Humain built Saudi Arabia's new Arabic model on a Chinese startup's open-source system

Two purchases sit in one budget line and run on different clocks. Contracting for computing capacity means a build schedule and a rack layout designed around one vendor's networking. Choosing a base model means cloning a repository, fine-tuning it and putting it behind your own API. Humain did both, from suppliers on opposite sides of an export-control regime ^[2][3].

Brazil made the same division in cash. Of the roughly $444 million announced in August, about $255 million goes to supercomputing in Rio de Janeiro with Huawei and iFlytek. About $189 million goes to a separate machine Brazil is tendering and expects to source from Nvidia ^[4][5][6]. The Chinese-partnered side takes 57 percent of the money, the American side 43 ^[7]. "The strategy is not to depend on a single company, technology or country," President Luiz Inacio Lula da Silva's administration said in a statement. It added that the investments were intended to strengthen Brazil's sovereignty over its data ^[8][9].

The buyers doing this are governments that can pay for chips and cannot pay to train a frontier model. Jesse Marks, a China-Middle East research scholar at the Australian National University, told Rest of World that competing at the frontier has become prohibitively expensive. That leaves governments to decide which technologies they can develop themselves and where they must rely on foreign suppliers ^[21]. Marks said Washington holds greater leverage where American technology is harder to replace, including advanced semiconductors and networking equipment. He said it will struggle with countries that want AI but cannot afford American systems or software ^[10][11]. "Outside of the network of American partners who are trusted or can afford it, it's Chinese AI. It's open models," he said ^[12].

Kyle Chan, a fellow at the Brookings Institution's John L. Thornton China Center, told Rest of World that countries "want to keep their options open and not be too dependent on any single source of technology." Many of them, he said, do not attach the same security stigma to Chinese technology that Washington does ^[14][15].

The hardware side is contested too. Huawei has bid to build AI data centres for the Egyptian government against a proposal involving Nvidia, according to a source-based Bloomberg report cited by Rest of World ^[16]. Bloomberg did not report the cost gap at the model layer, and Marks framed the Saudi case conditionally, calling Chinese foundations a "natural decision" if they perform better on relevant benchmarks ^[22][17].

Washington has accused Chinese labs of "free-riding" and training their models from frontier American models at an "industrial scale" ^[18]. China rejected the accusations and warned it would retaliate if Washington curbs Chinese AI companies ^[19].

For a team picking suppliers for next year, the useful ranking is what replacing each one would cost eighteen months from now. Hardware sits high on that list: a signed build, a delivery slot, an export licence a policy change can suspend. Weights sit low: a fine-tune to redo and an eval suite to rerun. Humain buys its compute from Nvidia and its base model from MiniMax. Of the two, only the Nvidia relationship would take years to unwind ^[2][3].