ways. So Jason, I’ll start with you. Jason Slata is the Chief AI Digital Innovation Officer at UNIDO. If you’re not familiar with UNIDO, it’s about industrial development, very important for the topic. Jason.

Jason Slater: Thank you very much for having me here today. Just very briefly, my name is Jason Slater. I’m the Chief AI Innovation and Digital Officer for the United Nations Industrial Development Organization. We’re an organization that’s been around nearly 60 years now. We’re a specialized UN agency with a very specific focus on how can we ensure a sustainable industrial development.

Fabro Steibel: We go for Elena Estavillo Flores. She founded and leads the Centro AI para la Sociedad del Futuro, a think tank that works to build the digital future in an ethical, responsible and inclusive way. Elena, can you hear us?

Elena Estavillo Flores: Yes, yes, I hear you very well. Hello.

Jason Slater: So, please introduce yourself very briefly and then later we go for the two questions.

Elena Estavillo Flores: Introduce myself? Yes, of course. I live in the Centro AI. As you told, this is an independent think tank and we work to foster ethical digital technologies, inclusion, responsibility. And I myself have a long career in regulation, in public policy. I was a regulator for telecommunications in Mexico and I also have taught for many years. I’m an economist.

Fabro Steibel: Thank you very much. I like very much to have economists in the panel. I’ll move now to Ivy. Ivy, Ivy Lau Schindelboff, works in international policy and partnerships at OpenAI. Hi, thank you so much for having me and for moderating and organizing this panel.

Ivy Lau-Schindewolf: My name is Ivy Lau Schindelboff. I am part of the global affairs team at OpenAI, based in our San Francisco headquarter. I wear two hats. One of them is I coordinate our work and growth markets and that means Africa, Latin America and APEC And the other hat I wear is I help lead our multilateral engagements. Pleasure to be here

Fabro Steibel: Thank you very much, Ivy. So now for the fourth participant Alisson O’Beirne is the Director of International Telecommunications and Internet Policy in the Department of Innovation, Science and Economic Development in the Government of Canada Thanks so much. I have a hideously long title. It’s mostly because of our department name. So you did very well there

Alisson O’Beirne: Hi folks, Alisson O’Beirne as as mentioned I’m the Director for International Telecoms and Internet Policy for our little team Within our what is equivalent to the Industry Department in the Government of Canada We have responsibilities for both the ITU and for Internet governance files for Government of Canada And it’s a role that I’ve actually been in for just under a year now So this is my first IGF and I am delighted to be able to experience it Live and in person after hearing about how great this forum is for many years And I had previously spent about five years in our same Industry Department working on AI policy So this is an issue that’s near and dear to my heart for sure

Fabro Steibel: Thank you. And you know Canada was one of the first countries to jump in the AI regulation arena So let’s move to a first question What are the key geopolitical and technical barriers? Preventing equitable access to computational power for AI development and how can International cooperation in particular help to address them. So what the barriers we have to achieve Access Jason, would you like to start?

Jason Slater: Yeah, thank you very much. Yeah, I Guess I a few things from what we see here that In terms of in terms of AI computing power This is really concentrated in in only a few nations right now. I think it’s roughly around 30 Primarily in US and China, etc and the way that I want to tackle this is really by looking at it from UNIDO’s perspective is who are the member states that we are primarily trying to support and where do we see this digital divide or AI computing divide so we can if we look at Africa for example we still see there that you know roughly globally there’s nearly nearly 3 billion still unconnected this is a huge challenge you see in Africa alone that in terms of AI and digital tools the adoptions in the region of 25% so for me one is of course in terms of you thinking about computing power and we have what we call these compute deserts where we have these these zones where there’s just simply no connectivity we also have to see it from how do we ensure adoption how do we also look at it from a skills perspective so there is a huge skills gap that we’re seeing right now and again if I look at Africa in particular but there’s some positive news and I think we’ll come to that later yeah when our fellow panelists have had a chance to speak so there are those are some of the barriers that we are seeing right now now one of the thing this has been framed very clearly by the global digital compact that was that was endorsed last last September in the United Nations General Assembly I was very fortunate to be there at the ceremony that was was celebrating us coming together under the pact for the future and one of my actual roles in the UN is and that’s what I would use as a call for action today is I’m vice chairing the objective number two on an inclusive digital economy which closely links to objective number five on AI so mine is is that we know what those challenges are which challenges to convert themselves into the barriers and how can we then switch that into a much more positive and solution mode and I’ll hold back on that because I think later on I’d like to talk about some of the things not only we’re doing but with our private sector partners and other stakeholders we’re putting in place.

Fabro Steibel: Yeah thank you very much Jason so Finland has a very good experiences on quantum computing and the challenge to share And the topics you bring also bring us very closely to meaningful connectivity Which is a interesting way to see the problem from ten years today ago to today. So I move now to to Ivy

Ivy Lau-Schindewolf: Yeah, I Keep thinking about the way the phrase Equitable access and what the barriers are and I actually want to like take a step back and and think about like just Barriers to access period and And what comes to mind and I think it’s important. It’s important to like Take a moment to think about The gap between supply and demand that everyone faces It’s actually We have reached we have received a lot of outreach from countries with a question of How much demand there really is like do we actually you know, we have this idea that we need GPUs But how much do we really need it? Can you help us quantify? And to be honest, it is not a question. We have thought a lot about until We have seen how fast our models have the demand for our products have grown when we launched at GPT in 22 or 23 now, I can’t remember it seems like a long time ago We thought it was a low-key research preview that nobody would use and pay attention to But then we actually ended up having a hundred million users in one month And now we have five hundred million weekly active users and our CEO Posted on X saying like our GPUs are melting and a lot of people thought oh my gosh, are they actually melting? They know they were not but like we actually just have seen a huge inference demand And that is not we knew training models would involve a lot of GPUs, but even we ourselves have underestimated how much demand there will be on inference. And as we have seen, the cost has come down, and when the cost of serving these models come down, the demand and the use also went up. And when we plotted out in a model, we realized, oh, no, we are not on a trajectory to meet the demands. And so even we ourselves in the U.S. realize we need access to more GPUs. And that’s why we launched Stargate and why we launched OpenAI for countries, and I’ll share more about that later, but I think I want to throw out the framing that, like, yes, we should think about, you know, the divide and whether the access is equitable, but maybe if we can zoom out a little bit, the problem isn’t just inequitable access. The problem is everyone needs more. How do we solve for the gap between supply and demand everywhere?

Fabro Steibel: Thank you. And I like to think like Brazil, we need to prepare for 5G and no connectivity, and you have to take both paths together because we have both audiences. So I will go online now to Elena.

Elena Estavillo Flores: Yes, thank you. I was also reflecting on the question because when we ask if we have enough access, maybe we’re thinking that there is computational power and the problem is how to access it or how to access it equitably. But in this aspect, we have the problem of not enough computational power. No. So it’s a double question about meeting demand and supply and then having the mechanisms. So. that everybody has, well, everybody in the interesting parties has this equitable access. So it’s not just a case of making sure to access to something that exists, but how to produce it. And then we have many barriers that keep just reinforcing themselves. And I’m thinking mostly of Latin America, the case that I know better, because we don’t have enough access to basic infrastructure, to services, to capacity in those services, now for final users. But then this comes over all of this, the companies, the scientists, the academia, the startups that could produce more services, more AI. They have this barrier because there’s not enough compute power so that they can develop the AI that is focused on the region, culture, needs, ways of deciding on how to use AI for which needs, for which problems to solve. So this is just something, it’s like a circle that keeps reinforcing itself. And I see something that it’s something that makes me reflect on whom makes the necessary investments so that there is more compute power in some countries than in others. And it’s very clear that, for example, in the U.S., mainly investment in compute power comes from companies. This is private investment that is oriented to the market. And we don’t see the same dynamics in many of our countries in Latin America. And we expect much of the investment coming from governments. But governments have not enough resources for these huge investments that we would need. And so that’s where this idea of collaborating for producing this infrastructure, this computing infrastructure regionally, well, this is very attractive because individually governments don’t have these resources. But also it comes to questioning us if this will be enough. We need the other side of the investment side from the private sector. So we also have to work on that. And also we can think of collaborative efforts, but bringing together the private sector so that we can gather these sufficient resources coming from only the governments. I don’t think that those will be sufficient.

Fabro Steibel: Elena, thank you very much for your contribution. Alisson, do you want to go next?

Alisson O’Beirne: Yeah, absolutely. I’ll keep my comments relatively brief because I think my colleagues have done a very good job of covering the ground so far. Maybe just to raise a couple of other things. I think we’ve talked a little bit about. And I really want to pick up on what Elena was talking about with regard to the sort of regional and the geographic differences in terms of access, I think is one of the major challenges that we’re facing. Recognizing that particularly for sort of emerging economies for the global south, there are cost issues that can be associated with, you know, developing or establishing compute capacity, that the cost of creating compute capacity is not the same in every region, that there are issues that come to infrastructure, that come to latency that already exists, that have to be addressed in order to be able to establish compute capacity. There’s also, I want to acknowledge, as I think Elena did as well, that there’s a real concentration of the current compute capacity in the hands of a very few providers who have incentives because of the scale of the demand and because of their proximity have incentives to really focus on sort of North American and Western European markets. So with those barriers that I think various folks on the panel have talked about, the one thing that I want to note is that the real challenge that we face is that some of those barriers are not only complex, but they’re also, they’re compounding, they’re self-perpetuating. So as folks are left behind and as there’s a lack in compute capacity, or in some cases lacking infrastructure, or in some cases even lacking demand side, like we talked about, you know, lacking skills or lacking awareness of the necessity of compute power, those that are already behind the game are going to be left further and further behind because as the demand increases in those places that already have compute capacity, we’re going to see just a continuation of response to that instead of a more equitable approach. So I think that’s, as Elena said, a place where governments and particularly where international discussions and international dialogues are critical because you can’t rely sort of on market forces to be able to correct for that need for equity and access.

Fabro Steibel: Thank you Alison and I think you remember that the legacy of the digital divide keeps going and then we have new challenges and the old challenge all together. Exactly. So I’ll go for the second question of the panel. What lessons can be drawn from Global Public Goods Initiatives, such as GAVI that I mentioned before Lessons to design a multilateral framework that ensures fair distribution of compute resources needed for AI development So Jason, we start with you

Jason Slater: Yeah, thank you Hmm, I Think there’s a lot of positives that we can look at that. I mean we when you talk about Multilateral or multi-stakeholder frameworks, it’s we actually launched a similar thing already Near two years ago now. It’s known as AI for manufacturing. It’s a global alliance and What the purpose of this is is that as a you as a as a UN agency that you actually as a trusted advisor if You like to convene and bring together stakeholders. We have around 140 members now from over 40 countries and it’s a complete mix of academia think tanks private sector we’ve got people such as You know Google Huawei imagining those two sitting around the table and what is it that they’re trying to do is to see and identify how we can Leverage AI that can support manufacturing Primarily in those countries that we want to support. So these are you know developing our middle-income countries? So in terms of how Gavi and what it did in bringing together this framework I think there’s a lot a lot of positives that one can see in it I just see from the perspective of what we’re doing around this aim global alliance not to be confused with the title is We want to make it much more solution orientated so when we talk around here now of digital divide of lack of skills of not on Let’s understand where we need to deploy such solutions, etc. So really become much more solution orientated We hear enough about the problems that are going on when the title here. We’re talking about geopolitical issues We’re in the UN and fully understanding that right now, but let’s get on the front foot. Yeah Because we do know that there are gaps. We know for example when we talk about computing power We know where there are challenges right now when it comes to growing a tomato and getting a tomato in Kenya To the market and how 20% of that is lost and here is perfect a use case for a way I can help it So for me bringing all those stakeholders together from the people who need all the way those who provide and Frankly from the UN to have this convening role that actually is a trusted advisor is a phenomenal thing that we can learn Based on the experiences of what happened with Gabi and that scheme before. Thank you Thank you, Jason, and I love that you bring the tomato from Kenya because it moves away from the prompting user Yeah, normal that resembles rural areas Manufacturing intermediaries sustainability and others. I’m going to talk about coffee as well in a few minutes as well.

Fabro Steibel: So Elena, do you want to go next?

Elena Estavillo Flores: Yes, of course about lessons. I think that we have many lessons one of how that an inclusive governance model works very well and in the sense that decisions were not solely shaped by by Some small group of nations of wealthier nations or a small group of corporate interests but that there were many multistakeholders just looking at a meaningful participation for for these different different actors and a civil society also played a very critical role in bringing transparency in monitoring and keeping this equity in Central to this monitoring know and and making sure that decision-making was very aware of of the need to ensure that distribution was fair. So we can learn from that, in that if we get to build this collaboration, that compute distribution will be not only looking at a technical efficiency, but also on socially fairness to distribute and to give access to this computing power. Also of being very aware of the importance of having corrective mechanisms to address historic inequality. So in this mechanisms that have to be designed, that we bring the factors of correcting and to bringing local institutions and research ecosystems to engage continually with these systems. So that it’s not only of bringing a technology or shipping in equipment to build the centers, but to really understand how compute is used for which needs and that local institutions and organizations are deciding on this.

Fabro Steibel: Thank you, Elena. And thank you for bringing the researchers and universities again to the topic. And I like the position you made from a technical solution to a fairness solution. It’s very difficult to define what fairness is, but it’s certainly something we have to pursue and define. So I like it very much. Ivy, do you want to go next?

Ivy Lau-Schindewolf: Sure. Yeah, it’s kind of hard to go after, you know, Elena. And that was a very, very good point and compelling. You know OpenAIR is just one company in this big world, in this ecosystem, but I think we will share what we see from our own vantage point and experience. I think what I couldn’t agree with more from both the Gavi example and what my co-panelists have shared is importance of working across sectors. And what I mean by that is we as a single company, for example, have learned from our Stargate experience that that definitely could not be accomplished by one party alone. And let me just maybe take a moment to explain what Stargate is and who is involved in it and why that convinced me that this has to be a multi-stakeholder solution. So Stargate is a 500 billion dollar infrastructure project, essentially, over the next four years. And we have started building data centers in Abilene, Texas. And this is who and how it all came together. Like Gavi, there is one group of partners that are technology and operations focused. So that includes OpenAI, Microsoft, NVIDIA, Oracle, ARM. And then we need another group of partners that can be finance focused. And for us, in this case, for Stargate, it’s SoftBank. They take up the financing responsibility. And the third group, very much like Gavi, how the responsibilities are distributed and structured, is the political side. This is something we are working with state governments on. And it extends to what we do internationally. And this is what we have talked to the U.S. government about. And I think another possibility I want to posit here is we have heard from a lot of countries that what they really, really want, or they’re not mutually exclusive. I think they want to make sure there is access to the benefits that compute and compute enabled models bring. And so in addition to thinking about infrastructure, we are also therefore very focused on how we can make access to the solutions that compute enable available. And that’s why when we launched OpenAI for countries, it is not copy paste Stargate everywhere. It actually is much more expansive than that. We think about how we work with different education partners, universities, schools, how we increase AI literacy. So we’re not just investing on the stuff, the hardware, right? We’re investing in people as well. We make our products freely available. We have a partnership with WhatsApp. So in low connectivity settings, people can still access the benefits of intelligence that compute enables. And all of that is something that, you know, like as one company, this is what we’re thinking about doing. But we also know that we need funding partners. We need operations partners, technology partners, political partners, very much like Gavi, so that this is something we can coordinate and offer to the world from very different levels of the stack, at the infrastructure level, at the solutions level, and at like even like a people level, so that we are as a society evolving along with the technology.

Fabro Steibel: Thank you. And that reminds me that bridging the compute power might or might not involve localities. So we can process data remotely. we have done. Brazil has a partnership with Spain for supercomputers. Finland has one. Estonia, I reckon, has one as well. So, we can solve problems where the energy is a possibility, where water is a possibility, but also where the supply and demand, it’s another part of the scale and can be achieved. Thank you. So, I move now to the third question of the panel. So, how can Global South nations shape AI governance and infrastructure policies to reduce dependence of foreign compute providers and build sustainable local AI capacity? What lessons can be learned from the Dipsy case or other cases that kind of rephrase it, how we ask questions about computational power? Jason, you want to start?

Jason Slater: No. No, thank you. I’m still wanting to answer the previous question. Tell us about coffee. Tell us about the coffee. Shall I? Because that may be something that can be, I think it’s actually very relevant to this when you’re talking about the Global South. So, I talked about tomatoes before. What we did last year on the sidelines of the General Assembly, we basically built what we call a Lighthouse solution. It links to AI. It was a consortium of two governments. So, Italy, Ethiopia, Ilia Levatsa, Google with one of its implementing partners, NGS and an international coffee organization. What was it we were trying to solve? There’s roughly around 3 billion cups of coffee drank a day. Ethiopia is the fifth largest coffee producers, the number one coffee producer in Africa. And there’s a new EU directive being issued called deforestation. And so, the challenge was, how can you leverage AI and digitalization to support coffee farmers? So, that’s why we built this consortium and brought everybody together. So, again, back to the GAVI example, the AIM Global, this was a very specific one. And there’s many, many other examples. That’s why it links nicely to what you mentioned with Stargate. It’s a huge case where people came together because you could not possibly Solve it as a UN agency or individually. So that was the coffee example. We brought it together. We have a solution We went to Addis. We actually met and Presented this to the coffee farmers association themselves. They didn’t like it. We had McKinsey present didn’t like it Why because they didn’t understand the incentive our land is not subjected to deforestation Only a couple two percent so but when one understands that in order to comply with this directive that it opens up of Opportunities along the supply chain that openly is about, you know, increasing productivity and what-have-you Then we tapped into the real incentive behind it So that’s another thing when we know that there are I when you when you talk about the usages etc And then GPUs and what-have-you I think about there’s so many faces out there where they don’t yet know how AI And digitalization is going to support them. So that was the very that was the coffee example Where are we now 12 months on? We are now looking to see whether we can actually implement some pilot that within Ethiopia in addition to that that consortium are opening up and an AI initiative in Italy Bringing those examples together primarily in the area of coffee. So that was the specific example. I’ll I’ll answer this question later

Fabro Steibel: I Think I Like very much the coffee example because it’s hands-on and in deforestation we bring the AI and climate Close to each other. So you have wonderful uses for mapping deforestation mapping disasters predicting climate and many other things and If you do it together, if you have elements of these that are open that can be shared It certainly is an asset for global global climate change so Could you go next? Yes.

Elena Estavillo Flores: Well, I’ve already said about something that I wanted to focus on in the importance of investing in people, in investing in local talent, and not just thinking of the hard work and and supporting community-driven research, because countries are innovating despite this infrastructure gaps. And this, I find this very interesting, you know, something that we always are, repeat now in Mexico, but that I believe that it is true of many other countries, is that given that we don’t have many resources, so then we have managed to develop ingenuity and contextual intelligence to find solutions with very limited resources. And I see that this is happening also in AI development, that we look at researchers, small developers in civil society that are experimenting and they are using open source and small scales, hybrid models and finding interesting developments. So I think that this is something to learn from and these are opportunities. And if this ingenuity is met with more infrastructure, because we definitely need it, and we can find these collaborative ways to find this infrastructure, then there is an opportunity to meeting. ingenuity with infrastructure. And and also, I find another thing that I find very interesting is that these regions of the world, not like like Latin America, are are pushing for more plural and justice center vision of AI. And also where we emphasize not only into an individual, but collective rights. And this can also help us build government, a strong governance for AI and reshaping this this AI, this governance for AI in a way that that protects these different models of innovation and different models of of protecting individual and collective rights.

Fabro Steibel: Thank you, Elena. And you remind me of how from the global south, we just have to be creative with scarce resources. We just have to hack it. We just have to to make it happen. If we level up the opportunities for certainly for sure, we’re going to have more better results. So, Ivy, do you want to go next? Sure.

Ivy Lau-Schindewolf: All this talk of infrastructure just reminded me that when I first landed in Oslo Airport and then took the train to the city, I’m like, wow, like when trains work, it is an amazing experience and it is such a great utility. And I wish my home city of San Francisco would offer the same thing. And I mentioned that example, because I think when we talk about chips and compute, sometimes it just, you know, we we might think of it as a different category, as other infrastructure that have built cities and nations. And when you ask the question, Fabio, about what to do about that lack of access. I think it’s kind of how we think about what everyone can do about the gap in supply and demand. How can we solve for the problem of access to the infrastructure and the problem of access to the benefits of infrastructure? If we can’t access to the infrastructure to the same degree, is there a creative way to access the benefits? And I want to maybe offer two more things in addition to talking about data centers. One is I think it’s important to incentivize and cultivate a vibrant local startup ecosystem. If you have chips and you have maybe some people who know how to use a tool, but there aren’t entrepreneurs and if you don’t incentivize and really cultivate that growth, then I feel like we’re not really facilitating access. We’re not facilitating innovation. We’re not facilitating access to the benefits of the technology. So I think that is really important. It’s like one of the prongs of what we offer when we say we’re launching OpenAI for countries. And then the other thing is kind of touching on what Elena also mentioned and what I said earlier about like investing in people. We launched OpenAI Academy earlier this year and we are just starting to scale. And if you haven’t seen some of our videos online already, I encourage you to do that. We have been in production like every week and they are for all sectors and they’re all for all skills levels. And sometimes we offer in-person events as well. And the main point here is not like, look, there are more free videos to watch. They are. And like, we’re proud to say that we have trained 1.4 million people already. But I but I think like to think when we think about access, there needs to be like a very concrete way to apply the technology. And we can’t do that if we don’t actually know what it is and what to do that with. I very much I’m excited that when we move from day zero to day one, well, maybe day one to day two, to be fair about all the progress we have seen already, that there will be a much more sophisticated and maybe even demanding like approach to like how we use the compute and the tools that the compute empower and not just talking about like the stuff itself.

Fabro Steibel: Thanks, Ivy. You reminded me that prompting brings humans able to easily talk to computers. So when you have descriptive AI, you usually need a team of technicians to code something. So we use the result. When we have generative AI, we regain this capacity to be a normal user, to make use of technology, which brings the issue of supply and demand to a great distress, because now we have way more need for technologies and uses. So Alison.

Alisson O’Beirne: Yeah, sure. Thanks. I think the question really is about how we include how we include Global South voices in conversations around AI. And I will continue to make it a policy not to tell the Global South how to do policy. But I will say that I think that one of the effective ways that we ensure that we have voices from the Global South in the conversation, and it really reflects back on what Elena was talking about in the Gavi model, is ensuring that we have. a multi-stakeholder approach and one that thinks about including a whole range of different potential partners. So when we’re talking about something like, you know, a global alliance for artificial intelligence, we have to think about ensuring that we’re talking to large and to small providers and users. We have to think about talking about both public and private space. Like it has to be the big AI institutions, the accelerators, the startups, the governments, all coming into the conversation together. That’ll be the most effective way for us to be able to make progress on ensuring equitable access to what’s needed for compute capacity. I’m going to give a little plug to a Canadian initiative in this regard because I can’t be stopped. What we have from our International Development Research Centre in Canada, IDRC, are working with the UK’s Foreign and Commonwealth Development Office that have committed sort of $10 million to the development of an equal compute network that strives to do exactly this. So strives to bring together a number of different partners and a number of different types of partners to think about equitable access to compute capacity. And one of the things that the network really hopes to achieve is by bringing together a number of different sort of partners from the global south together. It allows the possibility to kind of create a critical mass in order to obtain sort of better rates and better processes as they’re looking to establish compute capacity. When you have a number of different countries or a number of different institutions who are speaking together, they have greater power together than they can each individually have on their own. And that’s sort of the value of both the international or the global alliance model and also of the multi-stakeholder model. So it’s a nice little Canadian piece we’ve got.

Fabro Steibel: I like the Canadian piece. We have time for one last question. And one thing I like about Gavi is that they have a political group on how to share the assets. So imagine you buy 100 units of something and they will have to chip in, contribute, find consensus on how to share. So in the case of Gavi, part of it will go for countries that can pay for vaccines. So the model is sustainable. Part of the vaccines will go to countries that cannot pay for a vaccine so there’s access. Part goes to countries that are doing an effort to be more sustainable or to produce vaccines and so on. And all of these are political questions that address how to share. So if you can purchase, you can buy finance. If you know what we are buying, if it is a commodity that we can have it, there’s a political group that will decide on how to share it and mood stakeholder seems like a good approach to do that. So I’ll go for the last question so you can advance in one of the points that you made. Jason, for example, said a call to action in the first question if you want to bring back. If you want to bring from tomatoes to coffee to something else, feel free. And in case you have questions, I have further questions for you. So Jason, what would you like to expand?

Jason Slater: No, thank you. Well, linking it to what you just mentioned there and Gavi, we have, as I see a global digital compact in place, a pack for the future. We have a very clear way forward with those objectives. So that’s my first. That is a multi-stakeholder approach. That was governments, that was tech, that was academia. We all came together. We’re nearly 12 months in and there’s momentum. We under the objective number two regarding digital economy, which links to objective five AI, have a very clear call for solution to call for action in that links to what I mentioned before. We already have an alliance, a global alliance on AI for manufacturing, which has actually three pillars to it, which is one around smart manufacturing, which is not so much about tomatoes and coffee, but it’s when it starts to get a bit harder, you know, on the shop floor and how can you infuse AI in the production process. Then we have our AI lighthouse solutions that I just mentioned. The one that I, the example I gave was the one around coffee. We’re also building up around other products. I’m hoping we collaborate with open AI, by the way. and the third component and that as I mentioned is around digital economy that links it directly back so our mandate in UNIDO is to make sure that we implement what is being committed under the GDC and last but not least I there’s a point that Elena mentioned before that I also didn’t uh that I that I mentioned and you also mentioned in terms of regulation around um 4G 5G etc coming in is this open innovation we also have this program that supports that that again is not it’s not a UNIDO program it’s an open one that we convene we bring together this multi-stakeholder around innovation how can you help those great ideas those startups those innovators and bring corporates together as well and importantly investment there is a clear funding requirement here we come and go projects end but then so what so that’s basically what I just wanted to mention is in terms of this multi-stakeholder around the GDC the aim global and those components I also think as you mentioned there about this model for Gavi we are also trying to make sure there is a sustainable model in place that ensures that these great ideas that’s going to come becomes something that’s investable and sustainable so that it then can be ultimately replicated not just in Ethiopia on coffee but we can take it to Latin America you know tomatoes are grown in most places how can we come to COP and ensure that we bring a solution in place that helps mitigate climate change so that would be my call for action please please join us and we will do our utmost to promote the solutions that people are working on thank you.

Fabro Steibel: Thank you Jason and I was in Bonn last week for the SB 62 and the issue of how to use technology to advance climate agenda and how to make technology greener is top of mind maybe will be a big team for COP 30 this year so Helena moving to you I think we shared the challenges of talking from the civil society point of view And it’s always an interesting point of view, because we’re not government, we’re not companies, we’re kind of bringing issues to the debate. So what topic would you like to advance?

Elena Estavillo Flores: What topic? Well, governance, we have been talking about governance, but we’re supposing that the system already exists and we also need to create and to maintain the necessary incentives to create the model and then to sustain it. And one of it is to produce a model that is credible, that brings certainty and so that people trust it. No, the people trust it, and then we have this necessary trust to keep it together and to keep investment coming. And and the countries that invest in it have also need this, the right combination of of incentives to continue dedicating resources for this. And this this should come from different expectations. One is to to have a fair share of benefits from the model. And also to by the means of trust and credibility to keep believing now in this model that is bringing benefits to to the region. and Benefits for Inclusive Development and Sustainable Development. And that’s why I believe that the role of civil society as a component of this governance that produces trust and credibility is so important.

Fabro Steibel: Thank you. This reminds me that usually governments have an AI national plan for development. If I’m right, we see the Observatory ranked 81 countries that have a plan of action for AI. Brazil just released it last month. And one key thing is what kind of pillars we have in this development. So what Elena brings is that it needs to be multistakeholder. What you bring is that it needs to have open innovation. It has to be shared. And this is really interesting. So, Ivy, I go to you. I think one interesting point of view is that companies are very different from each other. So we think of them like the big techs or something like that, but they’re very different. The Americans are very different from Europeans, very different from Chinese. And if you look at one country, they’re very different. And all of them are trying to bring solutions for bridging the compute power from the perspective. So what point would you like to advance?

Ivy Lau-Schindewolf: Thank you for framing it that way, because I was sitting here thinking, you know, I really am not qualified to tell other people what to do. Right. We’re just one company. I can’t tell what governments want to do, what other companies to do and what all of global South should do. But from the perspective of just one company based in San Francisco, what I think I want to the one call to action, so to speak, is let’s use this technology to build a tool. And I’m here to talk to you about how we can use our tools to solve hard problems. I think, I don’t know if everyone has to worry about getting chips the same way. If all of us can get the benefits to the chips the same way. To accomplish that end is to make our tools accessible. That’s why we have, like I mentioned earlier, an integration with WhatsApp so that it’s workable even in low connectivity settings. We are rolling out open weights models later this summer. You mentioned the Brazil AI plan. And two of the examples that were mentioned in that plan were Favela GPT and Amazon GPT. Favela, you know, we made the tool usable in a lot of favelas. And we also have a partnership with the university and Amazon. So that our tool is enabling conservation and health insights that are already helping residents. And just to help preserve the largest rainforest on the planet. And this is the kind of story that really excites me. That we really, there is so much more to come. And that progress is, you know, it’s happening today, right now. And so let’s think about how we use what exists and what is to come. And to solve hard problems and advance the benefits of LLMs.

Fabro Steibel: Thanks, Ivy. And if you are online and want to post the questions, or if you are in the audience, please make a queue. After Alison’s contributions, we’re going to open the floor for participation.

Alisson O’Beirne: Perfect, thanks. I think, really following on from Ivy’s point, I think that… There is something very important to be said for international action to support equitability of access to compute capacity. I think that there’s a recognition, even among those who already sort of have good access to compute, there’s a recognition of the value of sort of sovereign abilities or sovereign access to compute capacity, even in kind of Western Europe and North America. And so it’s understandable that the Global South also ought to be able to play into that ecosystem, that there also ought to be an equitability of access there. But going beyond equitability of compute access, I think it’s also true that if we don’t have AI tools that are designed responsibly and that respond to the needs of local communities, access is not going to be sufficient. So having access to compute capacity doesn’t mean anything. If we think about equitability of access, we also have to be thinking about equitability of design, designing AI systems and tools that are free from bias, that reflect linguistic diversity, that are climate conscious in the way that they’re created. And then equitability of use as well, designing AI tools that protect the individuals that are really looking to seek the benefits of the use of artificial intelligence, that protect workers, that support development purposes, and that have benefits beyond kind of a small group that would already have access to kind of privileges in that regard. So I think really, as we think about equitability of access, it has to be part of a broader conversation about equitability in the design and the use of AI systems as well.

Fabro Steibel: Thank you, Alison. Let me check if there is anyone online or in the audience that want to make a comment. No? I’ll check here. So we go for a last question. And I think that I like the climate agenda and the relationship with it. And what you bring, Alison, is that compute power is the start of it. And the governments will have a huge role on how to control it or make it safer. make it more secure. And they have to relate. And there’s a huge challenge now for climate, which was a problem already. Now it’s a bigger problem. So as concluding remarks, what would you like to highlight or call to action?

Jason Slater: Again, I just underline what we’ve all in our different aspects been saying about is, you know, collaboration is absolutely critical in this moment now around AI, understanding what the needs of it are, what the computing power of it is, making sure that we don’t leave the global south, that the divide doesn’t get bigger, etc. So my final comment would just be that that, you know, let’s, let’s join this alliance. Because as Stargate mentioned, and by Ivy, as what’s going on now with the AI factory that’s being opened up in South Africa, collaboration between NVIDIA and Cassava, what’s going on in Italy, what’s happening in the country, I just left a few hours ago in Austria with AI Gigafactory. It is a consortium of people that’s coming together. And yes, taking all of those components around the ethics of AI, making sure that this is equitable, that it is transparent, it’s inclusive, it’s collaborative, it’s reliable, it’s safe, and it ensures privacy. Privacy was a big issue on the coffee example. I didn’t mention that before. So that would be my final one is just I could not underline any further than what we’ve all collectively said on all those various levels. All I would offer is enough from perspective of UNIDO, we do have a platform in place, I don’t say we’re happy to help and to support and join forces. Thank you.

Fabro Steibel: Thank you. Elena, do you want to go for

Elena Estavillo Flores: Yes, yes, of course. And I will build upon concluding remarks or highlights? on the same ideas. Because These technologies tend to concentrate and to build a scale, so countries, smaller countries or countries that have smaller access to resources have to collaborate, have to collaborate to attain enough scale to be part of the movement. Otherwise, we will have bigger gaps in our development. And so this collaboration, I believe that is a must right now, just to bridge gaps and to change this mode of development that has produced so many persistent gaps that will get wider. So that’s why collaboration has to be the new mode of development.

Fabro Steibel: Thank you. And I hope collaboration sparks for the global alliance from what is happening, could happen just outside of here. So Ivy, do you want to go for the concluding remarks?

Ivy Lau-Schindewolf: At the risk of just repeating myself and other people, I will actually still say there’s something maybe to the trifecta of political, financial and technical operational partners as the way we think about who should be at the table. So and I think that is necessary in all countries and in all fora. And so that we can truly, truly collaborate and take into account all the equities. Because we’re talking about a really a massive, massive scale of infrastructure and a massive, massive potential of transformation.

Fabro Steibel: Thank you. Alisson.

Alisson O’Beirne: Thank you. Always dangerous to go last for this thing. I think I really, I want to build on the idea of collaboration, which I think a lot of us have talked about today, and I think is super valuable as we’re discussing how to kind of encourage that sort of equitable access to the benefits of artificial intelligence. And one thing that I want to build on that concept of collaboration at the risk of being controversial is the need for collaboration to be in a spirit of openness and in a spirit of listening. And it doesn’t matter what kind of equitability challenge we’re talking about, whether it’s the sort of global south’s compute access, whether we’re talking about linguistic diversity in the internet or in AI, whether we’re talking about in Canada, Indigenous connectivity and Indigenous data sovereignty, when we are in these equitability conversations and when we are thinking about how we collaborate with partners, one of the biggest challenges that we see is that governments are not immune to this. We will often come to the table in a spirit of collaboration, meaning here’s my idea and everyone must agree with it in order for us to collaborate. And I think there really needs to be in a space like AI, where the technology is evolving and where our understanding of the capacities and the understanding of the benefits is evolving. We have to come in a spirit of listening and openness and in a spirit of compromise as well. If we want to have collective action, we are going to have to have compromise and we are going to have to have a recognition of the needs of others and a recognition that we don’t always understand the needs of folks who are outside our own context in our own community. So I think that is one of like, if I have a call to action, it really is to come to collaboration in a spirit of sometimes recognizing that maybe your own positioning is wrong or that you need to adjust your own approach in order to be able to meet the needs of other communities. If we’re not able to do that, then we won’t be able to take collective action on these equitability issues and they can’t be solved without collective action.

Fabro Steibel: Thank you. So, with this, I’ll do some concluding remarks and end the panel. I started the panel wondering if there is a need for a global alliance, and yes, not the global alliance, but many global alliances, a diversity of global alliances, and the global alliances, they’ll have shared problems, but also different communities, so they need to find this collaboration according to their communities. Maybe the community of IGF is different from COP, which is different from G20, and so on, and collaboration, off-collaboration, is really important. I still hope that we can increase the access to compute power in the global South somehow, either installing compute power locally, but or sharing compute power. I still hope that we can share compute powers amongst countries and in a collaboration, so Brazil has a collaboration with Spain, for example, for supercomputers, and I still hope that the issue of information society becomes enhanced by this access to a technology that can be transformative, but can also place risks and challenges. So, thank you very much for the participation, and we conclude. Thank you.

Agreement points

Multi-stakeholder collaboration is essential for addressing compute access challenges

Speakers

– Jason Slater
– Ivy Lau-Schindewolf
– Elena Estavillo Flores
– Alisson O’Beirne

Arguments

Multi-stakeholder frameworks like UNIDO’s AI for manufacturing alliance demonstrate the value of bringing diverse stakeholders together as trusted conveners

Multi-sector collaboration is essential, as demonstrated by Stargate’s structure with technology, finance, and political partners

Inclusive governance models with meaningful civil society participation ensure fairness over pure technical efficiency

Multi-stakeholder approaches including diverse partners from public and private sectors ensure Global South voices in AI governance

Summary

All speakers agreed that addressing compute access requires bringing together diverse stakeholders including governments, private sector, academia, and civil society organizations, with each contributing different capabilities and perspectives

Topics

Development | Legal and regulatory

The compute divide creates self-perpetuating disadvantages that require collective action

Speakers

– Jason Slater
– Elena Estavillo Flores
– Alisson O’Beirne

Arguments

Compute power is concentrated in only a few nations, creating compute deserts with no connectivity and significant skills gaps

Infrastructure barriers are compounded by lack of private investment in Latin America, where governments lack sufficient resources for necessary investments

Geographic cost differences and market concentration create self-perpetuating disadvantages for emerging economies

Summary

Speakers agreed that the concentration of compute power creates reinforcing cycles of disadvantage where those already behind fall further behind, requiring coordinated intervention to break these patterns

Topics

Development | Economic | Infrastructure

Solutions must go beyond hardware infrastructure to include human capacity building

Speakers

– Jason Slater
– Ivy Lau-Schindewolf
– Elena Estavillo Flores

Arguments

Compute power is concentrated in only a few nations, creating compute deserts with no connectivity and significant skills gaps

Cultivating vibrant startup ecosystems and investing in people through education programs are essential beyond just hardware infrastructure

Local ingenuity and contextual intelligence in resource-constrained environments create opportunities when combined with better infrastructure

Summary

All three speakers emphasized that addressing the compute divide requires investing in people, skills development, and local innovation ecosystems, not just physical infrastructure

Topics

Development | Sociocultural

Similar viewpoints

Both speakers emphasized the critical importance of ensuring Global South voices are meaningfully included in AI governance structures, with Elena focusing on civil society’s role in ensuring fairness and Alisson emphasizing comprehensive multi-stakeholder inclusion

Speakers

– Elena Estavillo Flores
– Alisson O’Beirne

Arguments

Inclusive governance models with meaningful civil society participation ensure fairness over pure technical efficiency

Multi-stakeholder approaches including diverse partners from public and private sectors ensure Global South voices in AI governance

Topics

Legal and regulatory | Development

Both speakers provided concrete examples of large-scale collaborative projects that demonstrate how different types of partners (technical, financial, political) must work together to achieve infrastructure goals

Speakers

– Jason Slater
– Ivy Lau-Schindewolf

Arguments

Practical solutions like the Ethiopia coffee project show how consortiums can address specific local challenges while building capacity

Multi-sector collaboration is essential, as demonstrated by Stargate’s structure with technology, finance, and political partners

Topics

Infrastructure | Economic

Both speakers emphasized that sustainable collaborative models require building trust and creating fair mechanisms for participation and benefit-sharing, with collective action providing stronger negotiating positions than individual country efforts

Speakers

– Elena Estavillo Flores
– Alisson O’Beirne

Arguments

Credible governance models require trust-building mechanisms and fair benefit-sharing to maintain long-term participation and investment

Critical mass through collective action gives countries greater negotiating power than individual efforts

Topics

Legal and regulatory | Development | Economic

Unexpected consensus

Universal supply and demand gap affecting all regions

Speakers

– Ivy Lau-Schindewolf
– Elena Estavillo Flores

Arguments

The gap between supply and demand affects everyone, with underestimated inference demand creating global GPU shortages

Infrastructure barriers are compounded by lack of private investment in Latin America, where governments lack sufficient resources for necessary investments

Explanation

Unexpectedly, both a major AI company representative and a Global South policy expert agreed that the compute shortage is a universal problem rather than just an equity issue, with Ivy noting that even OpenAI faces supply constraints and Elena acknowledging the global nature of insufficient compute power

Topics

Infrastructure | Economic

Importance of local innovation and contextual solutions

Speakers

– Elena Estavillo Flores
– Jason Slater

Arguments

Local ingenuity and contextual intelligence in resource-constrained environments create opportunities when combined with better infrastructure

Practical solutions like the Ethiopia coffee project show how consortiums can address specific local challenges while building capacity

Explanation

There was unexpected consensus between a civil society representative and a UN official on the value of locally-driven innovation, with both emphasizing that solutions must be contextually relevant rather than one-size-fits-all approaches

Topics

Development | Sociocultural

Overall assessment

Summary

The speakers demonstrated remarkably high consensus on the need for multi-stakeholder collaboration, the self-perpetuating nature of the compute divide, and the importance of human capacity building alongside infrastructure development. There was also strong agreement on governance principles emphasizing inclusivity and fairness.

Consensus level

High consensus with complementary perspectives rather than conflicting viewpoints. The speakers represented different sectors (UN, government, private sector, civil society) but shared fundamental agreement on problem diagnosis and solution approaches. This strong consensus suggests viable pathways for implementing global alliance models for AI compute access, with each sector bringing necessary but different capabilities to the collaboration.

Different viewpoints

Primary approach to solving compute access – infrastructure vs. access to benefits

Speakers

– Ivy Lau-Schindewolf
– Elena Estavillo Flores
– Jason Slater

Arguments

Making AI tools accessible through various means, including low-connectivity solutions, can provide benefits without requiring local compute infrastructure

Infrastructure barriers are compounded by lack of private investment in Latin America, where governments lack sufficient resources for necessary investments

Practical solutions like the Ethiopia coffee project show how consortiums can address specific local challenges while building capacity

Summary

Ivy emphasizes making AI tools accessible without necessarily requiring local infrastructure, focusing on creative solutions like WhatsApp integration. Elena and Jason emphasize the need for actual infrastructure development and local capacity building, with Elena particularly stressing the investment challenges in Latin America.

Topics

Development | Infrastructure | Economic

Role of private sector vs. government investment in compute infrastructure

Speakers

– Ivy Lau-Schindewolf
– Elena Estavillo Flores

Arguments

Multi-sector collaboration is essential, as demonstrated by Stargate’s structure with technology, finance, and political partners

Infrastructure barriers are compounded by lack of private investment in Latin America, where governments lack sufficient resources for necessary investments

Summary

Ivy presents a model where private sector leads with companies like OpenAI, Microsoft, and SoftBank taking primary roles, while Elena argues that Latin America lacks sufficient private investment and governments don’t have adequate resources, suggesting need for different collaborative models.

Topics

Economic | Development | Infrastructure

Unexpected differences

Framing of the core problem – supply shortage vs. access inequality

Speakers

– Ivy Lau-Schindewolf
– Elena Estavillo Flores
– Fabro Steibel

Arguments

The gap between supply and demand affects everyone, with underestimated inference demand creating global GPU shortages

Infrastructure barriers are compounded by lack of private investment in Latin America, where governments lack sufficient resources for necessary investments

Brazil has only 1% of global data centers and 0.2% of computational power, highlighting access challenges

Explanation

Unexpectedly, there’s disagreement on whether the fundamental issue is global supply shortage (Ivy’s position) versus unequal distribution and access (Elena and Fabro’s position). This is significant because it affects whether solutions should focus on increasing overall supply or redistributing existing capacity.

Topics

Infrastructure | Economic | Development

Overall assessment

Summary

The main areas of disagreement center on: 1) Whether to prioritize infrastructure development vs. tool accessibility, 2) The appropriate balance between private sector leadership vs. government/multilateral coordination, and 3) Whether the core problem is supply shortage vs. access inequality

Disagreement level

Low to moderate disagreement level. While speakers have different emphases and approaches, they share fundamental agreement on the need for collaboration and addressing compute access challenges. The disagreements are more about strategy and implementation rather than fundamental goals, which suggests potential for finding common ground and complementary approaches rather than conflicting solutions.

Partial agreements

Similar viewpoints

Both speakers emphasized the critical importance of ensuring Global South voices are meaningfully included in AI governance structures, with Elena focusing on civil society’s role in ensuring fairness and Alisson emphasizing comprehensive multi-stakeholder inclusion

Speakers

– Elena Estavillo Flores
– Alisson O’Beirne

Arguments

Inclusive governance models with meaningful civil society participation ensure fairness over pure technical efficiency

Multi-stakeholder approaches including diverse partners from public and private sectors ensure Global South voices in AI governance

Topics

Legal and regulatory | Development

Both speakers provided concrete examples of large-scale collaborative projects that demonstrate how different types of partners (technical, financial, political) must work together to achieve infrastructure goals

Speakers

– Jason Slater
– Ivy Lau-Schindewolf

Arguments

Practical solutions like the Ethiopia coffee project show how consortiums can address specific local challenges while building capacity

Multi-sector collaboration is essential, as demonstrated by Stargate’s structure with technology, finance, and political partners

Topics

Infrastructure | Economic

Both speakers emphasized that sustainable collaborative models require building trust and creating fair mechanisms for participation and benefit-sharing, with collective action providing stronger negotiating positions than individual country efforts

Speakers

– Elena Estavillo Flores
– Alisson O’Beirne

Arguments

Credible governance models require trust-building mechanisms and fair benefit-sharing to maintain long-term participation and investment

Critical mass through collective action gives countries greater negotiating power than individual efforts

Topics

Legal and regulatory | Development | Economic

Key takeaways

There is a critical need for multiple global alliances for AI compute access, similar to GAVI’s model for vaccines, but adapted to different communities and contexts

The compute divide is both a supply problem (insufficient global capacity) and an access problem (inequitable distribution), affecting even developed nations

Multi-stakeholder collaboration involving technology, finance, and political partners is essential for addressing compute access challenges

Local capacity building must go beyond infrastructure to include skills development, startup ecosystems, and culturally relevant AI solutions

Equitable access encompasses not just compute power but also equitable design and use of AI systems that reflect linguistic diversity and local needs

Alternative approaches like remote compute access and tool accessibility can provide AI benefits without requiring local infrastructure investment

Civil society plays a crucial role in ensuring governance models remain credible, transparent, and focused on fair benefit distribution

Successful collaboration requires openness, compromise, and recognition of diverse community needs rather than imposing single solutions

Resolutions and action items

Jason Slater called for joining UNIDO’s existing AI for manufacturing global alliance with 140 members from 40+ countries

Participants encouraged to engage with the Global Digital Compact framework for implementing multi-stakeholder AI solutions

OpenAI committed to expanding their ‘OpenAI for countries’ program and Academy training (already reaching 1.4 million people)

Canada’s IDRC and UK’s Foreign Commonwealth Development Office committed $10 million to develop an ‘equal compute network’

UNIDO to continue developing AI lighthouse solutions beyond the Ethiopia coffee project to other regions and use cases

Participants agreed to explore replicating successful consortium models like Stargate for international compute infrastructure projects

Unresolved issues

How to quantify actual compute demand versus perceived need in different regions and sectors

Specific mechanisms for fair distribution of compute resources among participating countries in a global alliance

Sustainable financing models that balance private investment with public sector participation in developing countries

Technical standards and interoperability requirements for shared compute infrastructure across borders

Governance structures that can effectively balance efficiency with equity in resource allocation decisions

How to address the climate impact of increased compute infrastructure while meeting development needs

Specific metrics and accountability mechanisms for measuring equitable access and benefit distribution

Integration challenges between different national AI development plans and international collaboration frameworks

Suggested compromises

Accepting that not every country needs local compute infrastructure if they can access benefits through remote processing and tool accessibility

Balancing market-driven efficiency with equity considerations through hybrid public-private partnership models

Combining infrastructure investment with people-focused programs (education, startups, local innovation) rather than hardware-only approaches

Allowing for diverse governance models across different regional alliances while maintaining interoperability and shared principles

Recognizing that collaboration requires adjusting individual country positions and approaches to meet collective needs

Accepting that multiple global alliances may be needed for different communities rather than seeking one universal solution

Integrating climate considerations with development needs rather than treating them as competing priorities

The problem isn’t just inequitable access. The problem is everyone needs more. How do we solve for the gap between supply and demand everywhere?

Speaker

Ivy Lau-Schindewolf

Reason

This comment fundamentally reframed the entire discussion by challenging the basic assumption that the issue is primarily about distribution of existing resources. Instead, it highlighted that even developed nations face compute scarcity, shifting the focus from a North-South divide to a universal supply-demand crisis.

Impact

This reframing influenced subsequent speakers to acknowledge the dual nature of the problem – both scarcity and inequitable access. It moved the conversation away from a simple redistribution model toward more complex solutions involving capacity building and creative access mechanisms.

We don’t have enough access to basic infrastructure… But then this comes over all of this, the companies, the scientists, the academia, the startups that could produce more services, more AI. They have this barrier because there’s not enough compute power so that they can develop the AI that is focused on the region, culture, needs… So this is just something, it’s like a circle that keeps reinforcing itself.

Speaker

Elena Estavillo Flores

Reason

This insight identified the self-perpetuating nature of the compute divide, showing how lack of access creates a vicious cycle that prevents local innovation and cultural adaptation of AI technologies. It connected infrastructure gaps to broader issues of technological sovereignty and cultural representation.

Impact

This comment deepened the discussion by introducing the concept of compound disadvantages and helped other panelists recognize that the problem extends beyond mere access to include innovation capacity and cultural relevance. It influenced later discussions about the need for local talent development and community-driven research.

Given that we don’t have many resources, so then we have managed to develop ingenuity and contextual intelligence to find solutions with very limited resources… if this ingenuity is met with more infrastructure… then there is an opportunity to meeting ingenuity with infrastructure.

Speaker

Elena Estavillo Flores

Reason

This comment flipped the narrative from deficit-focused to asset-based thinking, highlighting how resource constraints in the Global South have fostered innovation and creativity. It suggested that the solution isn’t just about providing resources but about amplifying existing capabilities.

Impact

This perspective shift influenced the conversation to consider Global South countries not just as recipients of aid but as sources of innovation. It contributed to a more nuanced understanding of collaboration that values different forms of intelligence and problem-solving approaches.

Some of those barriers are not only complex, but they’re also, they’re compounding, they’re self-perpetuating. So as folks are left behind and as there’s a lack in compute capacity… those that are already behind the game are going to be left further and further behind because as the demand increases in those places that already have compute capacity, we’re going to see just a continuation of response to that instead of a more equitable approach.

Speaker

Alisson O’Beirne

Reason

This comment introduced a temporal dimension to the inequality problem, showing how current disparities will exponentially worsen over time without intervention. It highlighted the urgency of action and the inadequacy of market-based solutions alone.

Impact

This insight reinforced the need for proactive international cooperation and helped justify why market forces alone cannot solve the equity problem. It strengthened arguments for coordinated global action and influenced the discussion toward more interventionist approaches.

Going beyond equitability of compute access… if we don’t have AI tools that are designed responsibly and that respond to the needs of local communities, access is not going to be sufficient. So having access to compute capacity doesn’t mean anything. If we think about equitability of access, we also have to be thinking about equitability of design… and equitability of use as well.

Speaker

Alisson O’Beirne

Reason

This comment expanded the scope of the discussion beyond infrastructure to include the entire AI development and deployment pipeline. It introduced the concept that true equity requires consideration of design, cultural relevance, and end-user needs, not just computational resources.

Impact

This broadened the conversation significantly, moving from a narrow focus on compute resources to a holistic view of AI equity. It influenced other speakers to consider the full ecosystem of AI development and helped establish that technical solutions alone are insufficient without addressing social and cultural dimensions.

We have to come in a spirit of listening and openness and in a spirit of compromise as well… We are going to have to have a recognition of the needs of others and a recognition that we don’t always understand the needs of folks who are outside our own context in our own community.

Speaker

Alisson O’Beirne

Reason

This comment addressed the meta-challenge of how to actually achieve meaningful collaboration, acknowledging that good intentions aren’t enough and that successful partnerships require humility and genuine openness to different perspectives and needs.

Impact

This served as a crucial reality check for the entire discussion, grounding the technical and policy conversations in the practical challenges of cross-cultural and cross-sector collaboration. It provided a framework for how the proposed global alliances should actually operate.

Overall assessment

These key comments fundamentally transformed the discussion from a relatively straightforward resource allocation problem into a complex, multi-dimensional challenge requiring systemic thinking. The conversation evolved from initial assumptions about redistribution of existing compute resources to a more sophisticated understanding that encompasses supply creation, cultural adaptation, innovation ecosystems, and collaborative governance. The most impactful insights came from reframing the problem (universal scarcity vs. just inequity), recognizing systemic barriers (self-reinforcing cycles), and identifying assets in unexpected places (Global South ingenuity). These comments collectively elevated the discussion from technical solutions to broader questions of equity, sovereignty, and sustainable development, ultimately making the case for why simple technology transfer is insufficient and why genuine partnership and systemic change are necessary.

How much compute demand do countries really need and can this be quantified?

Speaker

Ivy Lau-Schindewolf

Explanation

OpenAI has received outreach from countries asking for help quantifying their actual compute needs, indicating this is a critical gap in understanding that affects planning and resource allocation

How can we measure and address the gap between supply and demand for compute power globally?

Speaker

Ivy Lau-Schindewolf

Explanation

Even developed countries like the US are struggling with compute shortages, suggesting the problem extends beyond just inequitable distribution to overall supply constraints

What are the most effective mechanisms for ensuring fair distribution of compute resources while maintaining technical efficiency?

Speaker

Elena Estavillo Flores

Explanation

There’s a need to balance technical optimization with social fairness in compute distribution, requiring research into governance models that can achieve both goals

How can we design corrective mechanisms to address historic inequalities in compute access?

Speaker

Elena Estavillo Flores

Explanation

Historical digital divides are compounding with AI compute divides, requiring specific interventions to prevent further marginalization of already disadvantaged regions

What is the optimal balance between local compute infrastructure and remote access to compute resources?

Speaker

Fabro Steibel

Explanation

The discussion raised questions about whether countries need local compute capacity or if remote access through partnerships (like Brazil-Spain collaboration) could be sufficient

How can we better integrate private sector investment with government resources for compute infrastructure in developing countries?

Speaker

Elena Estavillo Flores

Explanation

Government resources alone are insufficient for the massive investments needed, requiring research into public-private partnership models for compute infrastructure

What are the most effective ways to cultivate local startup ecosystems and entrepreneurship around AI in the Global South?

Speaker

Ivy Lau-Schindewolf

Explanation

Access to compute infrastructure alone is insufficient without local innovation ecosystems to utilize it effectively

How can we measure and replicate the ‘ingenuity with limited resources’ innovations happening in the Global South?

Speaker

Elena Estavillo Flores

Explanation

There’s recognition that resource constraints are driving innovation in the Global South, but more research is needed on how to scale and support these approaches

What are the most effective models for multi-stakeholder governance in global compute resource allocation?

Speaker

Alisson O’Beirne

Explanation

The complexity of compute resource allocation requires bringing together diverse stakeholders, but the optimal governance structures need further development

How can we ensure AI tools are designed to reflect linguistic diversity and local community needs?

Speaker

Alisson O’Beirne

Explanation

Equitable access to compute must be coupled with equitable design of AI systems, requiring research into inclusive AI development practices