In the contest between US and Chinese artificial intelligence stacks in several markets, especially in Africa, state capacity and infrastructure deals will prove more decisive than model quality.
The contest will hinge on chips, data centers, energy, cloud regions, and the financing that binds them, together with the capacity of recipient states to regulate, procure, and absorb what is on offer. Topping the model leaderboards may count for less than the ability to finance, power, and permit the infrastructure stack underneath the models. Washington and Beijing have each built their strategy on that premise, and the African evidence so far bears it out.
According to one Atlantic Council analyst, a battle of opposing AI stacks is escalating, as national policymakers reach for control over critical digital infrastructure such as compute, cloud, chips, and regulation. Another Atlantic Council analyst expects Washington to sign more AI partnerships like the ones it established with Saudi Arabia and the United Arab Emirates, while working to counter Chinese influence in emerging markets, where Beijing’s growing strength in open-weight models could prove a winning formula for market share.
By late 2025, Qwen, the family of open-weight models developed by China’s Alibaba, had overtaken Meta’s Llama as the most downloaded model family on Hugging Face, the world’s largest repository for open-source AI models, while Chinese models accounted for more than 45 percent of downloads among leading open models. The significance is less that Chinese models are universally better than American alternatives, and more that they are increasingly becoming the default foundations on which developers around the world are choosing to build.
What Washington is doing
The US government has already organized policy around its own assumptions. Its American AI Exports Program encourages industry consortia to offer integrated AI stacks, backed by licensing support, financing, and advocacy.
The government’s call for proposals this spring reiterated the full stack as five mandatory layers, from AI-optimized hardware such as cloud, and network infrastructure to applications to specific sectors such as healthcare and agriculture. The call for proposals barred country-of-concern firms from any layer of a designated consortium. The Department of Commerce told applicants that a proposal presents a stronger national interest case where American engagement can tip a procuring entity choosing between American technology and technology from a country of concern. Naturally, analysts describe the program as aimed at emerging markets where American and Chinese firms compete for influence.
The design of the government’s call is also instructive: a package is presumed to satisfy the content test if American value accounts for at least 51 percent of the hardware layer. A model’s intellectual property must be majority American-owned, free of country-of-concern control, and where a package relies on open-weight models, an American firm has to supply the integration, fine-tuning, security, and support that make them usable. Nothing in the designation criteria measures how good the models are, and none of this export machinery would exist if Washington believed model quality alone won markets.
But Beijing acted first
Beijing assembled its own apparatus earlier. China’s Digital Silk Road has paired state-backed finance with exports of the Chinese stack to governments across Asia, Africa, and the Middle East. Huawei Cloud has marketed integrated cloud-and-AI packages in Saudi Arabia, the UAE, and North Africa, and has deeply integrated itself into Thailand’s digital infrastructure, with similarly substantial efforts in Malaysia and Indonesia. The terrestrial layer is already dominant across much of Africa: Chinese-built network equipment, affordable handsets through firms such as Transsion, and digital-payments infrastructure form a stack into which Chinese AI services arrive as an increment. At the model layer, Chinese platforms such as DeepSeek R1 and Qwen compete on cost and open weights instead of peak capability, a formula that could be well suited for capturing emerging-market share. The competitive logic on both sides is the same—finance, diplomacy, infrastructure, and distribution first, and models riding on top.
This pattern has appeared in Africa before: Huawei succeeded in displacing established competitors like Nokia and Ericsson to become a leading supplier of telecommunications infrastructure in the continent, using a model driven by vendor finance, integrated infrastructure packages, and rapid deployment rather than overwhelming technical superiority. The emerging contest over AI stacks increasingly resembles that earlier competition.
African outcomes turn on state capacity
Take Kenya. Microsoft’s planned one-billion-dollar data center with G42, the Emirati national champion behind the UAE’s technology partnership with Washington, stalled after the Kenyan government balked at the long-term fiscal guarantees and energy infrastructure commitments needed to anchor the project. The technology and capital were available, but the constraint lay in the state’s absorptive ability, or willingness to provide the conditions required for deployment. Once a project is blocked at the infrastructure layer, improvements in model quality become largely immaterial.
Or consider South Africa. One of the continent’s most advanced economies has kept Starlink out for years over local-ownership rules. And it withdrew its own draft national AI policy in April after the draft was found to have contained fabricated citations, which some suspect were generated by AI tools. Regulatory design decides which providers may enter; regulatory capacity decides whether the state can even articulate a strategy of its own.
Across the continent, the sovereignty agenda is, in practice, an exercise in state capacity. Nigeria, Kenya, and other African nations increasingly frame AI and other critical technologies through the language of sovereignty, resilience, and strategic autonomy, seeking greater control over data, infrastructure, and digital capability, to avoid long-term dependence on foreign technology ecosystems. Experts advising those governments report a push for data sovereignty, local talent, and better terms from foreign providers. The sixty-billion-dollar Africa AI Fund announced at the 2025 Global AI Summit on Africa targets infrastructure, talent, and startup development ahead of frontier model creation.
The same emphasis is visible in the private sector. The NVIDIA-Cassava partnership, a reported $720-million program, is rolling out NVIDIA-powered AI infrastructure across South Africa, Nigeria, Kenya, Egypt, and Morocco, beginning with three thousand graphics processing units at a South African facility and twelve thousand planned across the remaining markets over the following three to four years. It was possible because Cassava, founded by Zimbabwean entrepreneur Strive Masiyiwa, already owned the scarce assets: a fiber network stretching from Cape Town to Cairo, data centers, and a renewable-energy business.
Emerging-market data centers will succeed where there is energy flexibility, coordinated permitting, and workforce pipelines, with structural risk concentrated in grid redundancy, pricing volatility, and uneven regulation. Each of those ultimately reflects state capacity. African ministers’ explicit interest in efficient models suited to local energy and connectivity constraints reinforces the same conclusion, that the decisive determinants of AI adoption will be infrastructure conditions, absorptive capacity, and governance arrangements rather than incremental gains in model performance.
Where the American AI stack does land in Africa, it does so through deals whose outcomes are determined more by power, control, permitting, connectivity, and state capacity, increasingly accompanied by partnerships to build public-sector AI capability.
The decisive layer
At the point of adoption, model quality is increasingly fungible. For many public-sector and enterprise use cases in emerging markets, American and Chinese stacks are good enough, and the advantage of China’s open-weight, low-cost offerings can probably make up for whatever gap remains between the two options. Model performance has converged sufficiently that infrastructure and deployment constraints matter more.
But the bottom of the stack is not so easily swappable; chips clear export licensing, or they do not. Data centers get financed, powered, and permitted, or they stall. Cloud regions anchor to long-term procurement, or they collapse, as Kenya showed. Market entry is granted or refused by regulation, as South Africa demonstrated. And dependency, once installed, is sticky, because the switching costs sit in the infrastructure layer more so than software. Two superpowers have organized their statecraft around this. In the cases examined here, infrastructure, procurement capacity, and regulation have proven more consequential than differences in model performance.
The open question is whether the American AI Exports Program’s consortium-and-financing model can win in states where the true bottleneck is procurement capacity, energy systems, and regulatory institutions, and what investment in those institutions, ahead of investment in the stack itself, would do to tilt the field.
There is a caveat, though, because if model capability diverges sharply again at the frontier, through agentic systems with large productivity differentials. There, the assumption that models are largely interchangeable weakens. In that world, access to the most capable models could matter more than avoiding lock-in to any single stack. For now, the evidence points the other way, but that holds only under current conditions, and may not hold forever.
Amaka Yvonne Onyemenam is an advisor and technology thematic lead at Africa Practice, where she advises on strategy, risk, policy and engagement. She is a co-author of the African Union Startup Model Law and Policy Framework. She has advised development institutions and multinationals including Netflix, Google, Meta, Mastercard, and inDrive.
Further reading
Tue, Jul 14, 2026
When the plug gets pulled: Frontier AI, export controls, and what Africa must do
AfricaSource By Yasmine Abdillahi
A foreign government could sever access to frontier AI capability overnight, putting African institutions at risk.
Fri, Jun 26, 2026
The ‘MCC Effect’ is real. The Trump administration must not overlook it.
AfricaSource By Richard K. Bell
Millennium Challenge Corporation and its scorecards-based model incentivizes countries to make reforms that are in the US interest, a former US ambassador argues.
Fri, Jun 12, 2026
Somalia is the dark horse in US defense and economic strategy
AfricaSource By Maureen Farrell, Rose Keravuori
US policymakers have long viewed Somalia primarily through a counterterrorism lens. But as major maritime supply routes come under strain, its position along key sea lanes is reshaping its geopolitical relevance—and external actors like Turkey are already moving to capitalize.
Image: An AI (Artificial Intelligence) sign is seen at the World Artificial Intelligence Conference (WAIC) in Shanghai, China. REUTERS/Aly Song