On Africa’s AI Energy Problem
Disclaimer.
I am not an AI or data center expert or energy. I am just an enthusiastic (and optimistic) guy who loves to engage in wholesome conversations about Tech and Africa. I also happen to have a graduate degree in Technology Innovations Management and background in building communities around AI and other emerging technologies
Many thanks to the members of the Africa Deep Tech Community. The wisdom I have gained from you whether through our conversations or just pure osmosis inspired me to attempt putting these thoughts out there.
Things will never be the same again
Even if you have been living under the proverbial rock since November 30, 2022 (the day ChatGPT was unleashed), you probably now know that we live in a Generative AI-first world.
The release of the world’s most popular AI chatbot sparked a resurgence of interest in AI. Despite the usual indications of a hype cycle, the consensus is that AI is here to stay.
Sign of the times: If you are a company, you must not only have an AI strategy, but you must include the letters “AI” in the first two phrases that describe what your company is about. Woe betide you if you happen to be a C-level exec without an AI strategy for your organization or a job seeker without AI somewhere in your resume. 😃
What differentiates this current hype cycle from past tech hype cycles (e.g. the blockchain hype cycle of 2020–2021 or the previous AI hype cycle of 2018–2019) is that this time, the technology in question has been consumerized. It’s not just talk. Everyone with an internet-enabled device has easy access to the raw power of generative AI through various everyday tools provided by the major internet companies ranging from ChatGpt on the web to Meta-AI on Whatsapp. Whatever your platform of choice there is a way to directly leverage AI through an interface that you are already familiar with. The feedback loop created by this phenomenon is one of the reasons many believe that in about 5–10 years the world will be a very different place and that the era of Artificial General Intelligence or AGI (AI that is indistinguishable from human intelligence ) may not be that far off.
It is not all roses, however. The prize of AGI (or whatever the ultimate AI-induced era turns out to be) is not just there for the taking. Several challenges stand in the way of those reaching for the next major AI milestone.
The challenges range from technical and ethical to regulatory and environmental. This article will focus on one challenge that has not been spoken about enough but could determine who (company, country, or region) will win the AI race.
Nope, it’s not the challenge of getting access to Graphical Processing Units or GPUs (a formidable challenge to be sure), neither is it the challenge of access to talent (also formidable)
In my opinion,a question to be addressed by those in the AI race is this:
How will we source enough energy to train and/or run the AI that will fuel all of this coming awesomeness?
Energy, not GPUs or Talent
This first came to my notice when I saw an excerpt from an interview granted earlier in 2024 by none other than Mark Zuckerberg where he said that Energy and not computing capacity will be the number one bottleneck to AI progress. In that snippet, he said that even if everyone had access to all the GPUs that they needed, getting the energy to run them would still be a bottleneck. He estimated that while your typical data center is in the 50MW — 100MW range, to do really serious work on AI models, one would need to be talking about building a 1GW data center.
Since then several things have shown that Mark was not just saying this to go viral. For example in the last year, we saw all the major tech titans make moves in this area. Amazon acquired a nuclear-powered data center for $650M and started investment in nuclear power technology. Microsoft made a deal to restart the 3 Mile Island nuclear power facility. Meta announced its interest in partnering with nuclear developers that can help it meet its AI goals. Google also penned a nuclear energy agreement in 2024.
So clearly, energy is high up (if not number 1) on the list of factors that will make a difference in the AI race and you do not have to look far to understand why. Training and running AI models is an energy-intensive game. This Wall Street Journal Article estimates that Open AI’s latest AI model (Orion) will cost about half a billion US dollars to train. We also know from an Electric Power Research Institute (EPRI) report that a regular Google search uses around 0.3 Wh of electricity per query. Meanwhile, ChatGPT requires about 2.9 Wh of electricity (almost 10x). So even when you have a trained AI model, running it is not cheap
Africa’s Energy Situation
If energy is a problem on a global scale, that problem is multiplied tenfold for Africa.
Why? Because Africa’s current installed electricity generation capacity is nowhere close to meeting the basic needs of its 1B+ population never mind powering the number of new Gen AI-ready HyperScale Data Centers that we need to be competitive in the next decade.
Consider this for the sake of illustration:
In December 2024, Meta announced plans to build a new 2 GW data center in Louisiana, USA. Using 2022 data from The Global Economy website, we can see that 2 Giga Watts represents:
- 0.8% of Africa’s 2022 installed capacity,
- 3.2% of South Africa’s (Africa’s largest electricity producer) 2022 capacity,
- 17% of Nigeria’s (my dear country) 2022 capacity
However, the same amount represents only 0.2% of the US 2022 capacity and 0.4% of India’s 2022 capacity
Why is this such a problem for Africa you ask?
Some may ask:
What is wrong with the status quo?
Can’t all Africans in Africa continue to leverage AI tools and services domiciled in other regions?
After all, AWS, Google, and Meta are just a few browser clicks away.
These are fair questions.
To explain it in simple terms I will borrow a statement by Alex Tsado who once told me that if data is the new oil, computing capacity is the oil refinery.
If your oil (data) is refined outside your region and sold back to you as refined petroleum products (AI-driven tools, applications, and services) then you are likely paying a heavy markup for those products.
There is also the issue of data sovereignty i.e. the idea that a nation or region has the right to govern the data generated within its territory. This is becoming an important issue for some African countries as witnessed by the statements from the president of Senegal during his visit to Silicon Valley in October 2024. It does not require a Ph.D in rocket science to know that it is hard to achieve data sovereignty in the era of AI and big data if you do not have scalable ways to manage and process that data in your country/region. And it is tough to figure out a scalable way to manage and process this data if you do not figure out an energy plan to power the computing infrastructure.
The way forward?
There is no clear-cut answer.
At a high level, there are two main options. Either we find ways to generate more electricity in a super cost-effective (and maybe environmentally friendly) way or figure out how to drastically reduce the energy consumption resulting from the creation and operation of the African AI models and applications of the future thereby enabling us to do more with less.
The rest of this article is dedicated to sharing some ideas for consideration. Some of these ideas are not new. Others have come to my attention either in moments of quiet reflection or in conversations with people with whom I share a common interest in this topic.
I must warn you however, that the further down the list you go, the deeper into the realms of science ( or even business/ political) fiction (or fantasy) you may find yourself. 😃
I will mention these ideas and consider how they could apply to Africa. The goal is not to be exhaustive or go into details. The goal is to spark healthy debate, conversation, and perhaps some inspiration.
Strap in!
Big Tech going Nuclear in Africa?
This is an obvious one (or is it?). From all indications, big tech companies will continue to turn more and more to Nuclear energy to power their data centers to the required scale. I already alluded to this earlier in the article with several examples of what Amazon, Microsoft, Google, and Meta are doing or looking to do.
Is there a chance that a Big Tech company would choose to invest in a nuclear-powered data center in Africa?
Currently, South Africa is the only African country with a commercial nuclear power plant and a proven track record of running one so it may already have a headstart. While it may be a tall order for many African countries to undertake a nuclear program at this time, there could be a future where one or more of the big tech companies could be persuaded to build or invest in a major data center in the region with the accompanying nuclear power supply. This would require the host country to determine the necessary incentives to make this happen. It is not too far-fetched to see Microsoft, Google, Meta, or a consortium of similar companies deciding to invest in a new or existing hyper-scale data center in Africa.
If this consortium-based approach worked for projects like the 2Africa submarine cable where the consortium included the likes of Google, Meta, Vodafone, Orange, and MTN, then it could work for AI.
…. or Goes Renewable?
Google Gemini told me that Africa has the greatest potential for renewable energy out of all the continents in the world due to its clear abundance in the solar, wind hydro, and Geothermal aspects of energy generation, however, lack of infrastructure, and political and technology challenges have prevented the region from making any significant headway in harnessing this energy (many African entrepreneurs have horror stories to back this up!). However, if there was ever a time to figure this out, this is it.
Assuming hyper-scale data centers are the thing to set Africa on the path to data (and AI ) sovereignty then we might as well start figuring this out. It was encouraging to see this mentioned in section 1.4.2.1 of the Draft Nigerian National AI Strategy showing that some governments are already thinking in this direction.
Solar Powered Data Centers in the Sahara anyone? 😃
Smaller and more efficient models
The folks building Gen AI models are not oblivious to the conversation about their energy-guzzling Large Language Models (LLMs). Considerable work has gone into making sure that the more recent LLMs have versions that are more energy-efficient through changes like architectural optimizations and reductions in model size. Examples include GPT-4o which uses less energy than its predecessor GPT-4 and Llama 3.3 a smaller 70 billion parameter model that is expected to perform just about as well as its predecessor Llama 3 (450 billion parameters).
We are also beginning to see conversations around Small Language Models (SLMs) being championed by the likes of ARM. These are models designed to perform natural language processing tasks using fewer parameters than larger models (a few million to a few billion). SLMs can provide higher quality and more accurate results for specific use cases because they are trained on curated data sources.
As the Language Model technology gets more refined it is fair to expect better energy efficiency. We may see models that are small or energy-efficient enough to run economically in the traditional data centers that we have today making them easier to deploy and run in regions like Africa. I would also expect that the increased global demand for AI will probably match or outweigh any gains in model efficiency.
This would be a good place to shout out the work of folks like Lelapa AI and Masakhane Project who are building foundational language-specific models for African languages. My hope here is that the more specific the models are, the smaller they are likely to be thus requiring a smaller energy footprint.
Move the computer away from the data center with edge computing
So far we have discussed solutions that involve finding energy to power larger data centers or models that make it possible to leverage smaller data centers thereby requiring less energy. Thankfully these are not the only options. There is also the school of thought that says that if you move some of the computation away from the big data center, you can reduce the size of the data center and the amount of energy needed to power it.
Let’s talk about Edge computing.
A simple definition of Edge computing is that instead of moving all the data to the centralized computer in the data center for processing, what if you could move the computer (or at least part of it) closer to the source of the data? This would enable you to do some of the processing at the point (or close to the point) of data collection, thereby reducing the load on the central data center or in some rare cases eliminating the need for a large centralized data center.
The main use cases for Edge computing are for applications where:
- Latency cannot be tolerated and you need real-time processing/decision making e.g object detection in a self-driving car
- Network latency/congestion is a problem
- You are building Internet of Things (IoT) devices similar to Google Home, Alexa, M-Kopa Motorbike, or your Anergy Solar Box
What does Edge have to do with Africa, AI, and energy?
Most edge devices are small which means that they consume relatively little energy making them easy to power with a battery, solar energy, or both. Yes, they have less computing power than the server in your data center. However, depending on the use case, they have enough computing power than what is required for the task at hand.
While Edge computing on its own is nothing new, crafting regional or national AI strategies based on Edge computing on the other hand is something that has been ignored or “Slept on” for a long time if I am to quote Tayo Adesanya, who believes so much in the future of edge computing in AI that he bet his startup, Lola Vision Systems on it. He and his company are advocating for a new type of AI device that leverages what is called a Neural processing unit (NPU) which is different and more energy efficient than the now popular Graphics Processing Unit (GPU).
It is also worth pointing out that Edge computing may not be as “slept on” as it once was given the recent action in this area. In December 2024, NVIDIA released what it called the most affordable Generative AI Supercomputer, the Jetson Orion Nano (GPU-based) at a price point of $249. This could be a game changer for countries, regions, and organizations that see edge computing as a way out of the AI-energy dilemma.
Combine edge devices with small language models and you get a low-cost, low-energy specialized AI computer that can change the game in regions like Africa by reducing the immediate need to build 500MW — 2GW data centers in the region.
Something to think about.
But getting electricity to the edge is still a non-trivial problem in Africa despite the massive strides in solar and battery capacity.
This is where analog computing solutions could help!
The Return of Analog Computing
I mentioned earlier in this article that we may get into the realm of fantasy. So consider this the borderline between reality and fantasy. Here we go!
One thing to note about computing on the edge is that while it has very low energy requirements compared to cloud computing, it still has energy requirements. For edge computing, energy efficiency is probably more business-critical than it is in the case of cloud computing. There is no point running the small supercomputer in the middle of the Sahara if you have to figure out a way to change the batteries every week or you need to figure out how to get a big enough solar panel and solar cell to keep it properly powered. So how do you figure out how to keep power consumption as low as possible on the edge? That is where analog computing comes in.
The key thing to note with Analog computers is that they can work on analog signals from the real world without needing to convert them to digital and back again. They also do not need to constantly switch on and off like digital circuits. This drastically reduces their power consumption. For example, mythic.ai claims that its new Edge computing AI Chip uses 3.8x less power than its digital equivalent and is 2.6x faster while being 10x more affordable. Aspinity also claims to have an analog computer that can deliver state-of-the-art acceleration for machine learning in size and power-constrained devices with power consumption in the order of 100 micro Watts (i.e in theory, their analog computer can run for more than 18 months on a single size AAA Duracell battery).
If you have 18 minutes, You should check out this highly recommended YouTube video from Matt Ferrell as it does a great job of explaining what analog computers are and the role they can play in the realm of AI.
At this time, the main use for analog computers is for edge computing in applications like signal detection and processing.
For example, most voice-activated digital devices are constantly on and listening for every sound. For each sound it hears, it needs to convert the sound waves to digital form and then process it, to see if it matches any of the wake words (e.g. “hey Google”, “Alexa”) before it captures the subsequent voice command. This is of course energy intensive. With an analog computer, you can build a device whose digital circuit is always in a sleep state and is only triggered when its low-power analog computer detects a voice saying the wake word thereby saving a lot of energy. Devices like these can be combined with IoT and Edge computing devices as described earlier in this article to deliver AI-powered solutions that reduce the workload on the cloud thereby reducing the energy requirements for these data centers in regions like Africa.
So now we have discussed Edge computing and hopefully, you are beginning to appreciate it more.
But does this mean we will always need to buy a whole new shipment of dedicated edge devices to power Africa’s AI dreams?
Cue the final fantasy!
Leveraging existing Internet-enabled devices to train AI models
Ok, now we are firmly in the realm of fantasy. See if you can spot the flying horse!
Remember that the challenge we are trying to solve is all about accessing affordably powered computing infrastructure for either training AI models or running inference from an existing model.
Now consider this, in our everyday world, we are surrounded by 10s, if not 100s of powerful internet-connected computers. According to Parks Associates, the average U.S. internet household had 17 connected devices in 2023. The number would be smaller in Africa but still substantial.
These devices range from our smartphones and tablets to our laptops, smart fridges, and smart TVs. These devices may not be supercomputers but each of them is 1000x more powerful than the computer that sent man to the moon for the first time in 1969 meaning that they are non-trivial in terms of their computing power. Also, each of these devices has a lot of idle time meaning that most of the time, their CPU utilization is low.
What if we could find a way to combine all that idle computing capacity, we could potentially have one of the largest distributed supercomputers known to man.
What if there was a way to use this distributed supercomputer to train an AI model?
This is an idea I have been thinking about for a long time. I thought I was crazy until it was validated in a conversation with Omoju Miller, founder of Fimio when she graced our podcast. It turned out that she had also been thinking of the same thing.
What if all these devices were nodes on a blockchain that could donate their CPU idle time to complex computations in return for some reward? It would be possible to train AI models for relatively little or no cost. Of course, this is easier said than done as there are multiple considerations like privacy and security but imagine for a moment if these were solved as I believe they can be. This could unlock a whole new era for AI in regions of the world like Africa that are strapped for computing capacity.
If I asked you to donate some of your smartphone or laptop capacity to help me train a model in return for some Ethereum, would you do it?
Africa needs to think outside the box
The AI game is being played on a global scale whether Africa chooses to participate or not.
There is an Igbo proverb that translates to:
“When hunters learn to shoot without aiming, birds must learn to fly without perching”
Africa has to adapt to the times.
This article was written for no other reason but to show that there is nothing in the rulebook that says Africa has to do it the way other regions are doing it.
I hope it inspires the ecosystem (entrepreneurs, VCs, researchers, policymakers, students, etc) to see our AI challenges from a different perspective.
Thank you for making it to this point.