On September 9, Magalu Cloud used its Cloud Futures 2026 event in São Paulo to announce the MGC AI Hub, an API marketplace for large language models running on infrastructure in Brazil. For nearshore engineering teams building AI products for US and European clients, a Brazilian LLM host is worth a look. It is also worth reading the fine print, because much of the announcement is still a promise.
What Magalu Cloud actually announced
According to Magalu’s own recap, the AI Hub offers inference for Llama, Qwen, DeepSeek and Gemma, plus Sabiá-4, a Brazilian model from Maritaca AI. Billing is pay-per-token and charged in reais. Speech synthesis, embeddings and rerankers are planned for later.
The same keynote introduced a partnership with Alibaba Cloud. The two companies plan to deliver generative AI services in Brazil together, including the Qwen family. Canaltech notes that the announcement did not say which Alibaba models arrive first, or give pricing or launch dates. Magalu’s recap does not mention the partnership at all, and I could not find an Alibaba-side press release, so treat the details as reported, not confirmed by both parties.
The 30% to 75% savings claim
Magalu says its model can cut LLM usage costs by 30% to 75%, depending on the option chosen. That is a vendor claim. The coverage does not say what the baseline is, which models land at which end of the range, or what per-token prices look like. A range that wide tells you the savings depend heavily on workload, so run your own numbers with your own prompts before putting it in a budget.
Where this could matter for nearshore teams
Three practical questions come out of the announcement. These are my reading, not claims from Magalu.
- Data residency. Magalu says the models run on infrastructure in Brazil, and its recap refers to national territory “and partners’ infrastructure.” If a client asks where prompts are processed, that second phrase needs a precise answer from Magalu, and the recap names no region for the AI Hub.
- Currency. Billing in reais removes dollar exposure for a Brazilian team, as Magalu points out. It does not help if your contract with a US client is priced in dollars and your costs now move with the real.
- Latency. Magalu cites under 6 ms for its br-se1 region, but that is a platform-wide figure, not an AI Hub measurement. For a US-facing product, inference time will dominate network time anyway, so measure end to end.
The models are not the differentiator
Llama, Qwen, DeepSeek and Gemma are open-weight model families that other hosting providers can serve too. What Magalu adds is the hosting terms, such as local billing and in-country infrastructure. Sabiá-4 is the one model here that serves a specific regional need, Portuguese-language quality. Whether the rest of the catalog is competitive comes down to price per token, throughput and uptime, and none of those have been published yet.
How to evaluate it
Magalu reports more than 2,000 external customers and says external revenue has grown more than tenfold since early 2025. Those are company figures, and they describe the cloud overall, not the AI Hub. A sensible pilot looks like this:
- Pick one production workload and replay a week of real prompts against the AI Hub and your current provider.
- Compare cost per thousand requests, p95 latency from your clients’ regions, and output quality on your own evals.
- Ask Magalu in writing which regions process requests and which partner infrastructure is involved.
Bottom line
A Brazilian LLM host with local billing is a real option for nearshore teams, but it is an announcement, not yet a benchmark. Until prices, regions and the Alibaba terms are public, the right move is a small, measured pilot. If you are weighing where to run AI workloads for a US or European client, talk to the Luby team.