Insight · Power

Load-shedding-proof automation for Harare's power reality

Harare went from 18-hour cuts in early 2025 to 138 days without load shedding by May 2026. A stack diagram, a worked outage calculation and a checklist.

By Evert VorsterPublished 2026-09-148 min read

What happened, in four dates

January 2025. Lake Kariba stood at 3% full on 13 January, down from 13% a year earlier. Kariba South, with 1,050 MW installed, was producing about 185 MW — a fall of more than 80% in three months. In Harare the power went off around 4am and came back around 10pm. The World Bank’s estimate for what shortages cost Zimbabwe is roughly 6% of GDP a year. The supply mix that winter was coal 42%, hydro 35%, independent producers 3%, imports 20%.

August 2023, for context. Hwange Units 7 and 8 — two 300 MW coal units financed by a US$1.2bn China Eximbank loan — had come online. They are why the 2025 crisis was not worse, and why the 2026 recovery was possible once the rains returned.

May 2026. ZESA’s group CEO, Cletus Nyachowe, told Parliament and the press that the country had gone 138 days without load shedding, that a US$210m Afreximbank facility was funding power trading on the Southern African Power Pool, and that load shedding would end by December 2026, imports by 2027. He also said about 500,000 urban houses are still unelectrified.

July 2026. ZERA’s CEO, Edington Mazambani, put net-metered private solar at about 94 MW of connected capacity and utility-scale IPPs at 117 MW, and announced a ZERA–ZimStat survey to measure how much demand private solar now meets — because nobody knows.

The honest reading: the grid is better, the promise is credible but unproven, and every Harare business that bought an inverter in 2024 still has it. Design accordingly.

The principle: assume the grid returns, design as if it will not

An automation that dies with the office router was never an automation; it was a desktop application with a chat window. The Harare test is simple: which steps in the workflow keep working during an 18-hour cut at your premises? The answer should be “all the ones the customer can see”.

That leads to a four-layer stack.

The load-shedding-proof stack. Original diagram, aiharare.co.zw.

Layer 1 — cloud. The order table, the appointment diary, the reminder scheduler, the transcription of voice notes, the invoice draft: all of it hosted outside the building. Zimbabwe’s fixed subscriptions average ≈490 GB a month, so the office can afford to be a thin client; what it cannot afford is to be the server.

Layer 2 — channel. WhatsApp on phones. Customers are on mobile data averaging ≈3.8 GB a month; a text message costs them nothing they notice. Towers have their own backup power and the operators have spent 2024–25 hardening it, which is more than most offices did.

Layer 3 — on-site minimum, optional. A router and an inverter sized for a laptop and — if you are VAT-registered — the fiscal device. Starlink kits and fibre ONTs both need power; put them on the inverter circuit, but accept that when the inverter is flat the business still runs through layers 1 and 2.

Layer 4 — the outage playbook. The paper (or pinned WhatsApp message) that says: when the lights go, these three automated messages still go out, these two do not, and this person answers exceptions from their phone. Most outage failures are not technical; they are nobody knowing who is on.

A worked example (illustrative)

Take a private clinic in Avondale with 40 appointments a day, in the January 2025 pattern of cuts from 4am to 10pm. Round numbers; nothing here is a client.

Step	On-premises design	Stack design
Overnight reminder run (for tomorrow’s 40 patients)	Scheduled on the reception PC. PC off at 4am → 0 of 40 sent.	Scheduled in the cloud → 40 of 40 sent at 6am regardless.
Patient replies “reschedule” (say 6 of 40)	Reply lands on a desktop app nobody can open until 10pm → 6 slots lost for the day.	Reply processed in the cloud; new slot offered from the diary; receptionist sees it on her phone → slots refilled.
Pre-authorisation drafts for medical aid (say 10 of 40)	Drafted on the PC from the practice system → written by hand at 10pm, sent next day.	Intake captured on the phone, drafted in the cloud, clinician signs on a phone → sent same day.
Fiscalised receipt at the desk	Fiscal device on the inverter — works either way if the inverter lasts.	Same; the device is layer 3 in both designs.

On that illustrative day the on-premises design loses six appointments and delays ten claims; the stack design loses nothing except the fiscal device if the inverter runs flat. Multiply by the number of cut days you believe in — the 2025 pattern or the 2026 promise — and you have your budget argument. The point is not the arithmetic; it is that the two designs cost roughly the same to build and only one of them survives the city.

What about Starlink and solar?

They are layer 3. By Q4 2025 there were 67,057 VSAT (Starlink) subscriptions nationally and Starlink had overtaken all local fibre; roughly 94 MW of rooftop solar was net-metered by July 2026, with far more behind the meter. Both make the on-site minimum more robust. Neither changes the rule: if the workflow needs the premises to be up, it is the wrong workflow. Put the satellite dish and the panels on the inverter, and put the logic somewhere else.

Checklist before you sign off a design

List every automated step. Mark which ones the customer can see. Every one of those must be layer 1 or 2.
Name the person who owns exceptions from their phone during a cut. Write it in the playbook.
Confirm the fiscal device, router and Starlink/ONT are on the inverter circuit and estimate how long that circuit lasts.
Test by switching the office off for a day. Not a tabletop exercise — the breaker.
Re-read ZESA’s promise every quarter. If December 2026 arrives without load shedding, keep the design anyway; it costs nothing extra and the 2029 electrification target still leaves 500,000 urban houses to connect.