Standfirst: Fourteen days of announcements across New Delhi, Cairo, Montreal and Strasbourg have made one thing plain: the fight over who governs artificial intelligence has moved from regulatory text to physical capacity and named languages. BRICS states are asking for independent evaluation of frontier models while insisting on sovereign compute of their own. Egypt has signed a national AI factory built on Nvidia silicon. Canada and Germany have bound research institutes together on safety and energy. For boards and public institutions, the practical issue is no longer whether the global AI order is fragmenting, but which dependencies will still be standing when it settles.
Featured image: the 18th BRICS Leaders’ Summit at Bharat Mandapam, New Delhi, where President Cyril Ramaphosa called for an international mechanism for the independent scientific evaluation of AI. Photograph: Siyabulela Duda / GCIS, published by Daily Maverick, 14 September 2026.
What the original report says
Daily Maverick reported on 14 September 2026 that President Cyril Ramaphosa, speaking at the 18th BRICS Leaders’ Summit in New Delhi on 13 September, called for an international mechanism for the independent scientific evaluation of artificial intelligence. He argued that oversight is standard practice in aviation, pharmaceuticals, nuclear power and financial institutions, and warned that technological capability must never move beyond the capacity to understand, supervise and control it. Reporting by iAfrica on 15 September added the four-part framework behind the proposal: meaningful human-control principles, mandatory reporting of serious incidents, progressively stronger safeguards as capability increases, and simultaneous investment in sovereign AI capacity for developing economies. South Africa is advancing this agenda without a national AI policy of its own; its draft was withdrawn on 26 April after fabricated citations were found in its reference list, and a redraft is due at Cabinet in November.
THE SCENARIO: A health ministry committee meets to approve a national screening programme. The clinical model is strong, the evaluation data are sound, and the roll-out covers three languages out of the eleven the service is legally required to serve. The vendor runs translation as a cloud call, priced per character. A finance officer asks what happens to the programme if the pricing changes or the endpoint is withdrawn during a diplomatic dispute. Nobody in the room knows who holds the termination right, and nobody has asked whether a smaller model could sit on the clinic’s own device. The decision dies quietly, pending clarification.
The Question
What happens to public institutions when the systems they depend upon are governed by countries and companies they cannot influence?
That question now sits behind a month of unusually concentrated news. In New Delhi on 13 September, South Africa’s president asked BRICS to build an international mechanism for the independent scientific evaluation of frontier AI. In Cairo on 8 September, Egypt signed the joint venture that will stand up its first national AI factory, built on Nvidia accelerated computing. In Montreal on 16 September, Canada’s AI minister told the country’s largest AI conference that nations which do not build will rent, and nations which do not make the rules will follow someone else’s. In Strasbourg the same week, the Canadian prime minister proposed pooling sovereign compute with Europe and warned that technology platforms now aspire to sovereignty themselves.
Read separately, these are procurement notices and conference speeches. Read together, they describe a single shift. Evaluation, compute and language are being claimed at the same time, by governments and by firms, often in the same month and sometimes in the same building. The question is not who has the best model. The question is who controls the three things a model needs in order to work at all: the authority to test it, the electricity and silicon to run it, and the language and data to make it relevant.
For leaders, the practical issue is that these three claims arrive through contracts rather than treaties. A capacity agreement, an evaluation clause and a language licence are all commercial instruments. They can be signed, renewed or withdrawn. They can also be designed. The sovereignty test is whether an institution knows which of the three it currently rents, and what it would take to own even one of them.
What Happened and Why It Matters
The BRICS proposal is the most explicit governance claim of the fortnight. Daily Maverick reported on 14 September that President Cyril Ramaphosa used the New Delhi summit to propose an international mechanism for the independent scientific evaluation of artificial intelligence, telling leaders that oversight is standard practice in aviation, pharmaceuticals, nuclear power and financial institutions. He quoted warnings from frontier developers themselves, including Anthropic chief executive Dario Amodei’s September 2026 essay calling for the pace of frontier development to slow, and OpenAI chief scientist Jakub Pachocki’s argument that scaling should be constrained by confidence in safety.
The framework matters more than the rhetoric. Reporting by iAfrica on 15 September set out four components: meaningful human-control principles, mandatory reporting of serious incidents, progressively stronger safeguards as capability increases, and simultaneous investment in sovereign AI capacity for developing economy countries. That final clause separates the proposal from most frontier safety advocacy. Ramaphosa is pairing oversight with capacity, arguing that developing economies should not be asked to accept governance of technologies they have no means to build. The venue is not incidental. BRICS adopted a Charter on Responsible AI at the 2024 Kazan summit, positioned as an alternative to the EU AI Act, and the New Development Bank launched a five billion dollar Digital Sovereignty Fund in 2025 for AI infrastructure.
There is a governance gap that the proposal does not answer. As iAfrica also reported, South Africa is advocating mandatory incident reporting internationally while its own national AI policy remains unwritten. The draft was withdrawn on 26 April after its reference list was found to contain fabricated academic citations, and the World Bank subsequently excluded South Africa from its map of countries with a national AI strategy. An expert panel chaired by Wits researcher Professor Benjamin Rosman is rebuilding the draft, with a redraft due at Cabinet in November and consultation targeted for January 2027. The critique writes itself, and it is also a fair illustration of the wider problem: the evaluation capacity countries are asking for internationally rarely exists domestically, by design or by neglect.
The compute layer moved in the same fortnight, in the opposite direction. Egypt signed a joint venture on 8 September bringing Vodafone Business, Cassava Technologies and Elsewedy Electric together as Africa Data Centres Egypt, with an initial investment of 200 million dollars rising to one billion on completion, capacity starting at 20 megawatts over three years and targeting 200 megawatts, and Nvidia accelerated computing delivered as GPU-as-a-Service. The same reporting notes the origin of the deal: Huawei had proposed 2,008 Ascend processors on a twelve-month build schedule, and Washington responded by sending Nvidia, AMD and Microsoft to Cairo in July to present a rival offer. Egypt concluded the Nvidia-backed deal without signing binding contracts with Beijing.
Language moved in parallel. Tether AI Research announced on 2 September that it had released open-source translation models for 19 Sub-Saharan African languages, including Hausa, Amharic, Yoruba, Swahili, Zulu, Shona, Nyanja, Tswana and Southern Sotho, covering roughly half of Africa’s population. The company states that the smallest model, at 800 million parameters, outperformed Qwen3.5-122B-A10B, TranslateGemma-27B and NLLB-3.3B across the FLORES-200, BOUQuET and SMOL benchmarks. Most of the claim has independent support: the underlying paper, accepted for EMNLP 2026 and available as a preprint, documents the benchmarks in full. The critical design choice, however, is not the benchmark score. The models are released as weights under an Apache 2.0 licence for research and educational use, run offline on ordinary phones, and keep data on the device rather than in a third-party cloud.
Canada and Germany supplied the alliance layer. On 16 September, the German Research Center for Artificial Intelligence and Mila, the Montreal institute founded by Yoshua Bengio, signed a memorandum of understanding at the ALL IN conference in Montreal, in the presence of German federal minister Karsten Wildberger and Canadian AI minister Evan Solomon. The partnership covers trustworthy and safe AI, agentic AI, and AI for the energy sector, with explicit commitments to methodology exchange and talent mobility. Two days earlier in Montreal, Reuters-quoted coverage of the same conference recorded Solomon’s framing: if the technology is not built here, it is bought from someone else; if the rules are not made here, someone else’s are followed. He noted that Canadian cloud dependence on American providers remains substantial, and that American authorities can reach data those providers hold.
That matters because the two tracks are additive, not separate. A governance framework without evaluation capacity is a signature. Compute without governance is a dependency with faster latency. Language access without device-level distribution is a service that can be repriced. The announcements of September 2026 are not wrong; they are incomplete, and the gap between the three tracks is where sovereignty is quietly decided.
The Sovereignty Risk
The first risk is evaluation dependency. Independent scientific evaluation of frontier systems currently happens in a small number of jurisdictions, funded by a small number of governments and firms. A country that accepts an evaluation regime it did not help build is accepting someone else’s definition of acceptable risk. Ramaphosa’s proposal is an attempt to correct that, but it arrives before most of its likely members have domestic testing capacity. The sovereignty test is whether a state can commission a test, understand the result, and act on it without routing the question through a foreign laboratory or a vendor’s own report.
The second risk is compute dependency dressed as capacity. Egypt’s AI factory is real infrastructure with real megawatts, and its strategic value is genuine. It is also built on a supply chain that Cairo does not control. The Egyptian deal was won by the United States after a competing Huawei proposal, which tells leaders something useful: competitive bidding is available to states that can move quickly and hold credible diplomatic ambiguity. It also tells them that the outcome is a choice between two foreign suppliers, not the creation of an independent one. Capacity purchased from a partner reduces latency and increases political leverage. It does not remove the right of another government to restrict that partner’s future supply.
The third risk is linguistic and cultural. Translation and language models decide whose words are first-class inputs. The Tether release shows that the technical barrier has fallen sharply: a device-sized model can serve languages that frontier systems handled poorly, and it can do so offline. The governance question is what an institution does with that fact. Where language access remains a per-call cloud service, it is priced, metered and withdrawable. Where the weights sit on the endpoint, the service survives a connectivity failure and a commercial dispute. Public services in health, education and justice are precisely where that difference becomes visible to citizens.
The fourth risk is the accountability gap created by multilateral ambition. When evaluation is proposed at the level of a bloc while domestic legislation is absent, responsibility becomes diffuse. South Africa’s case is instructive rather than disqualifying: an expert panel is rebuilding a policy whose earlier draft failed on verification. That failure is a governance lesson about the technology being governed, and it argues for building the domestic capability before exporting the international standard, not instead of exporting it.
For leaders, the practical issue is that four distinct dependencies are usually managed as one procurement line. The evaluation relationship, the compute contract, the language layer and the alliance membership each carry a different failure mode and a different exit cost. Treating them as a single vendor decision is how institutions end up with resilience they can describe but cannot demonstrate.
The TEE Method Response
The TEE Method reads this fortnight through three lenses. Talent is the domestic capacity to specify, test, procure and supervise AI systems. Enterprise is the institutional architecture of contracts, controls and exit rights. Ecosystem is the network of partners, standards bodies and research relationships that supply what no single institution can build alone. All three are exercised by the September announcements, and each one is currently underfunded in the institutions that most depend on them.
On Talent, the response is to stop treating evaluation as something purchased with the model. South Africa’s proposal is a demand for independent evaluation; the domestic equivalent is a small, named team that can run adversarial tests on the systems an institution already deploys, document failures, and hold a termination vote. The DFKI and Mila agreement is the clearest template available, because it commits to methodology exchange and talent mobility rather than to joint marketing. An institution that cannot evaluate cannot govern, and an institution that outsources evaluation has outsourced the decision.
On Enterprise, the response is to make each dependency a named line item with an owner, a renewal date and a written exit. The Egypt deal demonstrates that capacity can be won on competitive terms by states that present real demand. An enterprise version is smaller but structurally identical: know who can log in, which legal orders the operator must answer, whether workloads are portable with weights and data intact, and what the notice period looks like. None of those questions require new technology. All of them require somebody to hold the pen.
On Ecosystem, the response is to buy the layer that removes a dependency, not the layer that reproduces it. Offline, device-resident language models change what a ministry can promise: coverage in languages that were previously uneconomic, at zero marginal cost per query, with data that never leaves the clinic or the school. That is ecosystem capability in the TEE sense, because it converts an external service into an internal asset. The same logic applies to joining evaluation consortia, standards working groups and bilateral research arrangements whose value is access to method rather than access to a price list.
Read together, the three lenses set a threshold. Where an institution cannot staff an evaluation, cannot name an exit, and cannot hold a model on its own hardware, the announcement cycle is not an opportunity. It is a schedule of future notices it will receive rather than issue.
Sovereignty Test Matrix
The matrix below scores five domains against the fortnight’s developments. Each domain is scored from one to five, where one indicates full external dependence and five indicates demonstrated internal capability. The scores are diagnostic, not declarative; an institution should expect to score differently across domains, and the lowest score is the one that sets the risk.
| Domain | Weak signal (1-2) | Stronger signal (4-5) | Score |
|---|---|---|---|
| Political | Accepts foreign evaluation findings without domestic review; no national AI statute in force | Commissioned or co-built evaluation mechanism; statutory powers with an enforcing regulator | 2 |
| Economic | Compute and inference purchased under contracts with no exit clause or portability term | Reserved sovereign capacity; exit and portability written into every capacity agreement | 2 |
| Cultural | Service languages set by vendor coverage; translation metered per call in the cloud | Model weights resident on institutional devices; language coverage treated as a statutory duty | 3 |
| Intellectual | Safety and evaluation method licensed from a vendor; no independent testing team | Named domestic evaluation team; published methods; research partnerships with method transfer | 2 |
| Technological | Single frontier supplier; no demonstrated switch; no fallback runtime | Tested fallback for critical workloads; supply-chain provenance known; firmware servicing path verified | 2 |
Two patterns are worth naming. First, cultural sovereignty scores highest in the current environment, and that is new. Until recently, the cost of supporting smaller languages made local coverage an aspiration; the Tether release shows the cost has fallen to a device download. A low score in this domain now indicates a decision, not a constraint. Second, technological sovereignty remains the weakest domain for almost every institution, because supply-chain provenance and servicing rights are rarely examined at the point of purchase.
The threshold for action is a score of two or below in any domain combined with the absence of a named owner. That combination means the dependency exists, is material, and has nobody accountable for it. Political and intellectual scores below three deserve the fastest attention, because evaluation authority is the lever that governs every other domain: an institution that cannot test what it runs cannot meaningfully negotiate what it buys.
What Leaders Should Do This Week
Seven actions, each assignable to a named team, cover the fortnight’s developments without waiting for policy to settle.
First, inventory the four dependencies separately. Ask procurement for the evaluation relationship, the compute and inference contract, the language or localisation layer, and any alliance or consortium membership to be listed as distinct line items with their own owners and renewal dates. Most institutions will discover that all four sit with one supplier and one signature.
Second, write the exit for the largest contract before the next renewal. The test is concrete: how long to leave, at what cost, and in what format the data, weights and pipeline arrive. A supplier that cannot answer in a sentence has given an answer. Add portability language to the next amendment rather than to the next negotiation cycle.
Third, commission one adversarial evaluation of a system already in production, run by staff rather than by the vendor. The scope should include ambiguous instructions, conflicting objectives and conditions that tempt the system beyond its assigned remit. Publish the method internally. This is the cheapest available step towards the evaluation capability that BRICS members are now proposing to build at state level.
Fourth, test one workload on device-resident models and record the results. Translation and language access are the natural first candidates because the models are small, the benchmarks are public, and the marginal cost per query is effectively zero. Record accuracy against the current cloud service, and note which languages become available that were previously out of scope.
Fifth, brief the board on the jurisdictional question in plain language. Which legal orders can compel the operator of your systems to act, and which of those orders can be served without your knowledge. Directors do not need the technical detail. They need to know that data residency and legal exposure are different questions, and that a regional data centre does not change who can reach the workload inside it.
Sixth, map alliance exposure. List every consortium, standards group, research partnership and bilateral arrangement the institution participates in, and record what each one provides: access to method, access to capacity, or access to a market. Alliances that provide only a price list should be reviewed against alliances that transfer capability.
Seventh, set a review date tied to the policy milestones now in play. South Africa’s redraft is due at Cabinet in November and consultation is targeted for January 2027, which makes the first quarter of 2027 a natural checkpoint for any African or Global South exposure. Institutions with European or North American operations should run the same exercise against their own regulatory calendars rather than assuming the questions resolve in the same direction.
The Question Revisited
The question was what happens to public institutions when the systems they depend upon are governed by countries and companies they cannot influence. The answer, in TEE Method terms, is that influence is manufactured rather than granted, and it is manufactured at three specific points.
On Talent, influence begins with the ability to test. An institution that can evaluate a system independently holds the only asset that matters in a governance dispute, because it can state what the system does rather than what the supplier says it does. Ramaphosa’s call for an international evaluation mechanism is that argument at state level, and the domestic version is a small team with a published method and the authority to recommend termination.
On Enterprise, influence is written into contracts. The Cairo deal shows that competition for capacity is real for buyers who present credible demand and can hold ambiguity between suppliers. The enterprise equivalent is not a larger negotiation; it is a shorter one, governed by portability terms, servicing rights and exit costs that were agreed while leverage still existed. Dependency becomes structural at the moment the exit clause is deferred, not at the moment the contract is signed.
On Ecosystem, influence comes from owning the layer that removes a dependency. Device-resident language models are the clearest current example: coverage in 19 African languages, offline, at no marginal cost, with data that stays on the endpoint. That is not a technical achievement to admire. It is a procurement option to exercise, and it converts a service that could be withdrawn into an asset that cannot be.
The fortnight’s announcements point in one direction. Evaluation, compute and language are being claimed simultaneously by states and firms that understand all three are forms of authority. Institutions that leave the three in a single vendor relationship will receive the consequences of decisions made elsewhere. Institutions that separate them, fund the smallest one first, and write the exit before the renewal will find that sovereignty is a sequence of unremarkable contract decisions rather than a policy position. The window for making those decisions cheaply is open now, and the November and January policy milestones indicate it will not stay open for long.
[tt_consultation_booking]
Editorial note: This is a news-led Sovereignty Briefing analysis produced with Hermes-assisted research and source checks.
Sources
Daily Maverick, Tamsin Metelerkamp, “Ramaphosa urges caution on AI advancements, echoing developers”, 14 September 2026 — dailymaverick.co.za (also source of the featured image: President Cyril Ramaphosa at the 18th BRICS Leaders’ Summit, New Delhi, photograph by Siyabulela Duda / GCIS).
iAfrica, “Ramaphosa Proposes BRICS Mechanism for Independent Scientific Evaluation of AI, Citing Anthropic’s Amodei”, 15 September 2026 — iafrica.com
Tether AI Research, “Tether Releases Open-Source AI Translation Models for African and European Languages”, 2 September 2026 — tether.io
TranslatePsy-AfriSLM: High-Quality Data Scaling For Low-Resource Machine Translation, Tether AI Research, EMNLP 2026 (preprint, arXiv 2608.18655) — arxiv.org
Mila and DFKI, “Canada and Germany Strengthen AI Cooperation Through Strategic Partnership Between DFKI and Mila”, 16 September 2026 — newswire.ca
BNN Bloomberg / The Canadian Press, “Build out AI here or ‘rent’ from neighbours, minister Solomon says”, 16 September 2026 — bnnbloomberg.ca
Humphrey Theodore K. Ng’ambi, “African AI Sovereignty: Egypt Factory, Nyanja Model”, September 2026 — humphreytheodore.com (investment and capacity figures for the Egypt AI factory, citing Egypt’s Ministry of Communications and Information Technology)
GovInfoSecurity, David Meyer, “Canadian PM Floats Tech Sovereignty Alliance With Europe”, 17 September 2026 — govinfosecurity.com (Carney’s European Parliament address on pooling sovereign compute and AI safety protocols)
This article draws on the TEE Method™ framework from SOVEREIGN: Who Owns the Future? For the complete framework, see tonishatagoe.com.
Sign in to save