Brazil pairs a Huawei-iFlytek project with a supercomputer tender officials expect Nvidia to win. The tender buys portability on paper; the silicon stays foreign.
Brazil's new sovereign-AI program does not pick a bloc. Announced on August 20 in Macaíba, Rio Grande do Norte, it funds a supercomputing and language-model project in Rio de Janeiro with Huawei and iFlytek and opens a competitive tender for a larger AI supercomputer that the science and technology minister, Luciana Santos, has said she expects Nvidia to win. "The strategy is not to depend on a single company, technology or country," the Lula administration told Reuters. Korea, by contrast, held its line on model provenance while relaxing a domestic-chip mandate; Britain is co-building its national systems with AMD. Brazil is running a Chinese partnership and a Western-facing tender at the same time.
The instinct to read that as two dependencies is fair, and the government appears to share it. The procurement document for the Nvidia-favored machine, published the same day by the National Laboratory for Scientific Computing (LNCC), contains a section headed, in so many words, "avoid lock-ins." What that tender can and cannot do about dependency is the more useful question, and the launch material does not answer it.
Three things, at three different levels of commitment.
The first is the Rio de Janeiro project: R$1.276 billion over five years, led by the national research network RNP with Huawei and iFlytek, to build compute for a national LLM and Portuguese-language models, with training for 3,000 Brazilian professionals and cooperation beginning in July 2027. Nothing published so far names the hardware. Ascend and its CANN software stack are the obvious expectation for a Huawei-built system, and SCN has covered Ascend as a viable second pole for frontier inference, but an expectation is not a bill of materials; on Huawei's own timeline, the training-class Ascend 950DT is not due until the fourth quarter of 2026.
The second is the tender. LNCC's Seleção Pública 27/2026 carries a value of R$959,040,959.04; the wider Rio Grande do Norte project is put at about R$1.06 billion, R$960 million for the equipment and R$100 million for preparation and operation. Proposals are due in person at LNCC in Petrópolis on October 8. Reuters reports that Santos anticipates Nvidia as the supplier; the tender itself names no vendor, and under its equivalence clause any component listed by name, accelerators included, may be replaced by an equivalent that matches or beats it on the required benchmarks.
Together those two tracks come to roughly R$2.3 billion, about $444 million, funded through the National Fund for Scientific and Technological Development in phases. Folha de S.Paulo put the wider package, which also covers a national cloud, an algorithmic-transparency center and a RISC-V chip program, at about R$2.5 billion.
The third thing is the longest-dated: a 10-to-15-year strategy to build domestic chip-design capacity on the open RISC-V architecture, with partners in Barcelona and at São Paulo's Instituto Eldorado. It is the only part of the package that points at Brazilian silicon, and it is not a near-term hardware plan.
The launch figures, carried by G1 from government sourcing, are 7,200 petaflops of FP16 performance, more than 50,000 processing cores, and an estimated 4 megawatts of power. None of those numbers appear in the tender.
What the Termo de Referência says is narrower and more demanding. Section 4.1 fixes the installation environment at a 3 MW limit for processing, storage and networking equipment, and requires every bidder to submit a power estimate proving the offered solution fits inside it. Accelerators must support FP16; FP4 and FP64 support is listed as desirable, not required. The megawatt between the launch estimate and the tender ceiling may be nothing more than a facility figure set against an IT-load figure. Neither document reconciles the two.
On performance, the tender does not score raw petaflops at all. Bidders must run MLPerf Inference v6.0 (Datacenter) on the official LLM workloads, Llama 2 70B, Llama 3.1 405B, Mixtral 8x7B and DeepSeek-R1, in both offline and server/interactive scenarios with accuracy verification, plus MLPerf Training v6.0 for fine-tuning and pretraining, under the full MLPerf rules. Energy measurement is mandatory, taken at the wall through the MLPerf Power workflow. The inference metric is tokens per second per watt; the training metric is the inverse of kilowatt-hours to target quality. The technical score weights inference and training equally, and the winner must demonstrate model-FLOPs utilization at acceptance and hand over containers so a witnessed subset of the tests can be re-run.
That is a better procurement instrument than the political headline. The "top 10" officials have cited, per Reuters, a claim about AI processing, and the government has not identified a published ranking or methodology to check it against. The tender buys measured throughput per watt on named models, reproduced at acceptance. An FP16 peak is a press-release number; tokens per watt under MLPerf rules is a contract term.
Section 4.6.5, "Technological Independence and Governance," is where the tender meets the dependency question head-on. Model exportability is an eliminatory criterion: every model trained on the supplied stack must be exportable in ONNX "or equivalent with proven market adoption," and the bidder must show at acceptance that at least one model trained on the delivered system runs successfully in an environment outside the supplier's. The stated reason: a national LLM that cannot leave the vendor's infrastructure is useless to public agencies, to Defense, or to partners on other hardware. Under the subheading "Evitar lock-ins," the document then requires open APIs and acceleration alternatives, naming CUDA, HIP, ROCm, SYCL and OpenCL "among others," with selection based on performance, compatibility and portability. Section 4.12 adds a minimum 320-hour training program for 40 engineers, up to four cohorts, and a supplier engineer resident in Brazil on exclusive assignment for the whole warranty period.
So a story claiming Brazil ignored lock-in would be wrong. The harder point is about layers.
ONNX portability is model-graph portability. It guarantees that the artifact, the trained weights and their computation graph, can be carried to another runtime. It says nothing about the things that make a training cluster fast: tuned kernels, collective-communication libraries, distributed-training behavior, profilers, and the operational reflexes of the team that runs it. Those live below the framework, and they are where accelerator ecosystems diverge. SCN has covered how the CUDA moat is becoming a composability problem and how expert-parallel collectives are turning into a lock-in layer of their own. A PyTorch model moved across through Huawei's torch_npu adapter will run on Ascend. Getting it to run at the utilization the tender demands is a separate engineering program.
The evidence on how large that program is should carry a date. A March 2025 CSIS analysis, citing DeepSeek's own assessment, judged it would be years before Ascend plus CANN was a viable alternative to CUDA, and compared the shift to Google's multi-year JAX/TPU migration. Since then, Huawei has committed to open the CANN compiler and virtual-instruction-set interfaces and open-source the rest of the stack by the end of 2025, and its PTO tile-instruction library did land as open source that December. None of that makes CUDA kernels run on Ascend. It does mean the CANN of August 2026 is not the CANN of early 2025, and a two-ecosystem operator is betting in part on how fast that gap keeps closing.
Brazil has written contractual defenses against application- and model-level lock-in. Those defenses do not reach accelerator- and toolchain-level lock-in, because no procurement clause can. Whether the training hours, the resident engineer, and the technology-transfer obligations add up to a team that can run both ecosystems at full efficiency cannot be judged from the documents; the government has not yet published a cross-stack staffing plan that would let anyone evaluate that burden.
"Sovereignty" means something different at each layer of this program.
Data residency is the strongest layer. Both systems sit in Brazil, and the national-cloud initiative is explicitly about government data. The tender's one soft spot, flagged by Capital Digital, is that the 5 percent interim capacity a bidder must provide within 90 days has no stated territorial requirement. Models and weights come under Brazilian control by contract, through the ONNX clause and the Rio project's national-LLM mandate.
The software layer is where control goes partial. Open APIs and a sovereign software stack are required, and the accelerator-native layer underneath stays the vendor's. Operations and skills are a work in progress, with 40 engineers at 320 hours on the tender side, 3,000 professionals over five years on the Rio side, and a resident supplier engineer in between. Silicon is foreign on both tracks for the life of both machines; the RISC-V program is the only domestic path, and it is measured in decades.
The geopolitics are real, and precision matters here. In July 2020, US ambassador Todd Chapman said Brazil could face "consequences" if it gave Huawei 5G access, while saying there would be no reprisals. In February 2021, the telecoms regulator Anatel approved 5G auction rules without a Huawei ban, after operators argued that Huawei supplied about half of the country's 3G and 4G networks. On July 15, 2026, USTR closed a Section 301 investigation opened a year earlier and imposed 25 percent tariffs on certain Brazilian goods on six grounds; digital trade and electronic-payment services were one, alongside preferential tariffs, anti-corruption enforcement, intellectual property, ethanol and deforestation. Lula said Brazil would begin proceedings under its Reciprocity Law and take the matter to WTO dispute settlement.
China is Brazil's largest trading partner. The United States is the largest ultimate source of direct-investment stock in Brazil, at about $232.8 billion in 2024 by the central bank's count. A national AI program that runs Chinese infrastructure on one track and US-origin accelerators on the other has exposure on both sides, but the exposures are not symmetric, and neither is automatic. On the US side, advanced-computing items exported to Brazil are subject to the EAR; BIS's May 31, 2026 guidance reaffirmed a license requirement for entities headquartered in Country Group D:5 or Macau, or ultimately controlled from there. Brazil is not in D:5, and nothing public suggests the Rio Grande do Norte machine trips a license requirement on its own. The Huawei track is the one to watch for a US reaction, given Huawei's long-standing Entity List status, but that is a risk to monitor rather than an intervention to predict.
China's own answer to the bloc divide is worth getting right. LineShine, the system that retook the TOP500 in June, was built entirely from domestic Arm processors and omits accelerators altogether, the one component China cannot reliably import. Brazil's wager runs the other way: operate inside both ecosystems and hold them to a common benchmark.
Brazil's program already understands that supplier diversification is not technological independence; the documents say as much. The open question is whether portability clauses, training obligations and a long-dated chip program convert two dependencies into bargaining power, or leave a small national team maintaining two ecosystems at once. The tender gives Brazil the instruments to find out, and the first reading comes with the award after October 8.
The SCN Staff is a small AI editorial squad working under human direction. Each agent owns one job.
Scout does the research. It runs down primary sources and checks what's already been published, on SCN and everywhere else, before a story gets written. If a claim can't be traced back to a real document, Scout flags it.
Forge writes. It takes what Scout found and turns it into a draft, argument and sentences and all. Every SCN piece starts here, then gets sharpened.
Cipher handles search: the titles, descriptions, and keyphrase work that decides whether a good article ever gets found. Least glamorous job on the squad. Also one that matters more than it looks.
Pixel makes the visuals. Images, charts, the occasional diagram, all built to SCN's brand instead of pulled from a stock library. When something's easier to see than to read, it goes to Pixel.
Editorial judgment and the final call stay with the humans. So does the fact-checking.
