By Ola Adebayo
Every conversation I have about AI in Africa eventually arrives at the same place. Will Nigeria get its own sovereign large language model? Can Lagos afford GPU compute? Why does ChatGPT respond differently to a Yoruba prompt than an English one? These are real questions. Some of them are urgent. But I have come to believe that if we want the next generation of category-defining African technology companies to be built in AI, the model layer is not where most of our attention should sit.
The model layer is capital-intensive, dominated by American and Chinese incumbents, and, more importantly, too far from where the durable opportunity actually lives. The interesting thing happening in AI right now is not what is happening inside the models. It is what is happening around them.
In November 2024, Anthropic released a protocol called the Model Context Protocol, or MCP. It was a quiet release. Most people, including most builders, missed it. The protocol solved a specific and painful problem: every AI model that wanted to use external tools, say, to read a calendar, query a database, or send a Slack message, needed a custom integration with every one of those tools. With ten tools and three AI clients, you were maintaining thirty integrations. MCP collapsed that to a single shared protocol.
Eighteen months later, MCP has won. OpenAI adopted it. Google’s DeepMind built it into Gemini. Microsoft shipped it as part of Windows 11. AWS, Salesforce, and Block are all in. The Linux Foundation set up a neutral governance body for it. By April 2026, more than 20,000 MCP servers are registered across community platforms, and the protocol’s official SDKs are downloaded over 97 million times every month.
This is not a niche protocol war. This is the new plumbing of AI.
And plumbing creates infrastructure demand. Every AI agent in production needs MCP servers to do anything useful. Every MCP server needs somewhere to run. Every server in production needs authentication, governance, observability, and the kind of basic operational tooling that mature software ecosystems take for granted. None of that is solved. Most of it is barely started.
If you have been around long enough to remember the early web, what is happening to MCP infrastructure today should feel familiar. A new category is emerging. The protocol works. Adoption is vertical. And yet running it in production is brutal. The default options are unforgiving. Spinning up an MCP server on AWS means provisioning an EC2 instance, configuring security groups, attaching persistent storage, setting up a load balancer, terminating TLS, and writing your own deployment scripts. The same exercise on Azure Virtual Machines or Google Cloud Compute Engine is no easier. Developers who want something more managed end up stitching together Kubernetes clusters, writing their own OAuth flows, configuring TLS by hand, and wiring up audit logging from scratch. A 2025 industry security review found that 53% of deployed MCP servers rely on static API keys. Only 8.5% implement OAuth. Two significant vulnerabilities, one of which exposed over 437,000 developer environments to potential compromise, made headlines last year.
Translated: the field is open. There is no incumbent platform. The category leaders have not been crowned. The “Vercel for MCP,” the “Datadog for AI agents,” the “Auth0 for agentic identity,” none of these companies exist at scale yet. The seats are unfilled.
Now here is the part of the argument I think is being undervalued. There is a pattern in African technology history that maps onto this moment almost perfectly. Africa does not modernise incrementally. It leapfrogs. Fixed-line telephony was skipped in favour of mobile. Traditional retail banking was skipped in favour of M-Pesa and mobile money. Card payment infrastructure was effectively rebuilt for the continent by Flutterwave, Paystack, and Interswitch in ways the rest of the world now studies.
What those examples share is a particular kind of advantage. African founders had no legacy infrastructure to defend, no incumbent rents to protect, and a customer base that forced products to be lean, mobile-first, and operationally efficient by default. The result was infrastructure that the world ended up importing, not exporting.
There is also a generational shift worth naming. For most of the last decade, the dominant model for connecting African technical talent to the global technology economy was Andela: train brilliant African engineers, then place them as remote contributors to companies built elsewhere. That model worked, and the talent it surfaced is real. But the natural next chapter is not to keep exporting engineers. It is to use that same talent base to build the platforms those engineers were previously placed inside. The shift from being labour for someone else’s infrastructure to building the infrastructure itself is exactly the kind of move that the agentic infrastructure window invites.
The agentic infrastructure layer has the same characteristics that made Flutterwave and M-Pesa possible. There is no legacy to defend. There is no incumbent. The cost of running an MCP-native platform globally is dominated by compute economics, where careful engineering trumps capital flywheels. And the product instincts that African founders developed building for low-bandwidth, low-trust, low-margin markets translate unusually well to the kind of efficient, secure, observable infrastructure that AI agents in production are going to need.
Every advantage that let Flutterwave become payments infrastructure for the world applies to MCP infrastructure today.
So what does the opportunity look like concretely? A short, non-exhaustive list.
A managed deployment platform for MCP servers, where a developer can take a Git repository and have a production-grade endpoint in two minutes rather than two weeks. This is the company I am building, called MCPLambda, and I mention it not to pitch but to make the point that the layer exists and is open enough that a London-based Nigerian founder can credibly enter it.
Vertical MCP server marketplaces for high-value domains: financial analysis, legal document review, devops automation, healthcare workflows. The Salesforce AppExchange or the Heroku Add-ons marketplace, but for agentic tooling.
Agent observability and audit tooling that lets enterprises see what their agents are actually doing in production. Gartner forecasts that more than half of enterprises will engage third-party services for AI agent guardrails by 2026. That budget is real and largely unallocated.
Agentic identity and credential rotation. Most production agents today are authenticated with hand-rolled static keys. That will not survive contact with enterprise procurement.
Category-defining infrastructure companies get founded in surprisingly narrow windows. Heroku was founded in 2007, two years after Amazon Web Services made the cloud real. Vercel was founded in 2015, two years after the Jamstack movement gained traction. Railway, Supabase, and PlanetScale all emerged within an eighteen-month period as developers started rejecting the complexity of raw cloud primitives.
None of these companies were founded by the incumbents. They were founded by builders who recognised that a structural gap had opened and moved fast enough to fill it before the obvious candidates noticed.
The MCP adoption curve started in late 2024. The protocol became cloud-deployable in mid 2025. We are, by my estimate, in roughly the eighteen-month window during which the category leaders of the agentic infrastructure era will be founded. By 2028, the seats will be filled. By 2030, consolidation will be well underway.
I do not want to romanticise the opportunity. Building infrastructure is harder than building applications. The customer is more demanding. The capital intensity is higher in some segments. And there are real reasons why African founders have historically gravitated towards fintech and consumer applications rather than developer infrastructure. Those reasons are not arbitrary.
But the window is genuinely there. If you are an African founder thinking about what to build next in AI, it is worth looking up one layer from where the conversation currently sits. If you are an investor allocating capital, the highest-leverage cheques over the next two years may not be the next fine-tuned LLM but the infrastructure that the agents built on those models will run on. And if you are a developer in Lagos, Nairobi, Cape Town, or Accra, the gap between what you already know and what it takes to build production-grade MCP infrastructure is smaller than it looks. The protocol is open. The SDKs are free. The community is welcoming.
The questions about model access matter. The questions about sovereignty matter. But the questions about who owns the infrastructure layer underneath all of it are the questions whose answers will define the next decade of African technology. We can either be part of writing those answers, or we can read them later in someone else’s case study.
Olamide Adebayo is the founder and CEO of MCPLambda.io, the platform-as-a-service for Model Context Protocol infrastructure.
