You can rent modern AI compute on this continent. But only one of the three big clouds will sell it to you here — and the African regions are the most expensive places on Earth to buy it. Every number below comes from the providers' own public price lists.
Lesson progress0%
0
The unit is a GPU-hour, and the map matters
Training or serving an AI model comes down to renting a GPU for an hour, then doing it a few thousand times. That is the unit everything is billed in, and it is the unit this page compares.
The comfortable assumption is that cloud is a global commodity: same machine, same price, pick whichever datacentre is nearest. Two of those three are false, and the way they are false is specific enough to change what you should build.
What is being measured. On-demand list prices for Linux instances, read from AWS's public price list, Microsoft's public retail price API, and Google's published GPU-region documentation. Nothing here is estimated. The scope limits matter and are spelled out in Section 7: list prices ignore spot, reserved and committed-use discounts, which serious buyers negotiate down.
1
The default answer is the worst one
Most people reach for AWS first. So start there. How many GPU machine types can you actually rent in each region, and what is the best chip on offer?
Bars are the number of distinct GPU instance types on sale. The badge is the most capable GPU you can rent in that region, with the year that chip launched.
AWS's only African region sells 8 GPU machine types. Virginia sells 85. And the newest GPU in Cape Town is the T4 — a chip launched in 2018. There is no A100, no H100, no L40S. Not expensive: absent. You cannot buy one there at any price.
Every other region measured, São Paulo and Mumbai included, will rent you at least an A100.
2
What a 16 GB ceiling actually costs you
"An older GPU" sounds like it just means slower. It doesn't. A GPU can only run a model that fits in its memory, and the T4 has 16 GB. Below that line, patience doesn't help — the job will not start.
Weights only, and that is the floor. Actually running a model also needs room for the CUDA context, activations and the KV cache, which grows with how much text you feed it. So a card whose memory the weights nearly fill is marked won’t run, not "fits" — the bar has to leave headroom to be usable. Quantising to int8 or int4 shrinks the weights and is the standard workaround, at some cost in quality.
Set it to Llama 3.1 8B at fp16 and look at the T4 row: won’t run. Its weights alone are 16 GB on a 16 GB card, leaving nothing for the KV cache. A mid-sized open model that anyone can download for free will not run on the best GPU AWS will rent you in Africa. That is the practical shape of the gap — not a larger bill, but a class of work that cannot start.
"Active parameters" are not the memory cost — this trips people up constantly. Most of the 2026 open models are mixture-of-experts: Only a small subset fires per token, which makes them fast. But every expert still has to be sitting in GPU memory, so the footprint is driven by the total, not the active count. Select it above and watch: the memory bar uses the big number, always.
That is why the biggest 2026 releases overflow even a full eight-GPU node at full precision.
And memory alone isn't enough — the chip has to speak the format. OpenAI's gpt-oss ships natively in MXFP4, a 4-bit format that needs CUDA compute capability or higher. That is Hopper-generation silicon and newer. The T4 is compute capability , so it cannot use the format: it has to dequantise back to bf16 first, and the footprint springs back from to .
And this is spreading rather than receding: Kimi K3 ships in MXFP4 too. So the model OpenAI describes as running in about 16 GB does not run in 16 GB on the only GPU AWS sells in Africa. This is the compute gap biting a second time: old hardware isn't just slower, it is locked out of the efficiency tricks that make new models cheap to run — and those tricks are exactly what a small team on a budget would want to use.
3
Where I was wrong — and it matters
The obvious conclusion from Section 1 is "you cannot rent a modern AI GPU in Africa." I believed that while building this page. It is false, and the correction is the most useful thing here.
Checking the other two providers:
Google's GPU-region documentation lists GPUs for regions worldwide. Not one of them is in Africa.
Microsoft will rent you an H100 in Johannesburg. Also an H200, and an A100 80 GB. Two African regions, real modern silicon, available today.
So the honest headline is not "Africa has no AI compute." It is that of the big three, exactly one will sell it to you here — and the default, the one most tutorials assume, is the one that won't. Had you taken AWS's answer as the continent's answer, you would have concluded the work was impossible when it merely needed a different vendor.
4
The price of buying it here
So it exists. What does it cost? Here is one identical machine — eight H100 GPUs, the standard unit for serious training — in every Azure region that publishes a price for it.
Same SKU, same eight GPUs, sorted cheapest first. Africa in orange.
The two African regions are the two most expensive on the entire list. Johannesburg is the Virginia price for the identical machine, and the second South African region is dearer still. Out of regions worldwide, Africa holds the top two places — the wrong top two.
5
What that means for a real run
Multipliers are abstract. Put in the length of a job and see the bill, and what staying on the continent adds to it.
—
On-demand list prices. A serious buyer would reserve capacity, use spot, or negotiate — all of which cut the absolute numbers, none of which close the gap between regions.
A sense of scale, not a budget. Ghana's GDP per capita is a year. One hour of that machine in Johannesburg costs roughly of average Ghanaian income. This is not a claim that nobody here can afford compute — plenty of organisations can. It is a point about who gets to experiment, and what a failed run costs the person who paid for it.
6
It isn't poverty, and it isn't distance
Two explanations suggest themselves. The data kills both.
GDP per capita, current US$, World Bank. Best GPU from the AWS price list measured in Section 1.
"Poor countries don't get modern GPUs." India's GDP per capita is and lower still than South Africa's — yet Mumbai will rent you a , the newest silicon in this comparison. Income does not explain it.
"Africa always pays the most." On AWS's T4 — the one GPU Cape Town does sell — São Paulo is more expensive than Cape Town: against the Virginia price. Distance from Virginia is not the rule either.
What both expensive regions share is a thin market: few datacentres, little local competition, high cost to build and power them, and not yet enough demand to justify more. That is a problem that changes when demand and investment arrive — a very different thing from a permanent one.
7
What to do about it, from here
Concrete moves, in the order they usually matter:
1Don't let one vendor's answer stand for the continent
The most expensive mistake available in this whole page would have been concluding from AWS that African GPU compute doesn't exist. Check all three providers for your region before concluding anything — the answers differ enormously, and the default is the worst one.
2Decide deliberately whether the data must stay on the continent
If it must — patient records, government data, anything with a residency rule — Johannesburg is the option, and you should price the premium in from the start rather than discovering it later. If it genuinely needn't, the same machine is far cheaper elsewhere.
3Size the model to the GPU you can actually get
A quantised 8B model that runs on hardware you can rent today beats a 70B model you cannot. Use Section 2 before choosing a model, not after.
4Rent by the hour before you buy by the year
This premium is on list price. Spot instances, reserved capacity and committed-use discounts all move it. Measure your own workload's real cost before signing anything.
5Treat this as a moving number
Prices and availability change. The script that produced every figure here is in our repository and re-runnable — re-measure before citing this page. If it has moved, that is itself the story.
The honest scope. On-demand list prices only, Linux instances, on the date this page was built. Excludes spot, reserved instances, capacity blocks, committed-use discounts and negotiated enterprise rates — all of which reduce real bills, sometimes a great deal. AWS is measured across six regions; Azure across every region publishing the H100 SKU; Google is checked for GPU availability only, not price. Other providers exist and were not measured.
✓
Check your understanding
Five quick ones, or the full set. Every answer explains itself, and your first answer is the one that counts.
Six questions across the key ideas.
The lesson at a glance
✓ Measured, not estimated
Where these numbers come from
Every figure was pulled from the providers' own public, unauthenticated price data by website/scripts/build-gpu-data.py, which is in our repository and re-runnable. Built .
Where this sits: this page is about what it costs to run a model. The token tax explainer measures what it costs to run one in Twi or Ga. The two multiply.
This is one of the interactive explainers we make at AgenticGHX — talks, learning tracks, and research out of Ghana.