Size Doesn't Matter

Hey there, Sheriff here 👋 

AI is here. It’s in our phones, our apps, and it does a chunk of our work already.

And every week, there’s another billion-dollar investment in AI capacity. I’m sure you’ve seen the headlines.

Do you ever see them and think, “Where is Africa?".

I know I do. The good news is, I may have found the best place for us to be.

Last year, AI companies across the world received $108 billion in VC funding, but only 0.1% of it came to Africa.

There’s an ongoing arms race to build the world’s smartest AI, and Africa is barely in that conversation.

• Anthropic spent “tens of millions of dollars” training its latest Claude model.

• The US announced Project Stargate, a $500 billion plan to build the biggest AI supercomputer to reach super-intelligence.

• And back in June, Elon Musk’s xAI was spending $1 billion a month building an AI data centre.

But what if Africa didn’t need to build AI supercomputers at all?

What if it could do more with less? 

Turns out, it’s possible. 

But before we tell you how, we need to unpack…

The gospel of “more”

See, AI is a giant prediction machine. 

It’s typically built by taking lots of data and teaching it to a computer program over a period of time.

At the end of this “training” process, the program develops a “brain” called a model, just like Claude Opus or GPT-4o.

These models have an interesting trait. The more data and processing power you give them, the smarter they get.

Researchers have a nice name for it, called the Scaling Law. We call it “The Gospel of More”.

And over the last four years, it’s fueled massive investment in chips, data centers, and AI talent.

But the problem is, the scaling law costs a lot to actually scale.

Every nine months, the cost of training the world’s most powerful AI models doubles. 

During training, OpenAI’s GPT-4 consumed enough electricity to power a small city. 

And the company spends up to $1 million trying to keep ChatGPT running every day.

But thanks to competition, these models go out of fashion very fast. 

So there’s always another huge bill up ahead.

According to Anthropic’s CEO, models could cost as much as $10 billion to build by 2027.

And these huge infrastructure costs? They’re a small price to pay on the path to building Skynet, especially if you’re in San Francisco or Waterloo.

But down in Lagos, Nairobi, or Kigali? It’s way too costly. And it makes one thing all too clear.

The shoe doesn’t fit

AI could have a huge impact on Africa. 

The promise is breathtaking: classrooms without teachers, doctors on every phone, and farmers armed with insights in their local language. 

But the way the world builds AI today is a far cry from Africa’s reality. 

An AI-enabled world means big data centres, abundant electricity and the best research talent.

But in Africa, we still face rolling blackouts and unreliable connectivity. 

And only 5% of our top research labs have access to the right processors.

There’s not much firepower for us to build the infrastructure to get up to speed, either.

Globally, over $112 billion was invested in AI hardware last year alone. This is from startups and big tech combined. 

But in the last 10 years, African tech has raised about $20 billion.

Simply put, our size isn’t up to par.

So, expecting Africa to compete in the global AI race, especially with the same playbook, is a stretch.

It’s like asking a boda driver to out-race a Formula 1 car; wrong vehicle, wrong terrain.

But what if we don’t need that car at all?

What if size didn’t matter?

Here’s something that’s often unsaid about The Scaling Law: it’s not absolute.

Sometimes, smaller models beat the bigger ones. Like the Phi-2.

It’s a small language model developed by Microsoft, and trained on just 2.7 billion parameters (think of these as neurons in the model’s brain).

That’s 74 times smaller than GPT-4o.

Yet, it’s on par in reasoning and coding tasks. 

And here’s the best part…Phi-2 is small enough to run offline on a modern Android phone.

But it's not the only one.

There’s Gemma from Google, Qwen from Alibaba, and DeepSeek from China.

These are open-source models that anyone can fine-tune, deploy, and even run offline. 

These models are compact, efficient, and increasingly competitive with the giants.

And guess what? With these models, Africa may have finally found its place in the AI race.

• The number of smartphones in Africa will hit 700 million this year. That’s half the continent’s population.

• The specs of the average smartphone in Africa are rising; a 16GB RAM with a fast chipset is not uncommon.

• So, that means we have the means to design and use these small language models.

So what if instead of spending billions to train large models, Africa focused on deploying smaller ones that can actually work in our context?

Models that run offline, in local languages, and are fine-tuned for agriculture, health, education, or trade.

That’s a race we can win.

And we’ve done it before. One look at African tech history, and you see one overarching theme.

Small things can be beautiful

Africa has always been the world’s best laboratory for doing more with less.

We turned SIM cards into banks with mobile money.

We turned airtime into a micropayments system.

We made social media our storefront, and YouTube our TV. 

While we’ve never had Silicon Valley’s deep pockets, we’ve never needed to.

We’ve used what was available to solve our problems every time.

And with these smaller models, AI could follow the same pattern. 

Instead of massive cloud models that cost the world to build and need constant internet access, imagine:

• A teacher’s assistant that runs offline on a school tablet, explaining lessons in Hausa, Kiswahili, or Yoruba.
  

• A health worker’s bot that diagnoses common symptoms and gives treatment guidance in remote villages with no network.
  

• A trader’s voice assistant that helps with pricing, accounting, and credit in Pidgin or Amharic.
  

• An agriculture model that scans cassava leaves or maize and flags diseases, all without connecting to the cloud.
  

These are not far-off fantasies. They’re within reach today.

Projects like Lelapa AI in South Africa are already building language models for African languages.

Masakhane is a community crowdsourcing translation data from across the continent. 

Zindi is helping local data scientists solve problems from crop disease to logistics optimisation.

Each of these is built on small, smart, and locally tuned AI.

And ironically, it puts us at an advantage.

Because while the rest of the world is building skyscrapers of data and computers, Africa could be planting thousands of intelligent seeds.

Tiny, local, resilient models that work anywhere, anytime.

And when this happens, we won’t be behind, we’d be running…

A different race

The global AI race today feels like an arms race: who can build the largest brain? 

But Africa doesn’t need to win that race to matter.

Our advantage lies in fit, not scale; in how well technology adapts to real problems.

Because in the end, the measure of progress isn’t how big our models are. It’s how many people their intelligence can actually reach.

So maybe the future of African AI isn’t about training the next GPT-4o. 

Instead, it’s about giving intelligence a passport, and letting it travel offline, into classrooms, clinics, farms, and marketplaces.

Because in this story, size doesn’t matter. Fit does.

What do you think about the potential of offline models in Africa?

Write back and let us know.

That’s it for this week. See you on Sunday for a breakdown on This Week in African Tech.

Cheers,

The Tech Safari Team

PS. refer five readers and you’ll get access to our private community. 👇🏾

Wow, still here?

You must really like the newsletter. Come hang out. 👇🏾