Why the economics of AI compute break on the public cloud at scale — and why, for African startups, they break sooner and harder.
It is far more economical for AI startups running at scale to move most or all of their workloads from the public cloud to a colocation provider. As a rule of thumb, 50% continuous utilisation is the cut-over point at which owning (or renting) your own iron in a colo beats renting capacity from a hyperscaler.
That is true everywhere. For an African AI startup it is truer still. You earn in Naira, Cedi or Shilling but pay your cloud bill in US dollars; you move large training datasets across oceans and pay punishing egress fees to get them back; you serve local users from a data-centre region thousands of kilometres away; and, increasingly, data-protection law requires that some of your data never leaves the country at all. Every one of those forces pulls the cloud-to-colo switch forward.
Any startup can begin on a public cloud with a credit card, a dream and an algorithm to test. Spin up GPU-accelerated instances, train a model, tear it down. AWS, Azure, Google, Alibaba and the rest have made this genuinely easy — and easier still to grow into what looks, to a young company, like always-available, infinite capacity.
It is an illusion the cloud giants work hard to maintain. There is no such thing as perfectly fungible, infinite capacity — and no AI startup was ever denied funding for running on a major cloud. But once the prototype works and the bills start arriving, two realities sink in fast: data gravity (training sets grow into the terabytes and petabytes) and data egress (moving that data off a cloud is slow and astonishingly expensive). Pick one provider's tooling, compute, storage and networking, and you have the same vendor lock-in that a proprietary minicomputer represented in the 1990s.
The generic cloud-vs-colo argument understates the case for an African startup. Four multipliers stack on top of it:
GPU-accelerated capacity is expensive to rent. Worse, a web-scale cloud is virtual compute, storage and networking scattered across a 100,000-server region — a sub-optimal platform for AI workloads that need deterministic, genuine performance. It is telling that NVIDIA runs its Selene supercomputer in its own data centre, and that Google does not run its core machine-learning workloads on Google Cloud's public Kubernetes layer. The people who built the clouds keep their heaviest AI iron on dedicated infrastructure.
In a multi-tenant cloud, demand often exceeds supply and one tenant's spike becomes your slowdown — "noisy neighbours." The risk of performance degradation grows exactly as a startup climbs the hockey-stick curve it hoped to ride. When an AI workload is in production and driving revenue, predictable performance and better performance both become non-negotiable — and both are hard to get on the cloud, at a reasonable price, all the time.
Before renting a single GPU instance, map how your compute needs evolve from prototype → product → production. Capacity grows over time on any metric you choose — MFLOPs, epochs, or kilowatts. Anytime you are using cloud capacity at scale for more than about 50% of any appreciable long term — a year or two or three, even with reserved instances — it is far more economical to run a base production load on iron you own or lease in a colocation facility, and burst into multiple clouds only for peaks. You buy the time to grow on-premises capacity by bursting to cloud, without tying yourself to any single vendor or plan.
Take an eight-GPU AI server (e.g. an NVIDIA DGX-class node) as the unit of comparison. In a colocation facility, the hardware plus storage and switching — depreciated straight-line over two years — lands near $10,000/month, plus roughly $2,000/month for ~10 kW of power and rack space: about $12,000/month all-in. The equivalent on-demand cloud instance can approach $1,000,000 over two years once storage is included.
| Cloud utilisation | On-demand cloud (2-yr) | 1-yr reserved cloud | Colo / in-house |
|---|---|---|---|
| 25% | $250,000 | $586,200 | $288,000 |
| 50% | $500,000 | $586,200 | $288,000 |
| 80% | $800,000 | $586,200 | $288,000 |
| 100% | $1,000,000 | $586,200 | $288,000 |
Illustrative framework (2-year, ~$12k/month colo baseline), adapted from public cloud vs colocation comparisons. Figures are directional, not a budget — but the conclusion is inescapable: past ~50% utilisation, colocation wins, and it widens from there. For an African startup, add the FX and egress multipliers of Section 02 on top of the cloud columns.
Everything above assumes a colocation facility exists that can actually keep the racks running. In much of Africa, it hasn't — because grid power is unreliable, and running a Tier III AI facility on diesel alone is neither economic nor sustainable. That single constraint has kept the continent's AI workloads stranded on distant clouds.
The SolarVolt data centre is built to remove it at the root — because the park it sits in was designed around power. The data centre's default supply is an on-site solar and battery microgrid under a licensed off-grid PPA, with SolarVolt-owned N+1 generators able to carry the full load. The park's compounds are laid out for up to 100 MW of co-located generation as capacity scales, and IUS is in discussion with Tetracore Energy — whose Atakobo Independent Power System in Ogun State began generating in April 2026 — on gas-fired supply. Tenants of SolarVolt's data centre are insulated from grid instability by design — power is contracted and generated for the park, not drawn from a fragile national grid.
SolarVolt is building and operating a 20 MW Tier III, AI-ready data centre inside Mawuko Technology Park, Odeda LGA, Ogun State — a park developed by site owner Integrated Uwa Services Limited (IUS), which provides the land, regulatory compliance, power and amenities. The data centre is phased 5 MW + 5 MW + 10 MW and engineered from day one for GPU-accelerated AI and high-performance computing.
It is the combination African AI has been missing: hyperscale-grade density and uptime, backed by power the operator owns rather than hopes for.
Whether you need a few GPUs, a rack, or a dedicated AI hall, the maths — and the map — point the same way. Talk to us early; Phase 1 capacity is being allocated to anchor tenants now.
Talk to our team ↓ Download this whitepaper (PDF)