The Centaora Signal — September 2026

The build-out becomes an asset class.

Spending plans for AI infrastructure rose throughout 2026, and the financing of that infrastructure began moving off corporate balance sheets into the hands of asset managers.

The escalation in spending plans

The four largest cloud providers (Amazon, Alphabet, Microsoft and Meta) began 2026 guiding to more than US$640 billion of combined capital expenditure. After second-quarter results in late July, that figure stood at US$735–750 billion (Platformonomics). By July, Amazon guided US$220 billion, Alphabet US$195–205 billion, Microsoft about US$190 billion for calendar 2026, and Meta US$130–145 billion. In the second quarter alone they spent almost US$172 billion, 77% more than a year earlier and more than they spent in the whole of 2023.

Three features of the revisions are worth understanding. First, the direction was overwhelmingly upward: Amazon, Alphabet and Meta all raised their ranges, and Microsoft held near its April level. Second, part of the increase is price, not volume: memory and component inflation explain a meaningful share. Third, markets have started to discriminate, rewarding some companies and punishing others depending on how convincingly they linked spending to revenue (Motley Fool). The question has shifted from “how much will they spend?” to “how quickly does the spending pay back?”

What the guidance implies for the rest of the year. By the end of June, the four companies had spent just under US$305 billion. If they meet full-year guidance, they must spend roughly US$430–445 billion in the second half, some 40–46% more than in the first. The guidance does not describe a plateau; it describes a further acceleration. That places heavy demands on everything below it in the stack: memory deliveries, power connections, construction crews and cooling equipment. Any shortfall will show up first as delayed projects, and later as revised guidance.

The financing layer

Alongside its August results, NVIDIA announced AI-infrastructure financing partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, designed to mobilise more than US$500 billion of third-party capital over time (Pulse 2.0).

This marks a change in how computing capacity is financed. Data centres are beginning to be treated like power stations, toll roads and ports: long-lived assets funded by pension funds, insurers and private credit, with returns tied to utilisation contracts. The approach spreads risk across more investors. It also spreads exposure: if AI demand or pricing disappoints, the losses would no longer sit only with technology companies.

Electricity: demand outruns forecasts

The International Energy Agency's Electricity Mid-Year Update 2026, published on 23 July, projects global electricity demand growth of 3.6% in 2026 and 3.8% in 2027, up from 3% in 2025. Consumption is forecast to reach about 30,700 terawatt-hours by 2027 (JM Financial). Industry, appliances, cooling, electric vehicles and data centres drive the growth, and renewables are expected to overtake coal as the world's largest source of electricity this year (Reuters via Engineering News).

The Middle East energy shock raised generation costs but did not slow demand. For the AI economy, power is simultaneously more needed and more expensive, a combination that favours locations with abundant, low-cost and reliable generation.

Why this matters: the dependency chain lengthens, and so does the list of winners

June's research discipline was to follow the dependencies. This quarter extends the chain. If AI spending grows, ask who supplies memory. If data centres grow, ask who finances them and on what terms. If power demand grows, ask which regions can offer reliable, affordable electricity. Each answer points to a different set of winners than the headline technology companies.

Evidence note. Strong for capex guidance, which is company-reported. Moderate for the financing partnerships, which should be read against NVIDIA's own announcement.

The counter-view. High capex is not the same as high returns. Free cash flow at major cloud providers is under pressure. If utilisation of new capacity lags, the financing structures now being built could amplify a downturn rather than cushion it.

Questions for further research

• What contract terms underpin third-party data-centre financing, and who bears utilisation risk?
• Which African grids could offer the reliable, low-cost power that compute now seeks?
• How much of 2026 capex growth is price inflation rather than added capacity?

Resilient growth, uneven gains.

The world economy absorbed a war-driven energy shock better than expected, but the growth that remained concentrated in economies connected to the technology cycle.

Growth: a V-shaped forecast with stalled disinflation

In its July 2026 World Economic Outlook Update, the IMF projected global growth of 3.0% in 2026 and 3.4% in 2027. Its deputy research director described the path as a V-shaped recovery: weaker growth this year than the pre-war forecast, followed by a rebound. Global headline inflation was revised up to 4.7% for 2026. Emerging and developing economies are forecast at 3.8%, the United States at 2.3%, advanced economies at 1.7% and the euro area at 0.9%. Sub-Saharan Africa is forecast at 4.3% in 2026, rising to 4.5% in 2027 (bne IntelliNews), though the IMF warns that higher food and energy prices will worsen poverty and food insecurity in exposed countries.

A note on comparability: June used the World Bank's 2.5% global forecast (market exchange rates); the IMF's 3.0% uses purchasing-power weights. The two are not directly comparable.

Across the continent: one average, many trajectories

Sub-Saharan Africa's 4.3% forecast is higher than the world's 3.0%, but the regional figure hides a spread wide enough to matter for any organisation operating across borders.

• Côte d'Ivoire is forecast by the EBRD to grow 6.1% in 2026 and 6.5% in 2027. It completed its IMF-supported programme in June and issued a US$1.3 billion Eurobond (EBRD).
• Ghana is forecast at 5.0%. Its public debt fell to 40.6% of GDP from 62% in 2024, and the IMF completed the final review of its programme in July.
• Nigeria is forecast at 4.1% by the IMF, supported by macroeconomic reforms and favourable terms of trade.
• South Africa is forecast at 1.1%, the slowest of the large economies, despite having the continent's deepest financial markets and most developed data-centre sector.

The lesson is that fiscal credibility is being rewarded. Ghana and Côte d'Ivoire, which have stabilised public finances, are seeing investor confidence return. Economies where debt service absorbs a large share of revenue have less room to invest in the infrastructure that would connect them to the technology cycle.

Kenya: steady growth under fiscal pressure

Kenya illustrates the fiscal side of this pattern. The EBRD's Regional Economic Prospects forecasts growth of 4.6% in 2026 and 4.9% in 2027, led by construction, services and mining, while agriculture and manufacturing lag. The constraints are fiscal and external: public debt at about 70% of GDP, debt service absorbing around half of government revenue, higher fuel and freight costs, and political uncertainty ahead of the August 2027 elections (Capital FM).

Trade: carried by electronics

The WTO's Goods Trade Barometer rose to 102.0 in its 9 September release, up from 101.7 in June. Electronic components remain the strongest driver at 104.9, reflecting demand for goods that support AI investment. Container shipping, at 99.6, is the only component below trend. The WTO estimates that sustained AI investment could add 0.5 percentage points to merchandise trade growth in 2026, against a baseline of 1.9% (summary).

Why this matters: globalisation is being rewired around high-value components

June argued that globalisation is being reconfigured rather than reversed. September's data sharpens the picture: trade in high-value technology components is holding up the aggregate, while flows of ordinary goods soften. Economies that participate in electronics and digital supply chains benefit; those that rely on shipping bulk commodities feel the headwinds.

Evidence note. Strong. IMF and WTO figures are official; country forecasts vary by institution.

The counter-view. The IMF's baseline assumed the Strait of Hormuz would reopen from mid-July. The shipping situation changed several times during the quarter, so energy-sensitive forecasts carry more uncertainty than usual.

Programmable finance: from debate to rulebooks

In June, The Signal asked which payment functions would become programmable first and what governance they would need. This quarter, regulators in the United States and Kenya answered the governance half of that question.

The United States writes the operating manual. The GENIUS Act, signed in July 2025, created the first US federal framework for payment stablecoins backed one-for-one by cash and short-term Treasury bills. In August, the Treasury proposed the core rules defining who may issue, offer and sell them. The prohibition on issuance by unlicensed firms takes effect on 18 January 2027, and the sale of coins from unlicensed issuers will be cut off in July 2028 (Mintmark Brief). The FDIC's proposal would confirm that deposits in tokenised form remain insured deposits (FDIC), giving banks two routes into programmable money.

Kenya builds a complete regime. On 22 July 2026, the National Treasury gazetted the Virtual Asset Service Providers Regulations, 2026 as Legal Notice No. 134, creating Kenya's first comprehensive licensing system for crypto-asset businesses (PwC Kenya). The design divides the market by function: the Central Bank of Kenya licenses wallet providers, payment processors and stablecoin issuers; the Capital Markets Authority licenses exchanges, brokers, token issuance and tokenisation (CM Advocates).

• Extraterritorial reach. A foreign platform that targets Kenyan users or earns revenue from Kenya must comply, whether or not it has an office in the country.
• Stablecoins as payment tools, not savings products. Issuers may not pay interest or holding rewards (AU-Startups).
• Monetary safeguards. Regulation 83 allows the CBK to direct licensed firms to restrict stablecoins issued outside Kenya (AU-Startups).
• Deadline. Existing operators must be licensed by 4 November 2026; company fines reach KSh 25 million.

Why this matters: Africa is writing the rules for the next layer of digital money

Kenya built one of the world's most widely used mobile money systems. It now has a rulebook for the next layer of digital money, with consumer protection and monetary control placed first.

Kenya is not acting alone. It now sits alongside South Africa and Nigeria as one of the continent's leaders in bringing digital-asset markets under formal regulation. South Africa took a different route: rather than a stand-alone regime, it requires crypto-asset service providers to hold a financial services provider licence and meet anti-money-laundering obligations under its existing financial intelligence law. A payment product designed for East Africa will therefore meet one rulebook in Nairobi, another in Johannesburg and a third in Lagos. Because Kenya's mobile-money experience has made it a reference point for policymakers, the way its regulators handle the first licences will be watched well beyond its borders.

There is a deeper question underneath. Much of the global stablecoin market is denominated in dollars. For economies whose currencies have been volatile, dollar-linked tokens offer households and businesses a store of value outside the domestic banking system. That is precisely why central banks are cautious. Kenya's power to restrict foreign stablecoins is, in effect, a statement about monetary sovereignty in a digital age. How African regulators balance access to stable digital money against control of their own monetary systems may become one of the defining policy questions of the next decade.

Evidence note. Strong. The regulations are gazetted law, and several law firms have published consistent analyses.

The counter-view. Licensing requirements may raise the cost of entry enough to consolidate the market among larger players, and strict controls on foreign stablecoins could push activity into unregulated peer-to-peer channels.

Questions for further research

• Which African economies are gaining a foothold in electronics or digital-service supply chains, and what enabled it?
• How many firms apply for Kenyan VASP licences before 4 November, and how many receive them?
• Could licensed African stablecoins lower the cost of remittances into the region?

The first rung narrows.

June identified the entry-level challenge as a risk. The third quarter produced the most detailed evidence yet that it is already happening, concentrated where AI replaces tasks rather than supporting workers.

The evidence: canaries in the coal mine

In August 2026, economists at the Stanford Digital Economy Lab published an update to “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence”. Using anonymised ADP payroll records covering millions of US workers through June 2026, Erik Brynjolfsson, Bharat Chandar and Ruyu Chen document six facts:

1. No economy-wide displacement. Overall employment in AI-exposed occupations has not fallen.
2. A large gap at the entry level. Employment of 22–25-year-olds in the most AI-exposed occupations is 19% below where it would be had it kept pace with less-exposed peers. Experienced workers show no comparable gap.
3. The gap is widening. On the same measure, it was 15% at the July 2025 data point.
4. Hiring, not firing. The divergence comes from fewer young people being hired, not from layoffs.
5. Substitution, not complementarity. Declines concentrate where AI mainly automates tasks; where AI mainly supports workers, employment is flat or rising.
6. Jobs, not pay. The adjustment is visible in employment, with little change in base pay.

In absolute terms, employment of 22–25-year-olds in the two most exposed groups of occupations fell about 11% between November 2022 and June 2026, while employment of the same age group in less-exposed occupations grew about 10%.

The mechanism

AI appears to substitute most effectively for codified knowledge: formal, documented knowledge of the kind taught in school. It complements tacit knowledge: the judgement and practical skill built through experience, mentoring and exposure to real situations. Junior roles have traditionally been where people convert the first into the second. As one analysis put it, a company that stops hiring 23-year-olds this year will have no 33-year-olds with a decade of AI-era judgement in 2036 (Alex Goryachev).

Why this matters: an experience gap that compounds over a decade

The consequences of reduced junior hiring do not appear immediately. In the first year, organisations may see only efficiency. The cost arrives later, and it compounds. Five years on, there are fewer mid-level professionals with hands-on experience. Ten years on, there are fewer candidates for senior roles that require judgement, accountability and the ability to supervise AI systems. The capabilities that AI currently complements are precisely the ones a thinner pipeline fails to produce.

For Africa the question takes a different shape. The continent has the world's youngest population, and many economies depend on creating large numbers of first jobs each year. The Stanford evidence is from the United States; no comparable payroll study yet exists for Kenya, Nigeria, South Africa or elsewhere on the continent. But the mechanism would matter even more in labour markets where the entry-level rung is already hard to reach. If outsourced services such as customer support, basic software work and data processing, which have provided many first jobs in Nairobi, Lagos, Cape Town and Kigali, are among the tasks most exposed to automation, the challenge is not only corporate. It is a question of national workforce strategy.

How strong is the evidence?

The authors describe the findings as descriptive, not causal. The gap is larger in the ADP sample than in national survey data, shrinks when controlling for education, and shows some divergence that predates generative AI. Researchers at the Federal Reserve Bank of New York found it difficult to attribute the entry-level slowdown to AI alone. Against this, the gap kept widening after interest rates stabilised, and it loads specifically on automation-type AI use with a clear age gradient. A note on reporting: much coverage cites a rise “from 13%”; that was an earlier regression estimate. The like-for-like comparison is 15% to 19%.

Governance: the EU buys time

On 7 May 2026, EU legislators provisionally agreed the AI Digital Omnibus, and the European Parliament adopted it on 16 June (Sidley). Obligations for stand-alone high-risk systems, including AI used in recruitment, credit scoring and education, move from 2 August 2026 to 2 December 2027; high-risk AI embedded in regulated products moves to 2 August 2028 (Hogan Lovells; Mondaq). The uses of AI most likely to shape who gets hired now have 16 more months before EU obligations apply. Organisations can use that time to build proper oversight, or to defer it.

Evidence note. Moderate to strong. A large administrative dataset and transparent methods, but US-only and explicitly non-causal.

The counter-view. Entry-level hiring may be weak for reasons that coincide with AI: post-pandemic normalisation, remote work and a “low-hire, low-fire” labour market. The gap could narrow if hiring conditions improve generally.

Questions for leadership teams

• Where have we quietly reduced junior hiring, and who will be our experienced staff in ten years?
• Which of our AI uses substitute for people's tasks, and which make people more effective?
• How do new staff now build tacit knowledge: apprenticeships, rotations, supervised review of AI work?
• Are we using regulatory delays to build oversight, or to postpone it?

Social & institutional change: the retreat of public development finance

While private capital concentrated around computation, official development assistance fell for a third consecutive year, and the heaviest cuts landed on health programmes and on the poorest regions.

Net official development assistance fell 23.3% in 2025, the largest annual drop on record, and is projected to fall a further 6.9% in 2026, its lowest level since 2014 (OECD). This is only the second time on record that aid has fallen three years running; the first was 1992–1995. The United States, Germany, the United Kingdom, Japan and France accounted for almost all of the 2025 decline, and US aid alone fell by 56.9% (African Business).

Bilateral aid to sub-Saharan Africa fell 26.3% in 2025 and is projected to fall another 11.6% in 2026, its lowest level since the early 2000s. Core funding to UN organisations is expected to fall about 31% between 2024 and 2026 (Third World Network). Programmes against malaria, tuberculosis and other infectious diseases, and for reproductive health, face projected reductions of 40–60% over two years. The cuts are felt across the whole continent: thirty-two of the 49 countries in sub-Saharan Africa are least developed countries, the group facing the steepest reductions (The EastAfrican).

Why this matters: the two faces of the human-capacity thesis

June's closing line argued that progress remains a human undertaking. This quarter shows both faces of that claim. Capital for machines is abundant. Capital for human health and development in the poorest regions is shrinking fast. For non-profits, research institutions and public bodies, three consequences follow:

1. Concentration risk becomes existential. Organisations funded mainly by one or two bilateral donors face the greatest exposure.
2. The evidence bar rises. Funders will ask harder questions about cost per outcome and sustainability.
3. New capital sources become strategy. Domestic philanthropy, diaspora giving, blended finance, earned revenue and private-sector partnerships move from supplements to core strategy.

Evidence note. Strong. OECD figures are official, though 2026 values are projections.

The counter-view. Some argue that lower aid could accelerate domestic resource mobilisation and local ownership. Evidence on whether that happens fast enough to protect health outcomes is, so far, thin.