South Africa economic report · 18 November 2026
A factory may need electricity at 7 pm to finish a production run. A computing job due the next morning might wait. South Africa should test whether those two customers can share the capacity of a substation at Grootvlei. It should not promise a data centre round-the-clock power at a discount before finding out what the connection is worth to everyone else.
AI-generated. Not investment advice.
What changed at Grootvlei
On 15 November, the Department of Electricity and Energy proposed a 50 megawatt (MW) flexible AI load at the Mpumalanga power station. That is electricity a site might draw. The announcement identifies no contracted customer, financing, commissioning date or guaranteed reduction at peak. The 50 MW is neither new generation nor a measure of computing power delivered.
There is a useful comparison on the same land. In August, the department reported that Grootvlei’s hydroponic greenhouse was producing its first harvests. It separately described a proposed 195 MW solar project. The greenhouse has something to sell. The panels and the data centre, in the cited announcements, do not. Nor would 195 MW of solar nameplate capacity guarantee a 50 MW supply after sunset. Department’s Grootvlei site update, 19 August
Better plant availability gives Eskom more hours in which to find paying customers. Its open-cycle gas turbines generated 164.92 GWh from 1 April to 12 November, down 83.89% year on year. A developer still needs to know how much the local circuit can carry on a winter evening with one transformer unavailable. The national figures do not answer that question.
Sell two different promises
First establish what the existing substation can safely deliver, hour by hour, after protecting existing customers. Invite competing bids for two rights: a firm connection for work that cannot stop, and a cheaper connection whose draw the system operator can cut on agreed notice. Publish the capacity available, the rules for reaching the front of the queue and who pays to strengthen the circuit. A processor planning another shift should be able to compete on the same terms as a cloud company.
The discount has a ceiling: the cost that a dependable reduction saves the system. Charge for the firm portion even if the customer uses fewer units than expected. Require the developer to fund dedicated equipment or guarantee its unrecovered cost if the tenant leaves. That gives a lender contracted payments to assess and keeps an empty server hall from becoming a charge on other users’ bills.
If the operator made up all 45 MWh over the next six hours, it would add 7.5 MW on average to those hours. The operator must show that the circuit has room for that rebound. Set the comparison against a pre-agreed baseline at the grid meter; test response time, duration and repeatability. Record backup diesel separately. A drop at the meter could otherwise conceal a switch from grid power to fuel rather than a rescheduled job.
A 2025 field test in a working Phoenix data centre cut the power draw of a 256-GPU cluster by 25% for three hours while meeting the trial’s service requirements. That makes the idea worth testing. It does not establish the flexible share of an entire South African facility. Cooling and other equipment still run; a batch job may wait, while a bank’s live fraud check cannot. If a tenant needs firm power and will pay its full cost, a firm contract may be the better deal.
How much local income would the site buy?
Start with a deliberately simple price test. Suppose the 50 MW site drew 60% of its maximum on average over a 365-day year. It would buy 262.8 GWh. Every 10 cents per kWh left after variable supply costs would contribute R26.28 million a year towards fixed costs. That is a sensitivity, not a tariff or profit forecast. If annualised network works, maintenance and credit risk cost more than the margin plus guaranteed connection payments, electricity sales alone do not pay for the project.
A substation on the land can save construction time. It cannot supply the last stretch of fibre, a second network route or water during a dry spell. Before reserving capacity, compare a priced data-centre offer with the next feasible user. Ask both for a water budget at peak cooling demand and a funded connection plan. A cheap computer connection is expensive for the town if the municipality must pay for new water works or postpone a factory connection.
With official unemployment at 33.7% in the third quarter, a count of construction workers tells little about lasting income. Count apprentices who move into paid electrical or cooling work, permanent operating payroll and contracts local firms actually win. Imported servers can support service exports, but their purchase price is not South African income. The payoff for residents is wages, locally retained earnings and reliable services, less what households or other employers give up.
South Africa in three, five and ten years
In October, Treasury projected real GDP growth of 1.2% for 2026, 1.6% for 2027 and 2% by 2029. The scenarios below ask a different question: how much would real output per resident change if the country sustained each growth rate from 2027? They illustrate the stakes of nationwide investment decisions; none attributes national growth to Grootvlei.
| Scenario | 2029 3 yrs | 2031 5 yrs | 2036 10 yrs |
|---|---|---|---|
| Base 1.8% GDP/yr | 101.8 | 103.0 | 106.1 |
| Reform 3.0% GDP/yr | 105.4 | 109.2 | 119.3 |
| Downside 0.5% GDP/yr | 97.9 | 96.6 | 93.3 |
Calculation: each year, divide the real GDP growth factor by 1.012, then compound from 100. The population assumption of 1.2% a year holds constant the 2025–26 rate in Stats SA’s 2026 estimates, appendix 4. These are scenarios, not official forecasts. GDP per person measures production, not the median household’s pay.
Base: Generation gains persist, but firms still wait for municipal connections and freight. At 1.8% GDP growth a year, output per resident is only 3.0% higher after five years. Reform: Many sites shorten connection times, reuse assets where it is cheaper and charge users for upgrades; water and logistics improve alongside power. Firms can finance new shifts against sales they can actually deliver. Sustained 3% growth lifts output per person 19.3% over a decade. Downside: Public guarantees fund equipment before buyers commit and networks remain unreliable. At 0.5% growth, total output rises but output per resident falls 6.7% over the decade.
The next decision
Watch for the regulator’s decision on Eskom’s separate two-year tariff proposal for cryptocurrency miners and other flexible users. NERSA’s consultation, questions 3–7 and table 1 asks how to price the hour and location of consumption and who bears the loss if a customer leaves. Its provisional timetable listed a recommendation on 18 November and publication of a decision on 7 December. Neither date constitutes approval of a tariff for Grootvlei.
At Grootvlei, release the first tranche only against a signed customer, a costed local connection and a contract specifying the MW that can be taken off the peak. If the customer offers the illustrative 15 MW above, test it through the first full winter after commissioning. For example, require the full reduction within the agreed notice for all three hours on at least 90% of instructed events; reserve enough backup margin for missed calls. Expand only if the subsequent rebound stays within the circuit’s limit, other users keep their contracted service and invoices cover the incremental costs. If it cannot do that, keep the circuit available. A better use for an old coal station may be the customer willing to pay for the hours that are actually free.