AI ServicesZW
  1. Index
  2. Guides
  3. Where should your AI run? Power, connectivity and hosting choices for Zimbabwean businesses
Guide

Where should your AI run? Power, connectivity and hosting choices for Zimbabwean businesses

Vendors talk about models. Your system will fail because of a power cut at the premises, a fibre cut on the Beitbridge route, or a foreign hosting bill that cannot be paid this month. This guide puts those three things first and gives you a decision table.

Power in 2026

Zimbabwe’s grid has had a good year by its own standards. In May 2026 ZESA’s group chief executive, Cletus Nyachowe, said the utility had recorded 138 consecutive days without load shedding, supported by a US$210 million Afreximbank facility used partly for power trading on the Southern African Power Pool, and that ZESA aims to end load shedding by the end of 2026, end imports by the end of 2027 and reach full electrification by 2030. That is a target, not a guarantee, and it says nothing about the last mile: a transformer fault in Waterfalls or a rural site 40 km from the nearest substation.

Design rule: anything at your premises — a fiscal device, an edge computer running a camera model, the phone that receives WhatsApp — must survive hours without grid power. Batteries and solar are part of the AI budget.

Connectivity

POTRAZ’s first-quarter 2025 figures showed 5G base-station deployments up 53% in a quarter and VSAT subscriptions up 32%; by the fourth quarter of 2025 VSAT subscribers had reached 67,057 — up from under 4,000 before Starlink’s market entry in late 2024 — and POTRAZ has equipped 98 community digital centres with Starlink. Satellite has changed the calculation for rural and mining sites: a camera system that once needed a fibre build can now sync over a dish. It has not changed the calculation for the phone in the customer’s hand, which is on a mobile network with a small data bundle.

Design rule: every field-facing part of a system works offline and syncs later; every customer-facing part is light on data.

Where the compute can live

OptionWhat existsSuitsWatch
Edge (at your site)A small computer or the fiscal device itself; camera models run locallyVision, gate control, anything that must work when the link is downPower (solar/battery), physical security, someone to reboot it
Zimbabwe data centreTelOne describes its Harare facility as a “Tier 3 Data Centre Environment” and offers colocation, virtual machines and storage; Dandemutande operates facilities in Harare and Bulawayo (and Johannesburg) with public-cloud and DR services; the market is dominated by telcos and the state, with a large Econet facility reported as comingData you want to keep in-country; systems with modest compute; DRGPU capacity for large models is limited; ask what is actually available
Regional GPU cloudCassava Technologies’ NVIDIA-powered AI factory in South Africa, live in Cape Town with a 20 MW Johannesburg site planned; offered as GPU-as-a-service and AI APIs (CAIMEx); Econet AI, launched April 2026, gives Zimbabwean customers access to Cassava AiCloud computeTraining or hosting your own models; Southern-African data residencyCross-border data transfer terms; USD billing; latency is fine, forex is the issue
Global cloud and model APIsThe large model providers and hyperscalers; most listed local vendors build on theseLanguage models for chat, extraction, summarising; fastest to startData leaves the region; USD invoices from abroad; provider terms change

The four-question decision

Decision tree with four questions: must it work when the link is down (yes: edge), must the data stay in Zimbabwe (yes: local data centre), do you need GPUs for your own models (yes: regional GPU cloud), otherwise global model APIs through a local vendor who invoices in one currency. Q1 Must it work when the internet link is down?Q2 Must the data stay in Zimbabwe?Q3 Do you train or host your own models (GPUs)?Q4 Otherwise: who carries the foreign bill? YESYESYES→ EDGE at your site + battery/solar; sync later.Vision, gates, fiscal devices, field capture.ZIMBABWE DATA CENTRE (colocation / VM).Modest compute; declare location on the SI 155 licence.REGIONAL GPU CLOUD (South Africa).Cross-border transfer clause; USD billing.GLOBAL MODEL APIs via a local vendor.Vendor invoices you locally, in one currency, with a cap. nonono
Figure — Hosting decision in four questions. Original diagram. Most SME chat and document services end at Q4.

Most services on this catalogue — WhatsApp assistants, document extraction, report narratives — end at question four: they run on global model APIs through a Zimbabwean vendor. The decisions that matter there are contractual: where the data goes (declare it on your licence), who holds the foreign account, and whether the vendor invoices you locally in one currency with a cap on pass-through.

Vision, gate and field systems end at question one, and their budgets should say so: solar, batteries, an edge device and someone who can restart it.

Questions for the vendor

Take the six from the “hosting, power and continuity” section of the question bank. The three that matter most: What queues and what is lost during a 40-minute outage at our site? Where exactly does our data go, by country? How do we pay the foreign bills if forex is constrained this quarter?

The integration sheet (S-04) covers the connections themselves; this guide is about where the thing lives.

A budget line that is always missing

Whatever the answer to the four questions, put one line in every AI budget: keeping it up at our site. For a WhatsApp assistant that is nothing more than a charged phone and a data bundle. For a document pipeline it is a scanner or a phone-camera guide and a place for originals. For a camera system it is solar, batteries, an edge device, a locked cabinet and the name of the person who restarts it after an outage. Vendors rarely quote this line because it is not their product; it is, however, where most systems here quietly stop working.