Last week I read a piece by a senior developer asking "what's next for AI?", four years on from ChatGPT, watching vibe coding go from meme to method and wondering out loud whether the next wave of AI restrictions means the most powerful models will stay open to everyone or quietly become the preserve of a few companies and countries.
As I sit here in Kisumu, I just can't shake off this thought, it's been on my mind for a while now. But being in a different location has given me a fresh perspective, and I'm seeing things from a new angle. The question that's been bothering me isn't really about whether AI will remain open or not, that's not the issue anymore. What's really on my mind is who in Kenya will actually have access to this technology, and at what cost? How will this impact the rest of us who make a living building software? It's a different story when you're looking at it from here, and it's making me think about the bigger picture.
The token bill, in shillings
In a place like San Francisco, a model priced in dollars per million tokens might seem like no big deal. But when you're in a country where the money isn't as strong, like Kenya where the shilling has been losing value against the dollar, it's a different story. For someone earning a decent entry-level salary of 40,000 shillings a month, these costs can add up quickly. The cost of using a advanced reasoning model can be tens of dollars per million tokens, while a simpler, faster model might only cost a fraction of a cent. This huge difference in cost is not just a theory, for a startup founder or freelancer in Kenya, it's the difference between being able to launch a product that can handle a few thousand queries a day, or watching their entire month's revenue get eaten up by bills before the tenth of the month. It's a harsh reality that can make or break a business.
This is why so much of what's actually working on the ground here quietly routes around the expensive frontier models. Startups building for SMEs, for farmers, for public services, lean on the cheaper small-and-fast tiers, cache aggressively, batch what they can and reserve the pricier models for the one step in the pipeline where quality genuinely matters. Affordable compute and shared infrastructure have become explicit talking points in Kenya's own AI strategy conversations, not side notes; because token cost is, in practice, the real access-control mechanism, more than any export list.
Where the money and the mandate are pointing
Kenya has a clear plan for using artificial intelligence, and it's not just talking about it. The country's National AI Strategy, which runs from 2025 to 2030, is very specific about where it wants to see AI being used. It's looking at areas like farming, healthcare, education, and how the government delivers services to people. It's also looking at security, banking, small and medium-sized businesses, the arts, and how to make things more sustainable. This isn't just a list of goals, you can already see it happening in the projects that are being worked on.
In Kenya, Fintech is the area where technology has become second nature. A great example of this is Safaricom, which is now using a type of artificial intelligence called generative AI to improve M-Pesa. Other companies like Cellulant are also using AI to detect fraud and decide who should get credit, and they're doing it with real transactions happening every day. The idea that Fintech and AI go hand in hand is not new in Kenya, it's been widely accepted for years, ever since mobile money became a thing.
Agriculture in Kenya is getting a boost from partnerships that use geospatial and satellite data. Organizations like the FAO, NASA Harvest, Microsoft, and the Kenya Space Agency are working together to create maps of crops and forecast yields on a national scale. This information is not only helpful to farmers, but also to insurance companies and lenders. It's like having a crystal ball that shows what crops will do well and where. Twiga Foods is using a similar approach to improve supply chain and logistics, making it easier to get food from farms to tables.
- Healthcare and education are crucial areas where technology can make a big difference. We're not just talking about people using new tools, but actually changing the way things work. Think about AI-powered advisors, telemedicine, and diagnostics that can reach people in rural areas. Also, platforms that help students learn in a way that's tailored to their needs. In these areas, it's really important to have voice models that can understand and speak many languages. If a chatbot only works in English, it's not going to be helpful in a country where many people speak different languages. The government is working to improve how it delivers services to the public and how it runs things. One way it's doing this is by using chatbots that can talk to people in many languages. It's also trying to make its tools more efficient. The government has made this a top priority and has even held events like GITEX Kenya and the Nairobi AI Forum to bring people together to talk about it. This year, some big names like the Ministry of ICT, UNDP, and some important international partners got involved.
What's happening in Kenya isn't about the country trying to keep up with the latest artificial intelligence trends. Instead, it's about making deliberate choices on how to use limited and costly computing resources. The focus is on practical areas that can make a real difference, such as ensuring people have enough food, can move money around, and have access to healthcare. It's not about creating chatbots just for the sake of having them.
The developer question underneath all of it
This is where things take a personal turn for me, since coding is my profession.
The original article I read asked whether a 500-person engineering org becomes a 5-person org supervising agents. In Kenya right now, the honest answer is: it depends entirely on which "league" you're in, and that stratification is already visible.
There's a tier of Kenyan engineers being pulled directly into the global market, Microsoft's Africa Development Centre, Google Nairobi, Andela, Turing; earning dollar- or globally-benchmarked salaries, often working on the same agentic tooling everyone else is reading about. For that tier, AI isn't a threat to the job; it's the job. The "cost arbitrage" that used to make Kenyan talent attractive purely because it was cheap is reportedly closing, according to local compensation trackers, which is a good problem to have, because it means people are being paid for skill, not geography.
In Kenya, there's a big group of engineers and founders who are creating products for the local and regional market. They have to deal with clients who pay in shillings, and every feature that uses artificial intelligence has to be worth the cost. This is different from what happens in Silicon Valley. For these engineers and founders, a new way of coding called "vibe coding" is helpful in some ways, but not in others. It makes it cheaper to build the first version of a product, which is great for a market where a lot of software still needs to be created. However, it also means that the things that are really valuable are the things that these coding agents can't do well yet. This includes knowing which industry really needs a particular tool, making it work with payment systems like M-Pesa, dealing with slow internet connections and users who speak many languages, and being responsible when the model makes a mistake, like giving a wrong number in a farmer's loan application.
And there's a foundational tier that barely existed as a "tech job" four years ago: data annotation, AI operations, prompt engineering support, often the entry point into the industry now, useful work, but structurally the most exposed if the tooling above it gets good enough to need less human correction.
The next skill isn't "using AI" : it's using it well
Here's the twist I keep coming back to: if token cost is the real ceiling, then the developers and businesses who win in Kenya won't be the ones who use AI the most. They'll be the ones who use it the most precisely.
Currently, many local AI applications rely on using the same model for every problem, regardless of its complexity. This approach can be costly and inefficient, especially for companies with limited budgets. In many cases, a simpler and more affordable model can handle certain tasks, such as one-line classification, without sacrificing performance. However, for more complex tasks that require multi-step workflows, a more advanced model may be necessary. The most skilled developers are those who can instinctively determine which tasks require a more powerful model and which ones can be handled by a cheaper, faster alternative. They prioritize using the more affordable option unless there's a compelling reason to use a more advanced model, which helps to optimize resources and reduce costs. By taking a more thoughtful and nuanced approach to model selection, developers can create more efficient and effective AI applications that are better suited to their specific needs and budgets.
That's a different skill from "knowing how to prompt." It's closer to systems thinking:
Choosing the right model is a key part of designing a system, not something you do later on. It's like picking the right database for the job - you want to match the complexity of the task to the capabilities of the model. A good engineer wouldn't just use the biggest, most powerful database by default, they'd choose the one that's just right for the workload. Same thing with models, we should be thinking carefully about which one to use, rather than just going with the biggest or most popular one out of habit or because it's trendy.
- Token discipline as an engineering practice : trimming context, caching aggressively, batching where latency allows, routing simple calls to small models and reserving expensive ones for the one step where they actually earn their cost. This is quietly becoming as core to the job as query optimization was for the last generation of backend developers. When working on a project, it's not about sticking to just one tool. Instead, it's about bringing together several models and services to get the job done. You might have one model for finding information, another for making sense of it, and a third for handling different languages. Then there's the cheaper option for routine tasks, and maybe a local or open-source model for sensitive information that needs to stay within the country. The key is knowing how to connect all these tools in a way that makes sense, and also knowing when a task is simple enough that it doesn't need any fancy AI at all. This is becoming a special skill in itself - being able to orchestrate all these different tools and services to achieve your goals.
I believe this is exactly the kind of gap that will lead to a certification market, and I'm willing to bet it will happen here. Kenya already has a strong interest in this area, with bootcamps and upskilling programs growing rapidly alongside the rollout of AI strategies and initiatives that train large numbers of people in practical AI skills. The next phase of this training won't focus on basic skills like writing prompts, as that's becoming outdated. Instead, it will focus on more advanced topics like AI systems literacy, such as how to design a multi-model pipeline, how to manage and monitor token spend like cloud infrastructure, and how to decide when human intervention is still necessary. Whoever develops this curriculum, whether it's a university, a bootcamp, or a company like Safaricom or Microsoft training its own employees, will essentially define what it means to be an "AI-ready developer" in this market for the next few years. This is a crucial step, as it will help establish a standard for AI expertise and provide a clear path for developers to acquire the skills they need to succeed in this field. As the demand for AI professionals continues to grow, having a well-defined certification program will be essential for identifying and developing top talent.
So, same worry, different shape
The Polish piece I read ended without a firm answer, and I don't have one either, but here's my honest projection for Kenya specifically:
I think the real problem isn't that the most advanced AI technology will be kept away from us because of export controls, that's more about who gets to use the latest and greatest models first, and let's be honest, Kenya has never been at the front of that line anyway. The real risk I see is that there's a limit to how much it costs to use these advanced models, and that's going to create a two-tier economy. On one hand, you'll have the big global companies and their developers who can use the best AI models without any issues. On the other hand, you'll have the local companies that are trying to solve problems specific to Kenya, using Kenyan money, and they'll be stuck with cheaper, faster, but not-as-good models because that's all they can afford. This means the local sector will never be able to catch up, and that's a big problem.
It's not all bad news. Having cheap and fast models, along with a developer who really gets the problem and knows exactly when to use a model and when not to, is a pretty powerful combination. This could be a better way to build a strong AI sector in Kenya, rather than just relying on the most expensive and advanced tools for every task. But it does change what makes a Kenyan developer valuable over the next few years. It's no longer just about being able to use an API, anyone can do that. Now it's about having the judgment to know which model to use, how much it will cost, and how to make it work with the local constraints. And that's exactly what the next round of certifications will try to teach. The tools are now available to everyone, but it's the judgment and expertise that will set developers apart. Being able to prove that you have this judgment and expertise will be key. It's not just about being able to call an API, it's about being able to make informed decisions about which model to use and how to use it effectively. This requires a deep understanding of the problem, the models, and the local context. In the end, this shift could be a good thing for Kenyan developers. It will allow them to focus on what really matters, using their judgment and expertise to build effective AI solutions that meet the needs of their community. And it will give them a unique advantage in the market, as they will be able to provide tailored solutions that take into account the local constraints and requirements. So, while the tools may be democratized, the value of a skilled and knowledgeable developer is still very high. And it's this judgment and expertise that will be the key to success in the next round of certifications.
What do you think? Is Kenya's AI story going to be about catching up to the frontier or about getting really good at building brilliantly within a tighter budget?
Top comments (0)