Based on my KSCCN monograph The Sovereign Capability Compact: Re-Engineering Official Development Assistance for Research, Enterprise, and Sovereign AI in the Global South (DOI 10.5281/zenodo.24048235).
An accounting system built for a different century
Why do so few developing countries manage to finance their own research, innovation and enterprise through development assistance? My new KSCCN monograph argues that this is not an accident. It follows from the statistical architecture of aid itself.
When the OECD Development Assistance Committee formalised the definition of Official Development Assistance in 1972, it built an accounting system designed to track welfare transfers and donor effort, not the growth of technological capability in the countries receiving it. The rules on what counts as aid are still negotiated among donors. Recipients' industrial upgrading is not what those rules were written to measure.
Where the money actually goes
In 2023, total DAC aid reached USD 288 billion, about 0.37 per cent of donors' combined national income, barely half of the 0.7 per cent target. Aid actually reaching developing countries had already fallen for three straight years before the record contraction of 2025, with a further decline projected for 2026.
Three mechanisms then drain what remains before it can reach domestic laboratories and firms:
- Aid that never crosses the border. Country Programmable Aid, the share partner countries can actually programme, has hovered around half of bilateral aid. In-donor refugee costs alone came to about USD 31 billion across DAC members in 2023. When half of reported aid never arrives as programmable capital, ministries of science, ICT and industry cannot align it with multi-year research strategies.
- Research spent in the North. Scholarships in donor countries and funding for donor-country research institutes absorb a meaningful slice of bilateral aid, yet most of it is spent inside the donor economy. Scholarships often accelerate the permanent emigration of the strongest STEM and AI researchers, while Northern consortia keep the intellectual property, the model weights and the high-impact publications.
- The tying loophole. Formal untying compliance looks high, but only within the Recommendation's narrow scope. Frontier equipment, cloud compute, IP advisory and technical cooperation largely fall outside it. In one member's 2023 profile, only 24.2 per cent of out-of-scope aid was untied. The highest-value technology contracts can stay with the donor's own firms.
Why sovereign AI has to be built as a whole
As public administration and industry reorganise around AI, dependency moves from commodities to data, compute and algorithmic standards. Roughly two-thirds of developing countries, and 88 per cent of Least Developed Countries, had no national AI strategy by the end of 2023, according to UNCTAD.
In the monograph I model a country's sovereign AI capability across four pillars: Data, Compute, Models and Governance, including talent. The pillars are strong complements. Open-weight models cannot be adapted without curated local-language data; national compute cannot run safely without domestic cryptographers, safety engineers and liability law. Funding one pillar in isolation buys very little. The rational budget rule is to balance marginal returns across all four.
Sovereignty here does not mean autarky. It means choosing the degree of openness at which the marginal benefit of global scientific collaboration equals the marginal cost of dependency: exposure to jurisdictional coercion, unilateral service termination and offshore surveillance. Aid tied to foreign cloud subscriptions pushes a country far past that point.
The Sovereign Capability Compact
The monograph proposes a 60-month compact between recipient governments and development partners:
- A statistical audit and ring-fence. Classify all incoming aid across the eight DAC modalities, require that at least 85 per cent of research and technical-assistance funds are spent in domestic universities and laboratories, and end parallel project units and tied technical assistance.
- Patient risk capital, not debt. Capitalise Sovereign Research and Enterprise Co-Investment Facilities that use the grant-equivalent rules for private sector instruments to co-invest in renewable-powered compute, open-weight domain models, in-country advisory systems and indigenous deep-tech firms, backed by regulatory sandboxes and innovation-oriented public procurement.
- Endogenous autonomy. Revolve royalties and equity exits into permanent university and sovereign AI endowments, scale South–South and triangular cooperation on open models and post-quantum standards, and phase out external concessional borrowing for core development functions.
Measuring what matters
Output counts tell us little. The compact proposes four co-designed indicators: the annual decay rate of external dependency; the retention of research spending and intellectual property in-country; a sovereign AI maturity score covering compute, local-language data, public-sector adoption and algorithmic audit capacity; and the foreign-exchange earnings of co-invested enterprises measured against annual external debt service.
Shrinking aid envelopes can either deepen dependency or become the foundations of technological sovereignty. The difference lies in rules that are, in the end, a matter of design.
Read the full monograph, with its tables, formal models and phased implementation plan:
Zenodo (DOI): https://doi.org/10.5281/zenodo.24048235
Academia: https://www.academia.edu/180476751
KSCCN