Pax Silica — Why AI Governance Is Geopolitical Architecture and Not a Compliance Exercise

In December 2025, nine nations formalised that access to AI infrastructure is conditional on political alignment. The EU AI Act is being progressively softened under US lobbying pressure. The frameworks South African boards are adopting were designed inside an alliance South Africa did not join and was not asked to join.

Sign up for my Substack daily AI newsletter here.

See my AI Training course portfolio for corporate Business Leaders here.

Article link: https://open.substack.com/pub/johanosteyn/p/pax-silica-why-ai-governance-is-geopolitical?r=73gqa&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/

In December 2025, nine nations gathered in Washington and signed an agreement called Pax Silica. The signatories were the United States, the United Kingdom, Japan, South Korea, Singapore, the Netherlands, Israel, the United Arab Emirates, and Australia. The framework formalised what had previously been implicit: access to AI infrastructure — chips, computing power, frontier models — is conditional on political alignment. Sweden joined in March 2026. India joined in February. On 23 June 2026, three days before this article was written, Under Secretary of State Jacob Helberg stood at a ceremony in Washington and welcomed the European Union, Germany, and Greece. “We have gathered before to welcome a nation,” he said. “Today we gather to welcome a continent.” The European Commission confirmed the accession on 25 June 2026.

South Africa was not invited. South Africa was not consulted. South Africa is not among the signatories of an alliance that now includes every major Western economy and several of Africa’s most significant economic partners. And most South African boards do not know that Pax Silica exists.

CONTEXT AND BACKGROUND

Pax Silica is not a treaty. It is not legally binding. It is, in the description of the US State Department itself, a political statement — a declaration of alignment among countries committed to securing AI supply chains from semiconductors and critical minerals through to high-end manufacturing and AI models. But the distinction between binding and non-binding matters less than the structural reality the declaration creates. Membership shapes a country’s entire technological trajectory — from its computing capacity to its AI development pathway and its ability to access the frontier models on which digital transformation depends. This is closer in logic to a monetary system than a trade restriction: once the network is established, remaining outside it carries costs that compound over time.

France was the loudest sceptic in the EU’s internal deliberations. Paris called Pax Silica an attempt to colonise Europe and a direct contradiction of the digital sovereignty agenda Brussels had been promoting — arguing that joining locks Europe into an American-defined AI stack and export-control regime, trading autonomy for a seat at a table Washington built and chairs. The critics were overruled. The EU joined because the alternative — independent AI infrastructure at European scale — was judged infeasible within the timeframe the strategic competition requires. This is the sovereignty compromise: when the infrastructure you depend on is controlled by an alliance, the price of access is alignment. Europe made the calculation and signed.

The implication for South African organisations is direct: if the largest single market on earth, with GDP exceeding twenty trillion dollars, concluded it could not afford to remain outside Pax Silica, South African enterprises have zero leverage attempting to negotiate independent access to the same infrastructure through commercial service level agreements.

For South Africa, the EU’s accession is the most important development in this article’s argument. The EU was the only major governance bloc that had built a binding, risk-tiered AI regulatory framework — the EU AI Act — that was at least partially independent of the Pax Silica alliance’s commercial and strategic interests. That framework is now inside the alliance. The regulatory alternative that South Africa might have used as a governance reference point is now embedded in the same geopolitical structure that controls the infrastructure, the framework assumes.

The lobbying dimension makes the commercial dimension of this geopolitical architecture visible at the corporate level. In 2025, OpenAI, Meta, Google, and Nvidia spent a combined $50.9 million lobbying members of the US Congress. In Q1 2026 alone, Anthropic outspent OpenAI in federal lobbying for the first time, spending $1.56 million — a 344 per cent increase year on year. Groups linked to OpenAI and Anthropic collectively spent more than $23 million on 2026 midterm election messaging according to Federal Election Commission filings. The companies building frontier AI are simultaneously building the regulatory architecture that governs it — through lobbying, through policy proposals, through consultation processes, and through direct electoral spending.

For South Africa, the context is specific and urgent. South Africa’s Department of Communications and Digital Technologies gazetted its Draft National AI Policy on 10 April 2026. Minister Solly Malatsi withdrew it sixteen days later after it emerged that a portion of the policy’s academic reference list consisted of fabricated, AI-generated citations — sources attributed to authors and journals that did not exist. South Africa is targeting January 2027 for a revised policy. The consultation process for that revised policy is either underway or imminent. The question of who participates most actively in that process — which frameworks, which industry bodies, which international standards — is not a future governance concern. It is the most urgent governance question in South African AI policy right now. And it is being asked at the precise moment when the geopolitical architecture of AI governance has consolidated in ways that will be very difficult to reverse.

INSIGHT AND ANALYSIS

The battle for AI sovereignty analysis published by GIS Reports on 2 June 2026 names the emerging dynamic with precision: the new model is not regulation. It is a capture, but of an unfamiliar kind. Traditional regulatory capture describes firms that capture the state, bending public institutions to private interests. What is emerging in the AI era is a dual capture: states asserting control over firms whose technology has become critical to national security, while those firms simultaneously shape the regulatory frameworks that govern their commercial operations inside the alliance that gives them access to state power. The result is a governance landscape in which the distinction between public accountability and private interest has become almost impossible to maintain — and in which countries outside the alliance have no formal mechanism to influence the outcome.

The Anthropic case illustrates the enforcement mechanism precisely. On 27 February 2026, US Secretary of War Pete Hegseth directed the Department of War to designate Anthropic a supply chain risk to national security. The trigger was not a data breach or foreign ownership structure. It was a contract negotiation: Anthropic had refused to allow Claude to be used for autonomous lethal weapons systems and mass surveillance of American citizens. OpenAI signed a competing agreement accepting use for any lawful purpose and gained the contracts Anthropic lost. The message to the AI industry was unmistakable: when national security is invoked, the terms of AI deployment are set by the government. The company that maintained its safety policy was designated a security risk. The company that accepted government terms was rewarded with contracts. That is not regulation. It is capture — and it is the governance environment inside which the frameworks South African boards are implementing were produced.

I have previously written about Africa’s relationship to AI governance frameworks designed for other jurisdictions — arguing that the EU AI Act was built for mature digital markets with well-resourced regulators and that transplanting it to a continent defined by thin institutional capacity and informal data flows is not governance. The Pax Silica development deepens that argument significantly. It is not only that the EU AI Act was designed for European markets with European institutions. It is that the EU AI Act’s most independent element — its commitment to binding, risk-tiered regulation that constrained even the most powerful AI companies — has now been absorbed into an alliance whose primary stated purpose is securing AI supply chains from competition with China. The framework South Africa might have considered as a reference point for independent governance is now operating inside the same geopolitical architecture that controls the infrastructure it regulates.

The South African policy withdrawal is the most specific local illustration of what the absence of institutional capacity means in practice. The state used AI to draft the regulation meant to govern AI, without applying the human oversight it was advocating for. The result was a policy document whose academic foundations were fabrications — a governance vacuum created by the very technology the governance was supposed to address. When the revised policy is built, the frameworks most readily available to the drafters will be the ones most aggressively promoted by the Pax Silica members with the greatest interest in South Africa adopting governance that protects their commercial positions while appearing to address South African interests. The January 2027 deadline is not a planning horizon. It is a governance emergency with a clock.

IMPLICATIONS

For South African boards and executives, the Pax Silica accession carries three specific and practical governance implications.

The first is about framework provenance. When a South African risk committee reports that the organisation’s AI policy is aligned with the latest international guidelines, King IV’s principle of ethical and effective leadership requires the board to ask a specific and uncomfortable question: which international guidelines, produced by whom, through what process, and with what geopolitical interests embedded in their specific thresholds, exemptions, and enforcement design? The frameworks most available to South African boards were produced inside the governance contest the Pax Silica alliance is managing. They reflect the values, priorities, and competitive interests of their authors. A framework designed to protect the commercial interests of alliance members while appearing to protect the public is not a neutral governance standard. South African boards that adopt those frameworks without examining their provenance are not governing AI. They are inheriting someone else’s governance choices without understanding whose interests those choices were designed to serve.

The second is about South Africa’s January 2027 policy opportunity. The withdrawal of the April 2026 draft creates a specific and time-limited opportunity to build an AI governance framework that reflects South African interests, South African institutional capacity, and South African economic and social priorities. That opportunity is narrowing. The EU’s accession to Pax Silica means that the governance alternative most available to South Africa — a framework built around the EU AI Act’s binding risk-tiered approach — is now embedded in the same alliance architecture. South Africa’s revised policy will need to engage explicitly with the Pax Silica framework and its implications for South African organisations, rather than simply importing the global standards that Pax Silica members have shaped for their own purposes. The organisations that engage actively in the January 2027 consultation will shape the framework that governs AI deployment in South Africa for the next decade. Those who do not will inherit what others designed.

The third is about infrastructure dependency and its governance implications. South African organisations deploying US-hosted frontier AI models are operating infrastructure whose availability, terms of access, and regulatory governance are determined by an alliance South Africa has not joined. The June 2026 Anthropic shutdown demonstrated precisely what that dependency looks like in practice — models made unavailable globally without advance notice, without published criteria, and without recourse in any South African contract or service level agreement. The sovereignty compromise that France lost and Europe accepted illustrates the structural reality: when the infrastructure you depend on is controlled by an alliance, the price of independence is not merely inconvenience. It is exclusion. South African boards need to ask, specifically and urgently, what their AI continuity plans are if the models powering their operations become unavailable by a decision made in Washington — because that decision can be made at any time, by any government invoking national security, and South Africa will not be consulted.

CLOSING TAKEAWAY

On 23 June 2026, Under Secretary Helberg stood in Washington and told the assembled representatives of the EU, Germany, and Greece: “Look at what we have become. Not a treaty among the few, but a confederation of the capable — the makers, the shippers, the builders of the foundations on which this century will stand.” He was describing an alliance that now includes every major Western economy. South Africa is not among them. South Africa was not invited to become one of them.

Most South African boards are implementing AI governance frameworks designed by and for the confederation of the capable, without examining whose interests those frameworks were designed to serve, or whether those interests are the same as South Africa’s. The compliance checklist says to align with international standards. The governance question the compliance checklist does not ask is: international standards written by whom, for whom, and in whose interest?

AI governance is not primarily a technical exercise. It is primarily a geopolitical one. The frameworks being built inside Pax Silica were designed for Pax Silica members. They reflect Pax Silica priorities. South Africa’s January 2027 opportunity is to build something different — a governance framework that begins from South African conditions and South African interests, informed by but not subordinated to the frameworks that the confederation has designed for itself. That is not a rejection of international cooperation. It is the beginning of genuine governance.

Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net