The plasma in a fusion reactor can change far faster than a person can react. At the DIII-D National Fusion Facility in the United States, a new system developed by the Princeton Plasma Physics Laboratory and Princeton University is transferring part of that control to artificial intelligence models.
PACMAN — Prediction And Control using MAchiNe learning — was tested in five real-world experiments. Its complete cycle of observation, prediction and action takes approximately 20 milliseconds. In one experiment, it anticipated a plasma instability around 200 milliseconds before it occurred and changed the conditions to prevent it. In another, it simultaneously coordinated all six gyrotrons used to heat the plasma. 1
This does not mean that the machine has replaced the people operating the reactor. The system incorporates physical limits that the models cannot exceed, and people continue to define its overall goals and parameters. But it introduces an important distinction. There are situations in which maintaining human authority can no longer mean approving every individual decision. By the time a person could intervene, the event would already have happened.
The question therefore shifts from who presses the button to who establishes in advance the space within which a machine can decide.
That shift seems far removed from what is happening in South Africa. Yet there, too, authority is exercised before the system starts operating. In a reactor, it takes the form of physical limits and control parameters. In a data centre, it appears in permits, electricity connections, land and access to water. In both cases, a decisive part of governance happens before the automated decision.
The cloud takes up land
South Africa accounts for around 70% of the continent’s data centre capacity, according to the figure cited by President Cyril Ramaphosa. Expansion has attracted major international companies and made the country one of the main hubs of Africa’s digital infrastructure. 2
It has also opened up a debate about the resources that infrastructure needs.
This year, the South African Human Rights Commission launched a dedicated process examining the human rights implications of data centres. The consultation received more than 250 submissions. Issues raised include electricity and water consumption, land use, environmental impacts and the lack of sufficiently consistent public information about those resources. 23
Five civil society organisations — Housing Assembly, Foxglove, Open Secrets, Research + Action and Planetary AI Collective — have called for a national public inquiry and consideration of a temporary pause on new approvals for hyperscale facilities until a regulatory framework and an independent assessment of their costs and benefits are in place. 4
Their report estimates that South Africa already has more than 60 known facilities and around 500 MW of declared capacity. It also argues that, using the industry’s standard cooling method, the Equinix facilities approved in Cape Town could consume more than 4.4 billion litres of water a year. These are estimates by the organisations submitting the complaint, not measurements of actual consumption, and should be read as such. 4
The industry disputes parts of this assessment. It points to the growing use of renewable energy and technologies intended to reduce water consumption. Eskom has also reported a capacity surplus during the recent winter peak in demand, after years of load shedding. 2
The important debate, however, does not need to be settled by assuming that data centres are necessarily good or bad.
The question is different. Who decides how much land, water and electricity can be devoted to turning a country into computing infrastructure?
In practice, that decision is not abstract. It involves land-use approvals, environmental frameworks, access to water and the electricity grid, disclosure obligations and oversight mechanisms. What the five organisations are calling for is precisely a national framework, an independent assessment and public capacity to oversee water, electricity and land use before hyperscale facilities expand further. 4
Housing Assembly also places the question within a longer history. The organisation connects the struggle over well-located urban land — close to work, schools and transport — to the displacement of communities towards the outskirts since apartheid. It warns that this same land, together with the water and electricity its members have been demanding for years, is now being allocated to data centres. 4
This extends an earlier AIthropology discussion of digital sovereignty, knowledge and the capacity to set limits. Having servers, data centres or models of one’s own does not, by itself, amount to being able to decide the conditions under which they operate.
Extraction begins before the data centre
In early September, another set of data emerged in South Africa that allows us to examine the problem from the opposite end of the chain.
The Protocol Gap: South Africa, produced by the Media Leadership Think Tank at the University of Pretoria’s Gordon Institute of Business Science, Journalism Relay Project and the International Fund for Public Interest Media, examined the robots.txt files of 263 South African news outlets. 5
Of those outlets, 74.1% had such a file. In April 2026, only 30.4% explicitly blocked at least one AI crawler, compared with 29.7% in December 2025. The change between the two measurements is marginal. The gap remains. 5
But the aggregate figure conceals another difference.
The capacity to identify crawlers, maintain rules and assess their consequences is concentrated among better-resourced publishing groups. Small, community, independent and local-language media generally operate with less technical capacity. 56
This does not mean that every absence of blocking amounts to an inability to act. Some outlets also have commercial reasons to remain open. Dependence on search engines and other platforms for audiences turns an apparently technical decision into an economic one. Restricting certain uses can mean losing visibility. But even allowing access requires information and technical judgement to understand what is being allowed. 56
Traffic figures compound this asymmetry, although the report does not establish that AI directly caused the decline. Leading South African news outlets lost, on average, close to a fifth of their daily visits between May 2025 and May 2026. And, according to Competition Commission figures cited in the report, the most popular AI chatbots generated just under two million visits to 222 South African news outlets between January and May 2025 — less than 1% of the 256 million visits those same chatbots recorded during that period. 56
Moreover, robots.txt is not a lock. It is fundamentally a declaration asking a crawler not to access certain parts of a website. Its effectiveness depends on whoever operates that crawler choosing to respect it. 6
Even the apparently simple act of saying “I do not authorise this extraction” therefore requires infrastructure.
The relevant difference is not simply between those who block and those who do not. It is between those who can choose a policy and sustain its consequences, and those who negotiate that choice from a much weaker position.
A paradox emerges. Content produced by local media or in languages with a smaller digital presence may be among the most valuable material for building systems capable of working with those societies and languages. At the same time, the people producing that content may be among those with the least technical and economic capacity to negotiate the terms of its use.
The whole chain
These three cases seem to concern different things.
An experimental fusion reactor in the United States. Data centres in South Africa. A text file a few lines long placed at the root of a local newspaper’s website.
But they can be read as parts of the same infrastructure.
AI needs the capacity to act. PACMAN shows what happens when that capacity must operate faster than a person.
It also needs the capacity to compute. Data centres show that this capacity takes up land, consumes electricity, requires cooling and enters political systems that must decide how to distribute finite resources.
And it needs material from which to learn and produce answers. South African media show that the capacity to determine how that material circulates is not distributed evenly either.
The three cases also share the same pattern: a space of authorisation defined in advance. At DIII-D, people set the physical limits within which PACMAN can decide. Data centres are expanding in a setting where the organisations that have approached the South African Human Rights Commission argue that those regulatory limits remain incomplete. And a robots.txt file is ultimately worth only as much as whoever operates the crawler decides beforehand. The question is not only who acts at any given moment, but who writes the conditions before the operation begins.
Human authority thus shifts further up the chain. It consists not simply in intervening when something happens, but in having the capacity to define in advance what can happen, with which resources and under what conditions.
The word “intelligence” tends to focus our attention on the model. But before the model come journalists, archives, cables, electricity grids, cooling systems, planning permissions, water and land.
After the model come actuators, machines, public administrations and people affected by its decisions.
Perhaps that is why an anthropology of artificial intelligence should not begin by asking only what a machine knows.
It should ask what material relationships make it possible for a machine to know and act, who sustains those relationships, and who retains the capacity to set limits on them.
References
[1] Princeton Plasma Physics Laboratory. (2026, 2 September). PACMAN AI framework for controlling fusion systems safely makes key decisions in milliseconds.
https://www.pppl.gov/news/2026/pacman-ai-framework-controlling-fusion-systems-safely-makes-key-decisions-milliseconds
[2] Magome, M., & Gumede, M. (2026, 4 September). Civil rights groups urge a halt to South Africa data centers boom amid water and power fears. Associated Press.
https://apnews.com/article/south-africa-data-centers-human-rights-590b64141c2262bd9a1aacb6d313a343
[3] South African Human Rights Commission. (2026). Call for public submissions on data centres and human rights in South Africa.
https://sahrc.org.za/index.php/sahrc-media/news-2/item/4479-media-advisory-south-african-human-rights-commission-calls-for-public-submissions-on-data-centres-and-human-rights-in-south-africa
[4] Open Secrets. (2026, 25 August). Submission to the South African Human Rights Commission: Data Centres and Human Rights in South Africa.
https://www.opensecrets.org.za/https-www-opensecrets-org-za-wp-content-uploads-2026-08-2026-08-20-pub-foxglove-sahrc-data-centre-submission-v265-pdf-2/
[5] Media Leadership Think Tank, Journalism Relay Project & International Fund for Public Interest Media. (2026). The Protocol Gap: South Africa — How South African News Publishers Are Responding to AI Crawlers.
https://ifpim.org/who/publications-resources
[6] Davis, R. (2026, 1 September). “Extraction without compensation” — the AI dilemma facing SA’s digital media. Daily Maverick.
https://www.dailymaverick.co.za/article/2026-09-01-extraction-without-compensation-the-ai-dilemma-facing-sas-digital-media/