A Q&A with The Geneva Learning Foundation’s Reda Sadki, based on his remarks on the panel “After World Labs: Is the Standalone AI Model Era Ending?”.
What the panel asked
On 3 October 2026, Reda Sadki joined an online panel convened by the AI for Developing Countries Forum as part of its weekly AI review. The starting point was AMD’s announcement, on 28 September 2026, of an agreement to acquire World Labs, the AI research company led by Fei-Fei Li, in an all-stock transaction valued at about 8.2 billion dollars and expected to close by the end of 2026.
The panel asked whether a company can build a durable business around its models alone, or whether value is moving to infrastructure, distribution and applications. It also asked what this means for developing countries: which capabilities they should build and keep, where partnerships are enough, and where dependence on external platforms would limit their choices.
The other panelists were Professor Bruce Mellado, Institute of High Energy Physics, Chinese Academy of Sciences, South Africa; Dr. Veronika Stoka, Department of Biochemistry and Molecular and Structural Biology, Jožef Stefan Institute, Slovenia; Georg Zangl, Co-Founder and CTO, MudoZangl, Austria; and Shashank Tiwari, CEO, Bytical AI, United States.
Three key messages
- Invest in the layer nobody can sell you: your people, your workforce, and the local knowledge they hold.
- Pay for what you want, which is reliable, completed work, and measure the cost per successful task.
- Sovereignty is three abilities: to choose, to leave, and to judge. Hardware sovereignty is out of reach for most developing countries, but model sovereignty is still within reach. You can download a model. You cannot download a data centre.
Why should a government in a developing country care about one acquisition?
When I first saw this week’s topic, I wondered why we should discuss one AI industry acquisition in the context of AI for developing countries. But if you work for a government in a developing country, it is critical to track these industry deals. There are stakes here that will affect your country.
There are two discussions. The first is about the global stakes: multi-trillion-dollar companies competing with each other, and underneath that, the competition between the United States and China. The second is what all of this means if you are a minister of health or a prime minister in a country in Africa, Asia or Latin America. That second discussion is where I hope to contribute.
At the RAISE Summit in Paris in July 2025, I listened to Eric Schmidt, the former chief executive of Google, describe one of the key geopolitical risks: that either China or the United States jumps so far ahead of the other that the other cannot possibly catch up. In that case, he said, all you have left is a nuclear option. He said this in a public forum.
Whatever you think of that assessment, consider what it means for everyone else. Hardware sovereignty is likely to be out of reach for developing countries. Model sovereignty may still be within reach, but deals like this one could make it harder. If a superpower can fall too far behind to catch up, what does that mean for a country that is already far behind? That is why understanding what is happening in the industry is critical for governments in the Global South.
What does the AMD and World Labs deal signal?
The simplest reading is that model companies are being folded into integrated systems. I think of three things. The first is the competition between the United States and China, for example the export restrictions on chips. The second is Apple’s integration of hardware and software, and the different views on how long the success built on that integration will last. The third is robotics. I cannot imagine a chip company looking at a future of robotics that would not want to acquire models for its own needs, so that it does not have to depend on others.
The part I care about most is what this means for the sovereignty of governments, in particular in the Global South.
Nobody outside the deal knows whether it reflects necessity or opportunity, and it can be both. World Labs speaks of scaling its work in spatial and physical intelligence. AMD speaks of research as a guide to future workloads.
For developing countries, the practical lesson is that the company you buy from today may belong to someone else tomorrow. Choose tools you can leave, and keep the knowledge that makes them useful inside your own institutions.
Do independent model companies face a limit?
I do not see a limit right now, because the work around large language models remains so open-ended. Look at the way Jev came out of stealth on a 40 million dollar investment and became the talk of the town. How much of that was PR and how much was substance is a separate question. But a small company that seemed to come out of nowhere was able to bring something new to the table.
I speak as a user. My organization has transformed itself around an agentic-AI-first workflow. From that perspective, a quick look at what is new on the model repository Hugging Face gives you a sense of the level of activity and the pace of change in models. That work remains within reach of smaller organizations, which can use it and own it. The hardware is a different matter.
Are models becoming commodities, and is the model still the product?
Models are becoming commodities, in large part thanks to the Chinese AI industry. Hardware is much harder to commodify, in the sense of becoming a plentiful resource that everybody can access and afford. That has implications for developing countries.
There is a gap between what you can run locally with open models on relatively affordable hardware and what a frontier lab offers. The Uber story carries some irony. Uber ran out of its annual AI budget in the first four months of the year. At the same time, frontier labs are subsidizing token costs, just as Uber did when rides were cheap because it wanted to get you hooked.
If you are a ministry of health and you trust the wrong answer from a model, people die. You have difficult decisions to make. Hardware sovereignty seems increasingly out of reach. Model sovereignty may still be within reach. You can download a model. You cannot download a data centre.
What customers pay for is completed work they can trust. In July, our AI co-worker refused to release two finished reports because a verification script had found seventeen quotations from health workers that it could not match against the raw data. Three had been altered by the drafting model, which had tidied a stray space and smoothed a phrase. The drafting took minutes. The value was in the check, and in people who knew what the health workers had actually written.
Is the gap between local models and cloud models widening or closing?
I do not have the expertise to answer that question. I can speak from my personal experience. As a user, I am trying very hard to work out how to use local models, and what kind of hardware can give results I can use. What matters is the result you are seeking, how reliable the answer is, and what it costs to verify that answer, which together make the total cost. So far, I have not been able to use a local model, even on very expensive hardware, to produce results I can use in my work.
If value is moving beyond models, where can developing countries capture it?
I work with people at the national level in ministries of health. Most of the people I know still use Gmail or Yahoo email addresses to communicate. I take that as a canary in the coal mine for technology maturity, and as a possible proxy indicator for sovereignty. If that is how far we are from owning our own email, what does it mean for hardware and for models?
I still think model sovereignty is within reach, in part because models are being commoditized against the larger stakes of robotics and the integration of hardware and software. But if you sit in a ministry of health, you probably do not have the resources to pay for your staff to use a frontier model. You do not have the technical expertise to set up your own data centre, run it, or teach people to use local models. What is left is the knowledge.
That knowledge is undervalued, and a model without it is an empty vessel. At the Agentic AI Summit in Berkeley, a Microsoft architect said that models are a commodity, and that the advantage is what an organization knows. In health, that means what frontline workers know about their own districts, in their own languages. In Nigeria, our peer learning programme ran for six weeks with UNICEF and NPHCDA, the national primary health care agency. Knowledge flowed country-wide between 4,300 peers across 308 local government areas. That had a multiplier effect on the rate of progress to find zero-dose children.
I would draw a parallel with pharmaceuticals. Much of the potential pharmacology sits on the African continent. Many medicinal plants could be the source of new drugs, yet very few are studied far enough to become medicines, and when they are, they are usually not owned by Africa. A 2026 review of medicinal plants used by rural communities in Limpopo Province, South Africa, still calls for clinical investigation of the uses those communities describe. In southern Africa, the knowledge of the San people about Hoodia was patented without their consent. Even celebrated benefit-sharing agreements, such as the one for rooibos, have entrenched inequity. So while there is a great deal of discussion about building Africa’s own pharmaceutical manufacturing, the path from those resources to an industry Africa owns remains unclear. Africa has the pharmacy, and others own the lab and the factory. AI carries the same risk, and the same opportunity.
So, speaking from the perspective of a ministry of health: look at the knowledge and the data you have, do not underestimate their value, and do not give them away. Do not let local knowledge become the next raw export. If you invest, invest in ways that ensure you own the knowledge, and that you keep the people who can check the validity of what these systems generate. The rest is not a lost cause.
What does sovereignty mean in practice?
I would offer three specific messages.
- First, invest in the layer nobody can sell you, which is your people, your workforce, and the local knowledge you have.
- Second, pay for what you want, which is reliable, completed work. If you are calculating the cost of any AI transition, work out the cost per successful task, and which combination of models and hardware will give you that.
- Third, sovereignty comes down to three abilities: to choose, to leave or opt out, and to judge.
If you come from a low-resource environment, those are the three tests for every difficult decision: whether you can choose, whether you can leave, and whether you can judge the value of what is being produced.
The risk on the other side has a name in our work. Writing about the 2025 State of AI Report, I described a possible new era of digital colonialism, in which a sector becomes entirely dependent on a few private companies for its most critical technology. Where they can, institutions should build rather than rent the parts that hold their knowledge: their data, their evaluation and their workflows. The institutions able to steward this already exist: ministries of health with in-service training policy, regulatory councils with continuing education requirements, field epidemiology alumni networks. The open question is whether they can do it at a cost per practitioner that a national budget can carry.
The ability to judge depends on people. A country that owns the hardware but has lost the people who can check the output has gained very little control. I call the link between a machine’s output and the human ability to check it the validation tether. Automate away the work through which people learn, and nobody will be left who can check what you bought.
Is the real battle between proprietary frontier models and open models?
Reda Sadki: It is one of the real battles, and it is the one that decides whether model sovereignty is possible. Open weights can be downloaded, run locally, and fine-tuned on your own data, on your own premises. CB Insights reported in December 2025 that small open models are gaining ground in government, finance, healthcare and law, and that companies outside the United States and China account for 40% of the small-model market, against 33% for large models. In Berkeley, Jennifer Chayes argued that growing open weight models is the only way the United States, and much of the world, can keep its lead in AI. Andrew Ng described a security review that his team could finish only with open weight models.
Open weights make model sovereignty possible. They do not protect your knowledge. For a district health worker, the contest that decides outcomes runs between two paths. The first is digital colonialism: geo-locked tools, dependency and reinforced inequality. The second is networked intelligence, where AI is a co-worker and peer learning connects people solving problems in similar conditions. An open model used without local data, local evaluation and people who can judge it still produces dependence. I would back open ecosystems, and judge them by whether the people closest to the community own the evaluation of the tools that serve them.
What if AI infrastructure costs drop suddenly?
Then the price of an answer falls again, and the bottleneck moves even faster to checking it. Every cheap output creates a small obligation for someone to read it, judge it, or act on it. The cheaper the drafting, the more checks we need.
A sudden fall in cost would also weaken the compute squeeze that now concentrates power in a handful of hyperscalers. That is good news for developing countries. Cheaper hardware helps, and it still will not be theirs. The knowledge can be, on one condition: that they still have people able to use these tools well and to see when they are wrong. Cheaper compute does not create that capacity, which humans build through work and learning.
What should governments in developing countries prioritize?
We have been talking about chips and models. Many people have suggested, in various ways, that AI is a new species: the first higher-order, non-human intelligence we have encountered, and one we had a hand in creating. If you take that view, it may help to think in terms of immigration rather than sovereignty.
The first issue is border control, and whether you want open borders that let any AI into your country. The second is residency, and the basis on which an AI should be granted a residence permit, just as a foreign worker in most places must apply and qualify. The third is removal, and the basis on which you would remove an AI you have allowed to reside in your country.
I am being both a bit facetious and provocative. But sometimes moving the lens away from an IT problem, a supply problem or a market problem, and trying a different metaphor, is a useful way to think ahead.