AI at work, at scale, and under strain

Shoprite's chief technology officer has told TechCentral that around 12 000 of the retailer's employees are already using generative AI tools (software that produces text, code, or other content from prompts), and that the group has built its own internal system to manage and reduce the cost of running those tools. Looking further ahead, the CTO envisions a workplace where new hires arrive with personal AI agents, software programmes that can carry out tasks on a user's behalf, as a normal part of how they work. For South Africa, where Shoprite is one of the country's largest private employers, the scale of this rollout and the direction of travel raise practical questions about how AI will reshape work and hiring in the retail sector.

NEC XON, a South African technology company, has partnered with Hewlett Packard Enterprise (HPE) to offer what it calls a 'sovereign AI Pod' – a self-contained hardware and software system that lets organisations run AI workloads on their own infrastructure rather than through a third-party cloud. According to ITWeb, the offering grew out of R1 million in internal AI-related savings the company recorded. For South African organisations concerned about data residency and control over sensitive information, the product represents a locally available option for keeping AI processing in-house.

Meta's Facebook and Instagram platforms ran hundreds of advertisements promoting AI-powered applications that generate child sexual abuse material (CSAM), including ads featuring real photographs of named minors, according to an investigation by the Tech Transparency Project published this week. The platforms took days to remove the content, and 332 such ads went undetected by Meta's own systems this year. South Africans are among the many users of these platforms, and the findings raise direct questions about whether global platform operators are doing enough to protect children here, at a time when local regulators and civil society are still developing frameworks to address AI-generated abuse material.

South Africa's data centre sector faces material risks from electricity and water constraints, according to S&P Global, which flags low water resilience as a particular concern, including water stress in Johannesburg. Data centres are the physical infrastructure that AI systems run on, so limits on power and water directly affect how much AI capacity the country can build and sustain. The findings point to a structural challenge for South Africa as demand for AI computing grows.

African Languages Lab, an AI research company founded in 2020, has launched Mansa, a multilingual platform that supports around 30 African languages in production, including Zulu, with plans to expand toward 1,000 languages over time. The company says it has spent roughly a decade building what it describes as the largest African-language dataset in existence, covering more than 70 languages and over 100 billion curated tokens (the units of text used to train AI models), alongside more than 19,000 hours of validated speech recordings. For South African users, the work has direct relevance: Zulu is among the supported languages, and the company's research highlights a safety concern with practical implications here, namely that AI models are significantly more likely to produce harmful outputs in under-resourced African languages than in English, because safety training is weaker where training data is scarce.

Anthropic, the company behind the Claude AI system, says rival Chinese AI developers have used networks of tens of thousands of fake accounts to collect Claude's responses at scale and use them to train competing models, a technique known as model distillation. According to reporting by CNBC cited by Memeburn, stolen credentials sold on dark-web markets are now part of the infrastructure enabling this access, making it harder for Anthropic to block. Memeburn draws a practical note for local businesses: as geopolitical tensions shape which AI tools remain accessible and on what terms, understanding where a model's capabilities came from and what safety measures survived its training is becoming a routine part of choosing an AI vendor.