Brazil buys one Huawei machine and one Nvidia machine, and says the split is the point

Brazil is spending 2.3bn reais, about $444m, on two AI supercomputers: one built with Huawei, the other expected to go to Nvidia ^[1][2]. The unusual part is not the hedge but the disclosure. "The strategy is not to depend on a single company, technology or country," the Lula government said in setting out the plan ^[3].

The Chinese half is the bigger one. Huawei Technologies and the speech-recognition firm iFlytek will build a 1.3bn reais system in Rio de Janeiro, roughly $251m, intended for training large language models, with work scheduled to start in July 2027 ^[4]. That is about 57 percent of the total budget ^[1], and roughly 30 percent more money than the American machine ^[2]. The second system goes to a tender that unnamed officials and the science and technology minister, Luciana Santos, expect Nvidia to win, worth about 1bn reais or $193m and due to be operating by the end of 2027 ^[5]. On those dates, the Nvidia machine is meant to be running about six months after the Huawei build begins ^[3]. Neither system exists; the tender has not been formally awarded ^[19].

Siting tells you what the binding constraint is. The Nvidia machine goes to Rio Grande do Norte in the northeast, chosen for its wind and solar capacity rather than proximity to users ^[6]. Officials say it would rank among the most powerful AI machines in the world, a claim published without a specification ^[7]. Read that as an ambition.

The duplication is real and priced in. Two architectures mean two software stacks, two sets of trained engineers, and no ability to pool the machines ^[9]. Most national programmes work hard to avoid exactly that. What Brasilia gets in exchange is exposure to two policy regimes instead of one: American export controls have repeatedly changed which Nvidia parts can be sold where, while the Chinese domestic alternative has been advancing on custom silicon rather than GPUs ^[10]. Money comes from the National Fund for Scientific and Technological Development in phases rather than a lump sum, which leaves the option to slow either project without killing it ^[8]. That is the hedge underneath the hedge.

The open questions are the interesting ones. Brazil has not said whether US export rules apply to the Nvidia hardware it intends to buy, whether Washington has been consulted, or whether the Huawei system runs into any American restriction on the company ^[11]. Jensen Huang has argued publicly that Chinese models running on Chinese chips would be a bad outcome for the United States ^[12]; Brazil has now arranged for both stacks to sit under one national programme.

Scale check: Britain put £225m into a single machine, Denmark bought an Nvidia system outright, and the EU has a €30bn gigafactory programme ^[13]. Against that, $444m is a first move, though it is the largest sovereign AI compute commitment in Latin America, a region that has hosted private data centre investment for years without owning much of it ^[14][15]. The word doing the heavy lifting is sovereignty, and it is narrower than it sounds: Brazil manufactures neither machine, so what it controls is the data and the models, not the silicon ^[16]. Training a Portuguese-language model on Brazilian data inside Brazilian borders is still materially different from renting capacity in Virginia ^[17].

Watch for the access rules. Funding of this kind normally goes to universities and state institutes rather than companies, and Brasilia has published no allocation policy for either machine ^[18].