Open weights do not come with a cluster

A seller on the fourth floor of a components mall in Huaqiangbei poured me tea beside a graphics card that had been opened up, rebuilt and closed again. An RTX 4090 carrying 48GB of memory, a capacity Nvidia has never sold on that board, a little over 20,000 yuan if I took it that afternoon.

Nobody in that building is training a frontier model. The geography is the interesting part. The memory, the shop that reflowed it, the two-person team buying it to run fine-tunes for a logistics customer and the man selling it all sit inside about forty minutes of each other.

America's edge in artificial intelligence was never that it produced the most talent. It was that it was the place talent had to go. Those are different assets and they fail in different ways. The first erodes slowly, as other countries get good. The second collapses the moment staying home stops costing you your career.

Eleven percent stay home

MacroPolo's Global AI Talent Tracker is the least sentimental document in this argument. Version 3.0 takes 4,622 researchers who published at NeurIPS 2024, ICML 2024 and ICLR 2025, and traces each one back to where they did their first degree. China accounts for 38 percent of them. The United States, 24. India, 10.

Then it asks where those people work now. Fifty-nine percent are at American institutions, three points up on the previous edition. Of everyone who did their undergraduate degree in China, 72 percent now work in the United States and 11 percent work in China.

Put the two halves together and the American position stops looking like a factory and starts looking like a gravity well. The country produces roughly a quarter of the world's elite AI researchers and employs three-fifths of them. Every other country in that table is running a training programme for someone else's payroll.

The 2026 AI Index put a figure on the well itself. The number of AI scholars moving to the United States has fallen 89 percent since 2017, with an 80 percent fall in the past year alone. That is the shift. It did not happen on a benchmark.

Every other country in that table is running a training programme for someone else's payroll.

The sentence in the Waves interview

In July 2024, months before most people outside China had heard of the company, Liang Wenfeng sat down with the Chinese outlet 暗涌 Waves and was asked who had actually built DeepSeek-V2. His answer was that the team contained no returnees from overseas, that they were all local, and that they were mostly fresh graduates of good Chinese universities together with PhD students in their fourth or fifth year.

Asked about the world's top tier he said: "The top 50 experts might not be in China." Then he said his company could probably train such people itself. I read that in the original when it came out and filed it under bravado.

In December the DeepSeek-V3 technical report arrived with the receipts. A cluster of 2,048 H800 GPUs. 2.788 million GPU hours for the full training. At an assumed rental of two dollars per GPU hour, 5.576 million dollars for the final run. The report is honest that this covers the final run only, not the failed ones, not the architecture research and not the data work, which is the caveat almost everyone quoting the figure left out.

The cost was never the signal. The signal was that a complete loop closed inside one country. Undergraduate degree in Hangzhou or Beijing, cluster within domestic reach, salary competitive enough to matter, senior people in the same building, users down the road. Not one step in that sequence required a consular appointment.

0.83 percent

The 2026 AI Index gives sub-Saharan Africa 0.83 percent of global AI publications across 2013 to 2024. North Africa, counted alongside the Middle East, gets 4.27 percent. The continent holds about a fifth of the world's people.

Compute is the worse number. Researchers at the Oxford Internet Institute have a phrase for most of the world map: compute deserts, places where there is simply no accelerated capacity to rent at any price.

Africa's own industry body is blunter than any foreign study. The Africa Data Centres Association's 2026 economic report puts the continent at 0.6 percent of global data centre capacity, 360 megawatts live against roughly 55 gigawatts worldwide, and projects that Africa will hold that share rather than grow it, because hyperscale building is accelerating faster everywhere else.

None of that is a talent problem.

What the download does not include

Open weights were the right thing to be optimistic about and I have written as much. A checkpoint you can pull, inspect and fine-tune removes a genuine barrier. It removes exactly one.

It does not come with a cluster. Udutech's Africa GPU Hub in Lagos rents GPU time from under a dollar an hour, which is real progress and a different category of object from 2,048 interconnected cards held for eight straight weeks. Cassava, the first Nvidia cloud partner on the continent, is standing up AI factories beginning in South Africa with Nigeria, Kenya, Egypt and Morocco behind it. Nigeria today runs on roughly 86 megawatts of installed data centre capacity across 25 facilities, against South Africa's 61 and Kenya's 19, with self-generated power costing 28 to 33 US cents per kilowatt hour.

Nor does it come with capital of the right shape. Partech counted 4.1 billion dollars into African tech in 2025, up 25 percent on the year, which reads well until you open the split. Debt was 1.6 billion of it, 41 percent of everything deployed, up from 17 percent in 2019. Debt is priced against revenue you already have. Nobody underwrites a two-year research bet with a repayment schedule.

The seniors are not in the download either, and they are the part no ministry can procure. A model card will not tell you which ablation to run first. A paper will not tell you what its authors decided not to publish, or which of the three approaches they abandoned in month four and why.

And there is no customer attached. In Yaba I sat with an engineer who had fine-tuned a small model for Hausa-language intake forms, on rented hours, paid for on a card that was declined twice before it cleared. The thing worked. He had no buyer, no second card and nobody senior enough to tell him which of those two problems to solve first. He is in Berlin now.

The strongest version of the argument against me

The honest counter is that the map changed and I am reading an old one. Remote work means a senior engineer in Lekki can sit on a San Francisco payroll without boarding anything. Open weights mean the frontier arrives by git pull rather than by NDA. Diaspora networks move money and knowledge home at a scale no fund on the continent comes close to matching, and they do it every month, without a term sheet.

All of that is true and some of it carries real weight. Retention no longer requires that the work be local, only that the worker is.

Here is where it stops. Remittances are consumption transfers. They pay school fees and finish houses in Enugu, and they do not capitalise a cluster, because nobody's aunt is wiring nine figures. Remote employment keeps the body and exports the decision: the engineer stays in Lagos while the architecture choice, the training budget and the intellectual property all sit in another jurisdiction. That is a services outcome. Services outcomes are fine right up to the moment the buyer automates the service.

The open weights point is the one I concede furthest, because it is mine. A downloadable model genuinely changes what a small team in Lagos or Kigali can put in front of users. Deployment is not capability. You do not fine-tune your way into knowing what to do when the loss curve flattens at hour thirty of a run you cannot afford to restart, and you will not learn it from a README.

China opened a visa in October 2025

Last October China added a new category to its ordinary visa system, the K visa, aimed at young foreign science and technology graduates and unusual in not requiring an employer to sponsor the applicant.

That is the tell. A country that had solved this with retention alone would not have bothered. Beijing's read is that keeping your own is necessary and not sufficient, and that the other half of the trade is becoming somewhere other people's graduates actively want to enter. The K visa is an attempt to buy the American asset at the exact moment the Americans stopped maintaining it, and the argument it started inside China shows the price is not only fiscal.

Nigeria has no equivalent instrument and, for now, nothing to put behind one. That is a sequencing problem rather than a moral one, and sequencing problems get solved with power, land, cards and patient money, roughly in that order.

The word doing the work

I preside over the Nigerians in Shenzhen Association. Read one way, our membership roll is a directory of engineers, traders and postgraduates Nigeria educated and does not have. Very few of them are coming back on sentiment. Two told me last year, in almost identical words, that they would take a pay cut tomorrow if there were a cluster to come back to.

I could not give either of them a date.

Back in the mall, the seller asked where I was from, and when I said Nigeria he asked whether we bought many cards. I told him not yet. He nodded and turned back to his tea. The whole argument lives inside that word, and nothing in the five-year strategy tells me who is supposed to make it true.

Sources

MacroPolo, The Global AI Talent Tracker 3.0 : https://archivemacropolo.org/interactive/digital-projects/the-global-ai-talent-tracker

ChinaTalk, translation of Liang Wenfeng's interview with 暗涌 Waves : https://www.chinatalk.media/p/deepseek-ceo-interview-with-chinas

DeepSeek-V3 Technical Report (arXiv:2412.19437) : https://arxiv.org/pdf/2412.19437

Stanford HAI, Inside the AI Index: 12 Takeaways from the 2026 Report : https://hai.stanford.edu/news/inside-the-ai-index-12-takeaways-from-the-2026-report

Techweez on the 2026 AI Index: sub-Saharan Africa's share of global AI publications : https://techweez.com/2026/04/21/ai-publications-from-africa/

ITWeb on the Africa Data Centres Association 2026 Economic Report : https://www.itweb.co.za/article/africas-data-centre-capacity-on-back-foot-despite-investment-push/dgp45MaBbknqX9l8

Techpoint Africa, AI infrastructure in Nigeria: closing the GPU gap : https://techpoint.africa/guide/ai-infrastructure-in-nigeria-gpu-gap/

allAfrica on Partech's Africa Tech Venture Capital Report 2025 : https://allafrica.com/stories/202601250064.html