Dry cooling bought South African data centres time. The H200 wave ends the bargain.

South African data-centre operators spent the last few years pushing water usage effectiveness, the ratio of water a facility evaporates to the IT load it carries, close to zero. They did it by switching to closed-loop, air-cooled chillers that sidestep evaporation entirely. That bought time. The arrival of Nvidia's H200-class accelerators, the next generation of high-density AI chips, will test whether the trick survives at scale.

The trade-off is not new: cooling dense AI racks costs water, energy, or both. Air cooling uses no water but draws more electricity than evaporative systems, especially in hot, dry African climates where ambient temperatures run high. Evaporative cooling is cheaper to run but consumes water. South African operators bet that buying more megawatts was easier than negotiating water rights. For now, that bet has held.

The H200 wave changes the math. Each new chip generation pushes rack power densities higher, which means more heat to reject from a smaller footprint. Air-cooled systems hit a wall there: they can move heat, but only at a steep energy cost. Operators globally are moving back toward liquid cooling (direct-to-chip or immersion), which removes heat far more efficiently but reintroduces water through the chilled-water loop, the cooling tower, or both. Even closed-loop systems bleed water through make-up cycles and humidification.

A December 2024 preprint modelling water-use efficiency across 41 African countries frames the constraint directly. The paper notes that African municipal water leakage averages 46%, against a global average of 39%, a structural loss that magnifies any new industrial draw. The same research found that operators in water-scarce regions have adopted closed-loop, air-cooled systems reporting WUE close to zero, but cautioned that air-cooled systems demand significantly more energy than evaporative systems under high-ambient conditions. The water-energy wedge is real, and the chip cycle is tightening it.

The per-inference water cost, estimated in the same preprint, is small in absolute terms but scales with deployment volume. Writing a 10-page report with a Llama-3-70B-class model uses roughly 0.7 liters of water; the same task on GPT-4 is estimated at up to 60 liters. A 120–200 word email costs about 0.13 liters on Llama-3-70B and around 3 liters on GPT-4. These are model-derived estimates, not metered figures, and the preprint is not peer-reviewed. They are, however, the kind of numbers ESG analysts and infrastructure funds are starting to read.

South Africa is the test case. TimesLive reported in July 2026 that the country is in the middle of a "data centre gold rush" that risks a resource crunch, a framing the Japan Times picked up two days later as an international syndication of the original. The bottleneck is the combined water-and-power envelope that determines whether a hyperscale campus can be sited, permitted, and financed. Grid capacity is tight, but water is the variable that is harder to swap.

That envelope is starting to show up in financing terms. Water intensity is moving from a sustainability metric into due-diligence language. CMU Africa's engineering faculty, commenting on the same research, noted that financiers and ESG frameworks are increasingly scrutinising water-use disclosures, and that operators in water-stressed regions are piloting non-potable and recycled water to push toward water-neutral design. An open dataset of water-efficiency metrics for African countries is now public, which makes the disclosure lane easier to enter and harder to ignore.

The ESI Africa webinar scheduled for 31 August sits inside this moment: water efficiency is being framed as a hidden constraint on African data-centre competitiveness, and the trade-off is moving into procurement and investor conversations. The trade-off has a clock now. H200-class chips are the next density step, and if the dry-cooling bet breaks, the bill lands in the next capex cycle, not the one after.