The United Nations asked for three billion dollars. The only hard money on the table is one million.

On 23 September, Dario Amodei briefed the UN Security Council. He had published an essay days earlier arguing that the world must slow the rate at which AI capabilities improve.

At the same Security Council meeting, Pakistan's Deputy Prime Minister and Foreign Minister Ishaq Dar said this: "Pacing must not become a new form of gatekeeping."

Six words, in the same chamber on the same day as the pacing argument itself. I have not found them quoted in any Western coverage of UNGA 81, and they are the most important thing said about AI governance during high-level week.

They name what the inclusion language at these events is designed not to name. A slowdown negotiated among those who already hold the compute is not a neutral act. It is a decision about who gets to arrive, and when, taken by those who have already arrived.

I should be clear where I am standing. ImpactLab is Nordic-founded and I spend most of my working life inside European institutions. Issue 18 argued that the summer's agent incidents expose a structural gap in every jurisdiction's law; I think the risk case has strengthened. The argument here is narrower: that a governance regime which asks most of the world's population to wait, while offering a million dollars to help them prepare, is not a governance regime. It is a queue with a press release.

The arithmetic

What was asked for. On 11 September the Secretary-General's report A/80/817 proposed a Global Fund for AI Capacity-Building with "initial funding commitments in the range of $3 billion over two to four years" — described by the UN's own summary as an indicative starting point rather than a settled design. The report states that "financing for artificial intelligence capacity-building remains fragmented and leaves significant geographical gaps." Status: proposed. No pledges, no governance design, no General Assembly endorsement. ^[1]

What arrived. Nothing new that I can find. The one hard figure in circulation is a one million dollar project from the International Telecommunication Union with Google, Giga and will.i.am to train children in AI and robotics in South Africa, Kenya, Nigeria, Ghana and Côte d'Ivoire — and that was announced at UNGA 80, in September 2025, a year before the Secretary-General's report. As far as I can establish, no new hard money was attached to AI capacity-building during this year's high-level week.

What convened the universal forum. The UN Global Dialogue on AI Governance — the body giving all 193 Member States standing on AI, launched in Geneva on 6–7 July — was prepared, in its Co-Chairs' own words, with "limited dedicated resources." Travel support for developing-country participants came from voluntary contributions. The UN names the contributors: the ASEAN Foundation, the Friedrich-Ebert-Stiftung, the Heinrich Böll Foundation, and the Ministry of Foreign and European Affairs of the Slovak Republic. No core operating budget is published. ^[2]

Two German political foundations, a regional foundation and one small European foreign ministry paid to get delegates from developing countries into the room where the world's AI governance is formally equal.

And the scientific body is in the same position. The Independent International Scientific Panel on AI — 40 experts, co-chaired by Yoshua Bengio and Maria Ressa — runs on voluntary contributions routed through the Digital Cooperation Fund and, in the UN's own words, "not earmarked for the Panel's substantive work." ^[3]

What the speeches said

The rhetoric at UNGA 81 was better than at most of these events, and the consistency across delegations is the point rather than any single line. ^[4]

Mia Mottley, Prime Minister of Barbados, 24 September: AI is "a technology that will affect every State" and "must be governed by rules that every State has had the opportunity to shape."

Judith Suminwa Tuluka, Prime Minister of the Democratic Republic of the Congo, 23 September: "The digital divide cannot become an intelligence divide" — linking AI infrastructure to the critical minerals, cobalt among them, whose value is captured far from the communities that mine them.

Ronald Lamola, South Africa's Minister of International Relations, 24 September: "Africa must have a stake in shaping the rules that govern the development and deployment of Artificial Intelligence."

Kashim Shettima, Vice President of Nigeria, delivering President Tinubu's statement, 24 September: "We refuse to surrender our technological future to paranoia."

The President of the General Assembly, addressing ambassadors on 10 September: "the vast majority of us remain outside the rooms where many of the most consequential decisions are being made," and safeguards are needed so that "those with the least say on these systems do not bear the greatest risk."

The Group of 77 and China, to the Global Dialogue: AI is "first and foremost, a development issue," and "Financing should become an integral component of global artificial intelligence governance." ^[5]

And the five words that do the most work of anything I have read this year, from Loreto Bravo of Chile, a member of the Scientific Panel:

"Access is therefore not the same as benefit."

Why access is not benefit

Bravo's line has an empirical twin, from Sam Matekane, Prime Minister of Lesotho, at UN Digital Cooperation Day on 21 September: "Like many countries, we have a data center, but possessing a data center is not the same as being prepared for AI." ^[6]

That is three years of AI-and-development policy stated by a head of government in one sentence, about his own country.

The supporting numbers are stark and well sourced. Emerging and developing economies other than China account for 50% of the world's internet users and less than 10% of global data centre capacity — a figure published by the Observer Research Foundation, and among the most-cited compute-divide statistics available. ^[7] Private AI investment in 2025 ran at roughly $285.9 billion in the United States against $12.4 billion in China, with India at approximately $4.09 billion; the AI Index itself cautions that private-investment comparisons understate Chinese state capital. ^[8] Over 98% of African languages are unsupported by current models. ^[9]

But the finding that should trouble anyone in this field is not infrastructural. It is what happens when you audit the strategies.

The Africa AI Governance Index 2026, published by the Lawyers Hub in Nairobi, scored all 54 African states against 80 indicators — 4,320 evidence-linked observations, researched by 97 Africa AI Policy Fellows. Its central finding: every sub-region scores highest on Strategy and Vision and lowest on Implementation and Impact. Ring-fenced budgets, independent oversight bodies and monitoring mechanisms are rare. The launch framing is blunter than anything an outsider could say: Africa has learned to write AI strategy faster than it has learned to fund, staff, and enforce it. ^[10]

That index is the honest audit of the African Union's Continental AI Strategy, adopted in July 2024, whose first phase was meant to run through 2026. And since the figure circulates constantly: the $60 billion Africa AI Fund declared in Kigali in April 2025 has, eighteen months on, no identified funder and no evidence of capitalisation. Repeating it as money does the continent no favours.

Climate Week, and the two divides meeting

Climate Week NYC ran 20–27 September, overlapping high-level week, and the reporting consensus was unambiguous: AI dominated it. MIT Technology Review's headline on 24 September was "AI is dominating the conversation at Climate Week." TechCrunch's on the 28th: "The AI boom took over Climate Week and not everyone is happy about it." ^[11]

Simon Stiell, the UN's climate chief, warned that energy-intensive AI is driving up fossil-fuel pollution and energy costs, and said AI leaders are now on thin ice regarding public support and licence to operate. Bill McKibben, protesting outside Amazon's Manhattan headquarters on 22 September, said "the tech overlords planning our AI future can't think of anything more interesting to do than burn a lot of 18th-century gas and oil." ^[12]

And the clarifying fact of the whole month sits where the two divides meet. In May, Kenya's $1 billion Microsoft–G42 AI data centre ran into an energy wall. President Ruto's own description, as reported by Semafor: "To switch on that one data center, we would need to shut off power for half the country." Kenya has roughly 3,000 MW installed against a 10,000 MW target for 2030. The project's status is contested — Semafor reports it is unlikely to proceed and that the National Treasury withheld approvals, so funding was a factor alongside power, while Kenya's Special Tech Envoy has said the project was not suspended. ^[13]

Whatever its status, the arithmetic does not change: a single foreign-financed facility, in a country where millions lack reliable power, requiring a third of the grid.

That is what access looks like when it arrives without the rest. It is also the question Issue 14 argued Norway must answer about its own grid — and Norway, with 34 gigawatts of hydropower and a functioning consent process, still has no published allocation methodology. If the country with every advantage has not written the rule, the country being asked to surrender a third of its grid is negotiating without one.

Meanwhile the money meant to make the transition survivable has not moved. The $300 billion a year agreed at COP29 remains agreed and undelivered; COP30 did not reopen it. The Baku to Belém Roadmap to $1.3 trillion was only taken note of, and is designed to come mostly from private investment. At the UN's own climate finance dialogue on 22 September the figures on the table were that global energy investment reaches $3.4 trillion in 2026, of which $2.2 trillion is clean energy — but "roughly four out of every five dollars invested in clean energy have gone to advanced economies and China," while official development assistance declined 23.1% in 2025 against 2024. COP31 opens in Antalya on 9 November, four days after this issue. ^[14]

The discipline the argument needs

Here I want to be harder on my own side than is usual, because the inclusion case is often made in ways that do not survive contact with evidence.

"The Global South" is not a bloc. Tech Policy Press puts it precisely: the category lumps South Korea and Japan with low-income states and is defined by exclusion rather than shared position. The positions at UNGA 81 diverged sharply. Pakistan accepted the logic of pacing while opposing gatekeeping. Javier Milei of Argentina rejected preemptive regulation outright, warning it would grant "some people… the power to limit what everyone else can create." AI Now, Aapti and The Maybe were blunter still in January: "Under the veneer of Global South solidarity lie divergent interests that obscure the darker picture of extractivism," naming Rwanda and Nigeria as states positioning themselves as scaling hubs. ^[15]

Exclusion is not only imposed from outside. The sharpest version comes from Dhaka. Nabib Bin Zahid of BIPSS notes that South Asia "supplies the data, the labor, and a fast-growing share of the user base" that frontier AI depends on — then observes that India went from Bletchley to co-chairing Paris to hosting the 2026 summit, which demonstrates that a state can secure a seat when it organises for one. The rest of the region has not. ^[16] That is a Global South author declining the comfortable argument.

And the publics are not where the framing assumes. GlobeScan's survey across 33 markets — 2025 fieldwork, published in 2026 — puts India, Indonesia, South Africa and Kenya among the publics most likely to rate AI's dangers as very serious, alongside the United States and Australia, with Thailand, Japan, China, Italy and the Netherlands at the bottom. Pew's 37-country study finds concern about AI's economic effects is higher in wealthier countries. ^[17] There is no clean North–South gradient in concern about AI. Anyone asserting one is asserting something the data contradicts.

And while the UN deliberates, the blocs recruit. The World Artificial Intelligence Cooperation Organization, founded in Shanghai on 16 July, has signatories that CSIS describes as "overwhelmingly low- and middle-income countries in the Global South," and notes that "no G7 economy or EU member state has joined." CSIS is explicit about why: these governments prioritise compute access, AI skills and local-language data over frontier safety, and China is offering exactly that. Of the UN track, CSIS writes that "a third architecture for AI governance, led by the United Nations, is without capacity." ^[18]

That is the competitive reality. A forum with universal legitimacy and no budget is competing against one with a narrower mandate and a product offer.

For policymakers

Fund what you already built. The Global Dialogue and the Scientific Panel exist, carry near-universal legitimacy, and cost very little. Norway spent €20.4 million in September for roughly 3% of a European supercomputer. A fraction of that, as predictable core funding rather than project money, would transform a forum currently running on foundation grants. For a Nordic government seeking a high-leverage, low-cost contribution to AI governance, this is the best purchase available.

Underwrite Barbados. The Scientific Panel's second in-person meeting is hosted by Barbados this month. It is a genuine small-island entry point into frontier AI governance, and it happened because Barbados asked. Support it visibly.

Stop announcing billions and start disbursing millions. The $60 billion Africa AI Fund has no funder. The AI 10 Billion Initiative is described by its own host's page as a proposal, not committed funding. The $3 billion capacity fund has no pledges. Against all of that, the ITU AI Skills Coalition has over 70 partner organisations, more than 180 courses and has reached 20,000 learners — double its first-year target — on roughly million-dollar projects. The small real thing is worth more than the large announced one, and the pattern is now itself a credibility problem.

And if you argue for pacing, publish what the pacing is for. Dar's objection is answerable, but only with specifics: what capability threshold, assessed by whom, with what provision for countries that have not yet built anything. Issue 18 argued that no jurisdiction currently requires pre-deployment evaluation of frontier models. The same standard applies here: "slow down" without a published threshold, a named assessor and a transition provision is, on its face, gatekeeping — and the countries being asked to accept it are right to say so.

For businesses

Compute siting is now a licence-to-operate question. Stiell's warning and Kenya's grid arithmetic are the same signal. Any organisation siting AI infrastructure in a capacity-constrained jurisdiction needs a published answer to what that electricity would otherwise have done, and a public-interest carve-out it can point to.

Language coverage is a market, not a charity. Over 98% of African languages are unsupported. For anyone building public-service AI in these markets, that is the moat, and it is unclaimed.

And read the Africa AI Governance Index before your next market-entry memo. Fifty-four country workbooks, published openly, built by African researchers. There is nothing comparable for any other region and it is free.

For civil society

Hold the specific number, not the general grievance. "The Global South is excluded" has stopped doing work. "The Secretary-General asked for three billion dollars and the only hard money on the table is one million" does. So does "under 10% of data centre capacity against 50% of internet users."

Audit the strategies, not just the infrastructure. The Lawyers Hub found that African AI strategies score highest on vision and lowest on delivery — a finding African civil society produced about its own governments, which is worth more than anything produced about them from outside. It should be replicated in South Asia, Latin America and the Caribbean.

And resist the bloc framing, including when it flatters. Being spoken for as a single political subject is its own exclusion. The most effective interventions at UNGA 81 were national and specific — Lesotho on data centres, Barbados on rule-shaping, Pakistan on gatekeeping, DR Congo on cobalt — not collective.

What ImpactLab is doing

This sits under Pillar V, Trust, Inclusion and Integrity, and Pillar VI, Climate and Sustainability.

The Middle East ATLAS carries the Arabic-language and Urdu-language vector for AI literacy work under Pillar I, with a scoping note publishing in the fourth quarter of 2026. The Norway–Pakistan publication track is analysis published annually with a named diaspora and academic contributor list; as stated in Issue 13, there is no institutional partnership counterpart and we are not claiming one.

The Nordic Responsible AI Summit's second edition, in autumn 2026, carries a trust and inclusion track producing an inclusion-explicit cohort standard, and an AI for the public good track producing a deployment-readiness checklist.

The forthcoming Equity and Inclusion Hub is the structural home for this work. The Capital Roundtable, forming in the first quarter of 2027, carries the AI-and-development capital screen described in Issue 15 — and the arithmetic in this issue is the clearest argument yet for why that screen is the missing artefact.