Passer au contenu principal

SAFOZi

GPU as a Service in Africa: The Compute Gap Behind the AI Factory Headlines

Summary

Africa’s GPU market is real, early and highly concentrated. Demand exists, but only operators with reliable power, patient capital and trusted local relationships are likely to turn it into a durable business.

Bar chart: Africa holds 18% of world population, 2% of AI training data and 0.6% of data centre capacity.
World Bank Africa Economic Update, October 2026, as summarised by FourWeekMBA
  • Supply: South Africa leads. Most large North African capacity is due from 2027, while Lagos anchors West Africa.
  • Buyers: regulated enterprises and governments can sign contracts. Researchers and startups have the clearest unmet need but the least budget. Near-term demand favours inference and fine-tuning over large-scale training.
  • Investment: at $60,000 a GPU and $1.80 to $2.98 an hour, roughly 46% to 76% utilisation is needed to repay the capital in five years, before running costs.
  • Sovereignty: local hosting delivers residency and some control, but chips and core software remain imported, with NVIDIA dominant.
  • Outlook: most hosters and cloud providers should survive, but in fewer, more specialised forms. Pure GPU resellers carry the highest risk.
  • Opportunity: the market is hard, not speculative. Few players can credibly serve it, which makes trust the scarcest asset.

Africa’s GPU market in late 2026 has real capital behind it, but very little capacity that a developer can actually rent. The continent holds nearly 18% of the world’s population, 0.6% of global data centre capacity, and only about 5% of its data centres are AI-ready, according to the World Bank’s October 2026 Africa Economic Update.

This article maps who supplies GPU as a service (GPUaaS), who buys it and why the business is hard, with a focus on North and West Africa. The core argument is that the bottleneck has moved from buildings to economics. A GPU cloud needs cheap power, cheap capital and steady utilisation at once, and most African markets struggle to offer all three. That makes GPUaaS a hard investment in an immature market, but not a speculative one: the demand is real, and few players are trusted to serve it.

Africa at a glance: compute does not follow GDP growth

East Africa grows fastest, yet deployed AI compute sits in slow-growing Southern Africa. South Africa has more than 60% of the continent’s operational data centre capacity. Compute follows power availability, cost of capital and existing data centre ecosystems.

Region2026 GDP growth forecastCompute position, November 2026
East Africa5.9%Nairobi has AI-ready racks and a GPU deployment; the 1 GW Microsoft and G42 project did not proceed
West Africa4.7%Lagos dominates, with new AI-ready campuses and telco GPU projects
North Africa4.0%Strong sovereign plans, with most large capacity due from 2027
Central Africa3.8%Thin market, led by operators such as ST Digital in Cameroon
Southern Africa2.1%The only region with GPU clusters at meaningful scale

Growth figures are from the African Development Bank’s 2026 African Economic Outlook, which puts the continent at 4.2% in 2026.

The numbers that matter most for a GPU provider:

The last point is easy to overlook. A GPU is bought in dollars and loses value quickly, so a high local cost of capital hits this business harder than most infrastructure.

The provider landscape: many announcements, few price lists

No public, continent-wide GPU count exists, so the provider picture has to be built from individual disclosures. The table separates what is operating from what is planned, because press coverage often blurs the two.

ProviderBaseHardware or scaleStatus
Cassava AIPan-African, NVIDIA Cloud PartnerH200; 3,000-GPU first phase and 12,000-GPU total are company figures; up to $720MCassava reports Cape Town operational; Egypt next; no public price list
AltronJohannesburg, inside TeracoGPU count not disclosedOperating since October 2025
HostAfricaSamrand, South AfricaRTX PRO 6000 instances from $735 a monthLaunched April 2026
Atlancis / ServernahNairobi, inside iXAfricaGPU count and price not disclosedAnnounced November 2025
Kasi CloudLagos4.5 to 7.2 MW IT load reported for LOS1; H100 and H200 cited by a Lagos officialLOS1 commissioned on 19 May 2026
MTNLagosPhase one part of a $235M project; phase two GPU build estimated at $240M to $250M and 9 to 20 MW; one report puts the total above $285MPhase one live; phase two reported for H2 2026
Airtel NxtraLagos38 MW, up to 25 kW per rackSlipped from Q1 2026; full service now targeted for 2028
UdutechLagosMarketplace from under $1 an hourLaunched August 2025
ST DigitalCameroon, Côte d’IvoireNVIDIA GPUs; count not disclosedCommercial launch was planned for Q2 2025
OneCloudMoroccoL40S nodes with up to 48 GPUs, on Oracle infrastructureListed on the provider’s website

Global hyperscalers are a smaller part of this market than their African regions suggest. They offer no current-generation self-service GPUs in Africa, and AWS Cape Town has only older T4 instances. An independent review in September 2026 found no GPU listing for Google’s Johannesburg region and no African provider publishing an hourly rate for training-class hardware.

Published pricing is starting to appear at the smaller end. HostAfrica lists GPU plans from $735 a month, though its own FAQ compares them to premium virtual servers rather than dedicated hardware. Delivery dates are also moving: Airtel now points to a 2028 go-live. Cassava reports its Cape Town AI Factory operational, but the 3,000-GPU first phase and 12,000-GPU total are company figures.

Descriptions of Cassava’s commercial availability differ. One competitor’s review calls its capacity enterprise-committed, while the independent review describes GPUaaS sold through a sales process. Both agree that self-service rental at published prices is not the norm.

North Africa: big sovereign plans, later delivery

North Africa has some of the continent’s largest AI projects, but most of the capacity arrives in 2027 or later. Each country is taking a different route.

CountryLead projectPlanned scaleTimingEnergy angle
EgyptAfrica Data Centers Egypt with Vodafone, Elsewedy and Cassava200 MW and $1B target; first phase 20 MW and $200MFirst phase over three yearsMust work around a long-running energy deficit
MoroccoNexus AI Factory16 MW at Nouaceur, then an additional 20 MW in the north, for 36 MW by end-2027; MAD 12B; up to 500 MW long termTwo phases, by end-2027Renewable supply from TAQA Morocco
TunisiaSoleCrypt AI Campus, Bizerte40 MW; first 7 MW with 1,728 Rubin GPUsEarly 202860 MW solar and storage project in Tozeur
AlgeriaAkid Lotfi centre, OranCapacity not disclosedFoundation stone March 2025State-led facility

Egypt shows that AI compute is now geopolitical. Huawei reportedly bid with 2,008 Ascend chips, while US officials were preparing a counteroffer. No submitted US offer is confirmed, and no award has been announced. Vodafone and Cassava plan local GPUaaS for government and business, with data processed in Egypt.

Morocco has the broadest mix of options. The Oracle region in Casablanca opened in April 2026, hosted by N+ONE, and UM6P’s Azul Cloud entered beta in June. Timelines still slipped: the Nexus plan first promised more than 40 MW in Q1 2026, and its April MoU now describes 36 MW by end-2027.

Tunisia has talent and connectivity, but limited AI-grade infrastructure. Its existing facilities remain mainly general cloud and hosting sites. The Bizerte site sits about 3.6 km from the Medusa cable landing, which supports a regional and European-facing role. InstaDeep remains the region’s best proof of what early GPU access can do.

Algeria is taking a public research route. Alongside Oran, ENSIA’s academic GPU system is reported to serve 52 universities and AI startups.

West Africa: Lagos leads, others build sovereign hosting first

West Africa’s GPU market is concentrated in Lagos, where demand is strongest and power is hardest. Elsewhere in the region, governments are mostly securing sovereign hosting before GPU capacity.

Nigeria is the regional centre of gravity. Lagos has 78.6 MW of capacity, with Estate Intel projecting more than 218 MW by 2030, and the state cites 146 MW in its pipeline. A Luminate-commissioned study by the Global Centre on AI Governance puts Nigerian spending on foreign cloud at $850M a year, which is the prize local providers are chasing.

The public sector is moving too. The Nigeria AI Scaling Hub launched shared compute in June 2026, hosted by Galaxy Backbone with Gates Foundation support. In October, the minister proposed pooling GPU capacity with Rwanda, Senegal and Kenya, which remains an ambition for now.

CountryMain developmentsStatus
Ghana$250M national AI compute centre; 30 MW committed at Agility campus, expandable to 150 MWCompute centre approved by Cabinet in March 2026
SenegalState data centre at Diamniadio; Stellarix expansion; reported Alibaba Cloud sovereign cloud workHosting base in place; dense GPU capacity not evident
Côte d’IvoireState data centre of 1.73 MW IT and 228 racks; Raxio CIV1, 3 MW at full capacity; ST DigitalState facility reported on track for 2027

The scale gap inside the region is striking. Côte d’Ivoire’s national facility is smaller than Kasi’s first phase in Lagos, which is itself small by global AI standards. Francophone West Africa may therefore depend on shared regional capacity rather than national clusters.

Timeline grid for North and West Africa showing most announced capacity arriving in 2027 or later.
Stated or reported dates, November 2026 · not to scale

The timeline shows why the next two years matter. Stated dates have already slipped for Nexus and Airtel, so delivery itself is the signal to watch.

Training, fine-tuning and inference: three different GPU businesses

GPUaaS is really three products with different hardware, customers and economics. Africa’s best near-term fit is inference and fine-tuning, not frontier training.

The global market is moving the same way. Gartner forecasts AI-optimised IaaS spending on inference at $23.3B in 2026, against $19B for training. Deloitte forecasts that inference will take about two-thirds of AI compute in 2026, up from about half in 2025.

WorkloadWhat it doesTypical hardwareRevenue patternFit in Africa
TrainingBuilds a model from scratchMany H100, H200 or B200 GPUs linked by NVLinkLarge, bursty projectsLimited: few buyers, and global clouds set the price
Fine-tuningAdapts an existing model to local data or languagesOne or a few L40S, RTX PRO 6000 or H100 GPUsShort, repeated jobsStrong where data is sensitive or language-specific
InferenceRuns a trained model for usersMemory-rich GPUs, often shared across workloadsSteady, recurring usageStrongest: helps fill capacity and benefits from local hosting

Hardware needs differ sharply. QLoRA fine-tuning of a 70B model needs about 48 GB to 65 GB, so a single 96 GB RTX PRO 6000 can handle it, while full fine-tuning and training generally call for linked data centre GPUs. This is vendor guidance and its figures vary, so real sizing depends on the model and batch size.

The business difference matters more than the technical one. Training can leave a cluster idle between projects, whereas production inference can run continuously and improve the utilisation that GPU economics depend on. MTN is pursuing this logic with a plan to place GPUs at tower sites as an inference grid.

GPU models running in Africa

Most confirmed African deployments use one- to three-generation-old hardware, while the newest chips appear mainly in plans. That is workable for inference and fine-tuning, but it limits large-scale training.

Bar chart of memory per accelerator, from 16 GB for the T4 to 288 GB for Rubin and Instinct MI350X.
Vendor specifications as reported by Spheron, VRLA Tech, Vultr and Clarigital · single-accelerator memory; B200 shown per chip (DGX B200 lists 180 GB per GPU); RTX PRO 6000 instances may be partitioned
GPUStatus in AfricaWhereBest workload
T4AvailableAWS Cape TownLight inference
A40, L40SIn useENSIA, Algeria; L40S also listed by OneCloud, MoroccoInference and small fine-tuning jobs
RTX PRO 6000On saleHostAfrica, JohannesburgInference and fine-tuning
A100, H100In useUM6P Toubkal, Morocco (20 A100s from 2022, 12 H100s added March 2025); H100 also at ENSIAResearch training and fine-tuning
H200Deployed or citedCassava; Kasi Cloud, per a Lagos officialTraining and large-model inference
GB200 BlackwellPlannedNexus Morocco, named in the NAVER planLarge-scale training and inference
RubinPlannedBizerte, Tunisia, early 2028Next-generation training and inference
Ascend 950, 910BProposedHuawei’s reported Egypt bid: 1,408 Ascend 950 for training plus 600 Ascend 950 or 910B for inference; no award announcedTraining and inference clusters

Who makes the chips

NVIDIA dominates, and African buyers depend on the same few producers as everyone else. No GPU in this article is designed or manufactured in Africa.

ProducerGlobal positionMain productsAfrica relevance
NVIDIAAbout 81% of AI data centre chip revenue in a widely cited estimate attributed to IDCH100, H200, Blackwell, Rubin; CUDA softwareBehind nearly every project named here
AMDEstimates range from 5% to 13%Instinct MI300 and MI350; ROCm softwareNo public African GPU cloud deployment found in this review
Huawei572,000 accelerators shipped in China in H1 2026Ascend 910 and 950; CANN softwareReported Egypt bid and GPU talks in Algeria
Cloud giantsCustom chips at about 5% in one estimateGoogle TPU, AWS TrainiumMainly available through their own clouds
IntelAbout 1%Discrete AI GPUsMarginal for African GPUaaS

For a provider, the choice of producer is also a software and geopolitical decision. NVIDIA offers the deepest ecosystem but exposes buyers to allocation and US rules. Huawei offers a second supply route, but its CANN tools remain less mature than CUDA. The safest plan is to buy for the workload you can contract today and keep software portable.

The AI users: four buyers with very different needs

Local GPU demand is real, but it is split between buyers who need sovereignty and buyers who can barely pay. No provider can serve all of them with one product.

BuyerTypical workloadWhat they valueNamed examples
Researchers and universitiesModel training, local language workFree or subsidised accessCSIR, Zindi community, ENSIA
StartupsFine-tuning and inferenceLow price, instant access, easy paymentLelapa AI, Dataviue, MathU
Regulated enterprisesFraud detection, customer service, document searchData residency, contracts, supportBanks, telcos and health groups
GovernmentsPublic services, sovereign AIControl, local hosting, anchor capacityEgypt, Nigeria’s NAISH

The largest latent group is the least able to pay. A UNDP analysis of Zindi users found only 5% had the compute they needed, with the rest relying on laptops or Google Colab. Payment rails make this worse: in Nigeria, GTBank raised its naira card limit to $40,000 a quarter in August 2026, while Stanbic IBTC’s stands at $100 a month.

Enterprises and governments are the likeliest source of bankable contracts, but their workloads may be modest. The World Bank sees the main opportunity in small, low-bandwidth, locally adapted applications, not frontier models. Its review of government procurement also found that only 1.4% of identified AI systems used generative AI.

That points to a clear market insight: near-term African demand looks more like steady inference and fine-tuning than giant training runs. Providers sizing clusters for frontier training risk building the wrong product. Some are also looking abroad, with Cassava planning to sell excess capacity to global Nvidia clients and Nexus targeting demand across EMEA.

Why GPUaaS is not an easy investment

GPUaaS is hard to underwrite because a fast-ageing dollar asset must earn its return in an immature market. Prices are set globally, costs are set locally, and the data investors need barely exists.

Bar chart: a $60,000 GPU breaks even at 76% utilisation at $1.80 an hour, 52% at $2.64, 46% at $2.98 and 23% at $5.88.
Author’s illustration. $60,000 per GPU matches $720M ÷ 12,000 GPUs (Cassava’s reported plan, which includes facilities) and one 8-GPU HGX H100 quote. Prices move daily; the global low uses the top of a $1.68 to $1.80 range. Excludes power, staff and financing.

The arithmetic is unforgiving. At $60,000 per GPU and a five-year life, a rental price of $2.98 an hour needs 46% utilisation just to return the capital, and $1.80 needs 76%. Power, staff and interest push the real break-even higher. The $60,000 matches one quote of $480,000 for an 8-GPU HGX H100 and Cassava’s reported plan, which includes facilities, so it is a proxy, not a disclosed unit cost. Break-even also depends on the hardware price: quotes of $250,000 to $320,000 for 8 GPUs would cut it to about 24% to 51% across those two rental prices.

Loop diagram: low utilisation leads to high prices, suppressed demand and costly capital; anchor tenants, pooled demand and phased capacity break the loop.
Author’s illustration, based on the World Economic Forum’s analysis of Africa’s compute market failures

Immature market signals

Structural costs

Hosting equipment: sovereignty is more than location

Putting GPUs in a local data centre delivers data residency, not full sovereignty. Control has to be tested layer by layer, because the bottom of the stack remains imported everywhere on the continent.

Five layers of sovereignty: data and operator access can be locally controlled, facilities are local with exposure, and software and accelerators are foreign dependent.
Author’s framework · five layers of control over hosted GPU equipment

The legal picture is also more nuanced than the marketing. More than 40 African countries have data protection laws, but most do not force all compute onshore. Nigeria has no single law requiring all data to stay in the country, and obligations vary by data type and sector.

Sector rules are where real demand forms. Nigeria’s central bank will require payment transaction data to be held locally from January 2027. Morocco’s decree 2-24-921 now sets requirements for cloud providers serving sensitive organisations, and Egypt’s regulations require licences for cross-border transfers.

Hosting modelData locationMain dependencyBest fit
Offshore neocloudAbroadForeign law and transfer safeguardsTraining on non-personal data
Foreign cloud, local regionIn-countryProvider remains under its home jurisdictionEnterprises needing residency and a mature platform
Local operator, imported stackIn-countryGPU vendor and export rulesRegulated enterprises and governments
Colocation of owned GPUsIn-countryOperator must fund, run and refresh the hardwareBanks, telcos and agencies wanting direct control
State facilityIn-countryPublic budget, foreign vendors, skillsSensitive government workloads

Colocation deserves more attention. Owners keep the hardware and the keys while a neutral operator supplies power and cooling. It is also more mature than local rental: Teraco offers DGX-ready hosting and iXAfrica takes 50 kW racks. The trade-off is that the owner carries depreciation, utilisation and refresh risk.

Before calling any hosted setup sovereign, buyers should ask:

  • Who owns the GPUs, and can they be repossessed or remotely restricted?
  • Who holds root access and encryption keys?
  • Which foreign laws reach the operator or its parent? US providers, for example, fall under the CLOUD Act.
  • Can the service keep running if vendor support or licences stop?
  • Are spare parts, trained engineers and backup power available locally?

Two cautions apply. Sovereignty often means choosing a dependency rather than removing one: Côte d’Ivoire’s national facility has US EXIM backing, while Senegal has reportedly worked with Alibaba Cloud. Local data can also make state access easier, so sovereignty for a government is not automatically privacy for its citizens.

Will web hosters and cloud providers survive?

Most will survive, but fewer will remain independent, and many will sell something different by 2030. This section is my forecast, not a measured result. The dividing line is simple: providers with a local advantage and limited GPU debt are safest.

Quadrant chart of local advantage against GPU exposure: managed local cloud and colocation are best placed; pure GPU resellers are most at risk.
Author’s forecast to 2030 · positions are judgement, not measurement
Provider typeForecast to 2030Main reasonBest survival move
Commodity web hostersSurvive through consolidationPrice pressure and loss of simple sites to SaaS buildersSell to a regional group, or move into managed services
Managed local cloudLikely to growLocal currency, support and sovereignty needsFocus on regulated sectors and SMEs
Telco cloudsLikely to surviveNetwork reach, balance sheet and enterprise accountsReuse network and edge assets for inference
Colocation operatorsLikely to growEvery model still needs power, racks and interconnectionAdd high-density and liquid-cooled capacity
Pure GPU resellersHighest failure riskThin margins, debt and global price competitionSecure contracts first, or partner rather than own
Global hyperscalersWill stay and expandScale, platform breadth and own capitalLocal partnerships and regional hosting

Web hosters are already consolidating. HostAfrica says it has acquired 19 businesses, including Web4Africa and GO54, while colocating rather than owning its data centres. Afrihost, which began as a web host, has since expanded into broadband and agreed to buy a majority stake in Echo.

The global picture is pressured, not terminal. In a worldwide survey of 446 hosting providers, 65% reported revenue growth in 2025, while 41% said customers were leaving for SaaS platforms. Half plan to expand professional services. The survey has no Africa breakdown and most respondents came from Europe and North America, so it is a directional signal, not African data.

Local cloud has a niche, not a takeover. In Europe, local providers have held about 15% of the market since 2022 even as sovereignty debates grew. African providers can expect a similar pattern, helped by local currency billing and regulated demand.

Sovereignty alone will not fund this shift yet. Gartner forecasts Sub-Saharan sovereign cloud IaaS at $31M in 2026 and $61M in 2027. Growth is fast, but the base is tiny next to Cassava’s plan of up to $720M.

GPU specialists face the sharpest test. Gartner expects more than 80% of neoclouds to lose independence by 2030, and executives quoted by The Information expect only seven to 10 survivors. Debt markets have tightened at the same moment. The infrastructure is more likely to change hands than disappear, which could create buying opportunities for well-funded African operators.

The practical conclusion is to treat GPUs as one product, not the whole company. Hosters and cloud providers should resell or rent capacity first, buy only against signed demand, and build their value in support, compliance and local relationships.

The real opportunity: scarce trust, not hype

The market is immature, but the opportunity is not speculative. Demand is backed by existing spending and regulation, and few players can credibly serve it.

Three demand signals exist independently of any AI hype cycle:

  • Spending to repatriate: Nigerian enterprises alone spend an estimated $850M a year on foreign cloud, according to a Luminate-commissioned study.
  • Regulated workloads: banks and public bodies in several markets face rules that a cheaper offshore GPU cannot satisfy.
  • Unserved users: the 7 million GPU hours of unmet demand shows latent need, though much of it will need subsidised or pooled access.

The same barriers that make this market hard also protect it. A credible provider must secure GPU supply, reliable power, local licences, government relationships and enterprise support, all at once. Few companies can do that, and a regulated customer will not hand sensitive data to an unproven operator. In these markets, trust is the scarce asset, and it takes years to earn.

This is my reading rather than a measured finding, but it changes the investment question. The test is not how large the market may become. It is whether a given provider has earned the right to serve the part of it that pays.

Speculative signalReal signal
Headline megawatts and long-term targetsLive capacity and a funded first phase
Memoranda of understandingSigned anchor customers and contract length
« AI-ready » or « sovereign » labelsDisclosed GPU model, count and rack density
Contact-us pricing with no track recordClear pricing and an audited service record
Grid power assumedContracted power with a backup plan
One large customer storyNamed customers across sectors

The most durable models will probably start small, as the World Bank’s view suggests. It finds most countries too small to carry these fixed costs alone and favours regional approaches. That supports phased capacity tied to contracts, anchor demand from government and regulated sectors, and pooled access for researchers and startups.

Four developments will show whether that model is taking hold:

  • MTN’s Lagos GPU phase, reported for late 2026, and whether it comes with public terms.
  • Nigeria’s payment data localisation deadline in January 2027.
  • Delivery of Egypt’s first 20 MW, Morocco’s 36 MW and Tunisia’s 7 MW against their stated dates.
  • Shared compute schemes such as the World Bank’s pay-as-you-go initiative and Smart Africa’s pooled GPU call.

Africa’s GPU market will not reward the loudest announcement. It should reward the operators that turn scarce capacity into dependable service, one contract at a time.

Notes on data and sources

Figures reflect sources available on 2 November 2026. Many project figures are company or government claims, and capacity totals differ by definition. Vendor and independent reviews, including Spheron’s and Francis Okafor’s, were used for availability checks because official GPU data is rare. Some World Bank figures are linked through FourWeekMBA’s summary; a second review checked them against the original report.

These facts could not be verified: Cassava’s GPU counts, Africa-specific results from the WebPros hosting survey, Huawei’s own specification for the Ascend 910B, whether any AMD GPU cloud is publicly available in Africa, and a primary IDC source for NVIDIA’s 81% share.