Artificial Intelligence (AI) has found its way into general conversation after the emergence of large language models like ChatGPT, capable of generating text on a near-infinite range of topics. However, the discussion is increasingly turning to the search for Artificial General Intelligence (AGI), which, once discovered, could entirely change the game.
Narrow AI is trained to perform a specialist function such as natural language understanding, playing chess, or identifying spam. In contrast, AGI will resemble a genius polymath, able to tackle a wide range of challenges across any field. AGI is expected to revolutionise sectors such as healthcare, scientific research, and automation. This automation could lead to increased productivity and efficiency but may also result in substantial job displacement. Naturally, AGI’s potential applications come with environmental, social, and governance (ESG) considerations that must be addressed. These considerations will be the subject of a separate article.
Before considering the ESG implications of AGI, it is worth noting that its discovery is currently constrained by three main factors: processing power, access to information, and energy supply.
AGI requires vast amounts of computational resources (i.e., chips) to mimic human-level intelligence, far surpassing the capabilities of today’s AI models. In addition, training AI and AGI requires access to large, high-quality datasets. As human-created data is a finite resource, AIs and AGIs may create new knowledge autonomously to teach themselves-raising important questions about privacy, control, and the implications of AI generating its own data.
However, none of this is possible without sufficient energy.
Training AI requires huge amounts of energy. Statistics published by OpenAI indicate that training ChatGPT-3 took 34 days and consumed approximately 1,248 MWh of electricity during that time. By contrast, the next iteration, ChatGPT-4, took 100 days to train, consuming an estimated 50 GWh of electricity. This amount of energy is roughly enough to power 17,000 South African homes for a year.
In simple terms, current models of AI typically train the program (resulting in a knowledge cut-off date), review the program, and then publish it to the general public for use. When an AI receives an instruction, it then generates a response based on the knowledge it acquired during training. The largest computational power and energy demand therefore occur during training, although processing power and energy are required when instructions are fed to the AI. However, AI systems are now evolving towards an “always learning” phase, where they continuously update and refine their knowledge in real-time. The energy demand associated with their operation will thus increase and become sustained, as this shift requires constant computational processing. Continuous learning involves ongoing data analysis, model adjustments, and real-time decision-making, all of which consume significant computational power and electricity. When AGI emerges, this effect will be amplified. AGI’s ability to understand, learn, and apply knowledge across a wide range of tasks in a human-like manner will necessitate even more complex computations and perpetual data processing, further escalating electricity demand exponentially.
The inability of regions or nations to meet these electricity demands could limit access to AI and AGI technology to regions and entities with stable, affordable, and increasingly green energy supplies. This will likely shape who can participate in the AI and AGI revolution, potentially reinforcing global and regional inequalities if not addressed equitably.
As AI becomes more prolific, more data centres will likely be constructed in South Africa both to train AI and to locally host AI software to reduce latency (the time delay between a user’s instruction and the system’s response). Consequently, both existing and future data centres in South Africa will need to be robust, secure, and always online (and always powered), or we risk missing out on the competitive advantages associated with the use and development of AI and AGI.
To ensure South Africa benefits from the AI revolution, it must adopt a proactive approach.
The first and most critical step is to prioritise and incentivise investment in renewable energy. This will prevent the stalling of AI development in South Africa. Powering data centres with renewable energy, coupled with storage solutions like green hydrogen fuel cells or batteries to address intermittency issues, has many benefits, including:
reducing greenhouse gas emissions;
reducing longer-term operational costs, particularly in light of the recent proposed Eskom tariff increases for the next three years;
alleviating strain on the national grid and energy losses due to wheeling if renewable facilities are co-located with data centres;
enhancing competitiveness in the global economy where the source of energy (renewable vs. fossil fuel) for data centres may attract or deter foreign investment; and
where possible, feeding excess power back into the grid to contribute to energy security for the broader community as well as earning revenue for the data centre.
South Africa is already taking steps to encourage energy self-sufficiency at data centres, as evidenced by the National Data and Cloud Policy, published on 31 May 2024. The policy states that data centres with self-sufficient energy and water sources should be “prioritised.”
To drive investment in renewables in the context of data centres and the country at large, there are always two options – carrot or stick. The carrot (rewarding proactive efforts) is often preferred, as it encourages investment and innovation, making adoption more appealing and reducing resistance. Thus, incentives such as tax breaks and reduced tariffs on renewable energy hardware should be considered to foster growth in the renewable energy sector, creating green jobs and green data centres.
Contrary to this, the South African government recently elected not to renew the 2023 tax incentives on solar panels and imposed a 10% tariff on all imported solar panels. These actions have the potential to discourage investment in renewable energy. South Africa should consider adopting the example of other African nations that have proven successful in incentivising renewable energy investment. For example, in 2021, Kenya exempted solar and wind energy specialised equipment from VAT; in Ghana, all imported solar panels are VAT-free; and in Botswana, equipment and machinery including solar panels and inverters have been exempted from import duty.
Embracing renewable energy investment is imperative for South Africa to fully capitalise on the opportunities presented by AI and AGI. By prioritising sustainable power sources for data centres, the country can meet the substantial and continuous energy demands of advancing AI technologies. Proactive action today will enable South Africa to harness AGI for economic growth and social development, paving the way for a more equitable and sustainable future.