China's African loan book is shrinking. Its models are free.

In April I rode from Ikeja to Yaba in a Toyota with something north of 380,000 kilometres on the clock, and the driver spent most of the crawl arguing with a free translation app on a Tecno handset. He wanted a Yoruba phrasing his customer had used and could not place. I asked what the app was. The listing on its page named a fine-tune of Qwen, Alibaba's open model family, wrapped by somebody in Nairobi and given away.

He had never heard of Alibaba. There is no reason he should have.

The road under the car was resurfaced by a Chinese contractor. The mast carrying his signal was, on the available numbers, most likely Huawei equipment. The handset was made by a company headquartered forty minutes from my flat in Shenzhen. And now the thing doing the thinking in his pocket had the same postcode as all of it. None of that arrived by treaty. Every piece of it arrived as the cheapest option available at the moment somebody needed one.

The loan book is smaller than the number in circulation

Start with the money, because the money is where most writing on this goes wrong.

The Boston University Global Development Policy Center maintains the Chinese Loans to Africa database, and its November 2025 update is the only figure I will use. Between 2000 and 2024, 42 Chinese lenders signed 1,319 loan commitments worth $180.87 billion with 49 African governments and seven regional institutions. That is a quarter of a century of lending, and it is a fraction of the totals that circulate on LinkedIn. In 2024 the entire continent drew just under $2.1 billion across six projects, with $1.45 billion of that going to Angola for transmission lines and roads.

Sovereign lending peaked years ago and has been contracting since. The Beijing summit in September 2024 pledged 360 billion yuan over three years, roughly $50 billion, and the composition says more than the headline does: credit lines and expected corporate investment carry most of it, not concessional state loans.

Nigeria's share is legible in concrete. The Abuja to Kaduna standard gauge line, 186.5 kilometres, cost around $874 million, of which $500 million came from China Eximbank as preferential buyer's credit. CCECC built it. Commercial service opened on 28 July 2016. Lagos to Ibadan followed in June 2021. I have ridden both, and the announcements are in English and Mandarin.

The lending is winding down. What the lending bought is not going anywhere.

Nobody in Beijing planned Luganda. That is the point.

Vendor lock-in is not a treaty

The most-quoted statistic in this whole argument is that Huawei built about 70 percent of Africa's 4G networks and roughly half of its 3G. It comes from an Atlantic Council piece published in April 2021, which in turn traces back to trade press and to Huawei's account of itself. I use it with the caveat attached, because it is five years old and nobody has audited it properly since. Order of magnitude, not measurement.

Power is harder to pin down and I will not pretend otherwise. The International Energy Agency reported a decade ago that Chinese companies had built around a third of new generation capacity added in sub-Saharan Africa over the preceding five years, weighted heavily towards hydro. I have not found a rigorous recent update, so treat it as direction rather than a current share.

Handsets are the firmest ground of the three. Transsion, headquartered in Shenzhen, sells under the Tecno, Infinix and itel brands, and took 48 percent of African smartphone shipments in 2025 on more than 40.5 million units. Devices priced under $200 were 81 percent of the market that year.

Now think about what a vendor actually leaves behind. Not a debt. An engineering staff trained on one management console. A spares chain. An interface the next procurement has to speak to. General Electric and the Bell system did exactly this across Latin America and East Asia for fifty years, and nobody needed to call it a trap, because it did not have to be one. It only had to arrive first and be cheap enough that evaluating alternatives felt like a waste of a budget cycle.

Weights are the cheapest export China has ever made

A submarine cable costs hundreds of millions of dollars and takes four years. A model costs a download.

The releases have stacked up fast. DeepSeek publishes its main weights under MIT, about as unconditional as software licensing gets. Alibaba puts most of Qwen under Apache 2.0. Z.ai released GLM-5 on 11 February 2026, a 744-billion-parameter mixture of experts with 40 billion active per token, MIT licensed and trained end to end on Huawei Ascend silicon using MindSpore rather than anything from Nvidia. Moonshot published the full Kimi K3 weights on 27 July 2026 at 2.8 trillion parameters, under a bespoke licence that permits self-hosting and only bites if you build a competing hosted service at real scale.

Alibaba says Qwen has passed three billion downloads and spawned more than 300,000 derivatives. That is a company describing itself, so take the independent count instead. Hugging Face's own August 2026 report on open models tallied about 2.045 billion Qwen downloads during 2026, against 418 million for Google's open models and 227 million for Meta's. Qwen is now the default starting point, and the gap is not narrow.

The most interesting consequence of that is not in any boardroom. Sunbird AI, a non-profit in Kampala, released Sunflower 14B and 32B in October 2025. The models handle 31 Ugandan languages, they beat ChatGPT and Gemini 2.5 Pro on translation for 24 of those 31, and they are built on Qwen 3. Coverage has since widened to 67 African languages.

Nobody in Beijing planned Luganda. That is the point. The base model was free, the licence permitted it and a team in Kampala did the rest. Which is precisely how a default becomes a default.

What this looks like from the Shenzhen end

I read the Chinese-language announcements before anyone translates them, and the gap between the two versions is instructive.

On 16 July 2026 officials from 29 countries signed the agreement establishing the World Artificial Intelligence Cooperation Organization, headquartered in Shanghai. English coverage framed it as a bid for governance influence, which it partly is. The Chinese material I read leaned much harder on application and on capacity building: 5,000 training places for people from developing countries over five years, plus cooperation centres with the African Union and other regional bodies. South Africa, Kenya, Ethiopia, Senegal, Algeria and Cameroon are among the signatories.

Training places are not glamorous. They are also how a standard propagates. I have watched a smaller version of this through the Nigerians in Shenzhen Association. The traders who came here fifteen years ago to buy phone components in Huaqiangbei now run freight and sourcing companies, and their children are reading computer science in Guangzhou on scholarships. When those graduates go home, the toolchain they know is the toolchain they deploy. That is not ideology. It is muscle memory.

There is a financial version of the same mechanism. In December 2024 Beijing and Abuja renewed a 15 billion yuan currency swap, about $2 billion, explicitly so bilateral trade could settle without touching dollars. Plumbing again. Nothing signed under flags. Just a marginally cheaper path that, taken often enough, stops being a choice.

The strongest version of the case against me

Three objections, and the first two are correct.

Infrastructure is not control. Deborah Brautigam and the China Africa Research Initiative went through thousands of Chinese loan contracts and more than 3,000 projects worldwide and found no evidence of deliberate over-lending in order to seize assets. The debt-trap story is a meme with a citation trail that loops back on itself. Anyone writing about this corridor who reaches for it has not read the documents.

African states have agency and the record shows them using it. Boston University's own 2024 work on Chinese-supported power projects found that host-state institutions, regulators, procurement rules and civil society materially shaped what got built and on what terms. Ethiopia renegotiated. Zambia restructured. Nigeria has cancelled and re-scoped Chinese rail segments more than once.

The third objection is the one I take most seriously. Open weights are genuinely the most sovereignty-preserving arrangement anyone is offering. An MIT-licensed file sitting on a server in Lagos cannot be rate-limited, price-raised, geo-blocked or switched off from abroad. It phones nobody. Compare that to an American frontier API, where your access is a commercial relationship subject to export policy and a terms-of-service revision you will read about afterwards. On the narrow question of who can revoke what, the Chinese open-weight option wins, and it is not close.

Nor is the picture monolithic. Nigeria's own state-backed model, N-ATLAS, launched at the UN General Assembly in September 2025 by NITDA's centre for AI and robotics with the Lagos startup Awarri, is a fine-tune of Llama-3 8B. American base, Nigerian data, custom licence with a thousand-user cap on free use. Nobody is captured yet.

Sovereignty you cannot afford to run

The trouble with the third objection is that it describes a right rather than a capability.

Self-hosting requires somewhere to host. Africa holds under one percent of global installed data centre capacity, against something in the region of 122 GW worldwide, and outside South Africa only about a third of what has been built is actually live and fitted out. Underneath that sits electricity. In April 2026 Nigeria's grid managed an average of 4,048 megawatt-hours per hour against 13,625 MW of installed generation, a plant availability factor of 31 percent. Roughly four gigawatts of delivered power, for more than 200 million people.

So the licence says you may run Kimi K3's 2.8 trillion parameters on your own hardware in Abuja. The grid says you will be renting it from somebody's cloud, and the nearest cheap one is increasingly Chinese. Huawei ran an AI data centre summit in Cairo in April 2026 and published a reference design for African AI facilities, then launched its data centre networking products in Johannesburg in July. That is a company selling the floor the sovereign model has to stand on.

The second trouble is that openness is a decision, and decisions get revisited. On 15 August 2026 Z.ai shipped GLM-5.3 through its API and held the weights back. Its stated reason was that the model scored 84.5 percent on the CyberGym benchmark, ahead of what Z.ai measured for Anthropic's Mythos 5, and that once weights are public a lab can no longer control what anyone does with them. Their phrasing, roughly: cyber capability developed faster than expected. It is the first time the company has ever delayed a GLM weight release.

The placeholder page

Z.ai said the weights would follow in about two weeks, around 28 August. That is today. As of yesterday the Hugging Face repository for GLM-5.3 was still a placeholder listing that date and nothing else.

Perhaps the file lands this afternoon and this whole worry ages badly. It probably will land. But the shape is visible now: the most sovereignty-preserving AI option available to an African government is sovereignty that a private Chinese lab, sitting under a Chinese regulator, chooses to keep granting, release by release, on a calendar it sets and for reasons it explains after the fact.

That is not a trap. Nobody set it. It is what happens when the cheapest thing is also the only thing that shows up, for long enough that you stop budgeting for an alternative, and then one morning it does not show up.

I will check the repository again tonight.

Sources

Boston University Global Development Policy Center, Chinese Loans to Africa Database 2000-2024 (November 2025) : https://www.bu.edu/gdp/2025/11/13/selective-engagement-and-strategic-retooling-chinese-loans-to-africa-database-2000-2024/

AidData, China Eximbank $500m preferential buyer's credit for the Abuja-Kaduna railway : https://china.aiddata.org/projects/195/

Atlantic Council, The digital infrastructure imperative in African markets (April 2021) : https://www.atlanticcouncil.org/blogs/africasource/the-digital-infrastructure-imperative-in-african-markets/

Fortune, Alibaba's Qwen models hit 3 billion downloads, passing Meta and Google (15 August 2026) : https://fortune.com/2026/08/15/alibaba-qwen-open-ai-models-3-billion-downloads-meta-google/

Hugging Face blog, GLM-5: China's first public AI company ships a frontier model : https://huggingface.co/blog/mlabonne/glm-5

Implicator.ai, Z.ai delays GLM-5.3 weights after CyberGym score tops Mythos 5 : https://www.implicator.ai/z-ai-delays-glm-5-3-weights-two-weeks-after-cyber-score-beats-mythos-5/

Sunflower: A New Approach To Expanding Coverage of African Languages in Large Language Models (arXiv, Sunbird AI) : https://arxiv.org/abs/2510.07203

NERC data via Leadership, Nigeria grid instability as GenCos operate at 31% capacity in April 2026 : https://leadership.ng/grid-instability-persists-as-gencos-operate-at-31-capacity-in-april/