AI Neoclouds in Latin America, Explained

A neocloud buys GPUs, racks them, and rents the compute by the hour. It is not a colocation provider, which sells floor space and power, and it is not a hyperscaler, which sells a hundred managed services. It owns silicon and sells compute, which means it carries the chip risk and collects the chip rent. Last week this site counted Latin America’s data-center megawatts. This piece asks who in the region is actually in that business.

Three companies say they were first. In the first week of February 2026, Brazil’s Claro announced it had become Latin America’s first NVIDIA Cloud Partner for data processing and AI. On 4 March, Axia Energia — the former Eletrobras — announced what it called Latin America’s first neocloud. And on 22 June, Mexico’s Bleeding Edge declared its Blackwell B300 AI Factory in Querétaro fully operational; DCD’s July report says it claims to be the first neocloud in Mexico and Latin America on NVIDIA Blackwell.

Axia’s platform was built with the electrical-energy research centre Cepel and installed in Rio de Janeiro. Launch coverage put it at 32 NVIDIA H200s in service, with expansion to 96 planned by the end of March, aimed at grid simulation, scientific computing and AI, with access to be opened to universities, research institutes and startups; we found no public confirmation that the expansion has been completed. Claro’s fleet is 256 GPUs, on its own director’s account. Bleeding Edge says its entire initial capacity is already contracted, but discloses no GPU count.

Now add the two Brazilian claimants together. Axia at 96 GPUs if fully built, Claro at 256: at most about 350 chips, and Claro’s share is not yet in the country. The other local operators that publish a number are smaller still — BITS says 64 GPUs; Uruguay’s DigitalIA runs ten V100s and five MI60s in Montevideo. The one local claim large enough to change this arithmetic is AI-GDC’s “thousands,” which rests on its own website.

Compare it chip for chip. Epoch AI credits Colossus 2 in Memphis with 110,000 B200-class and 330,000 B300-class chips — about 440,000 accelerators as of 15 June 2026, across a three-building campus whose third building was still under construction, sourced to SpaceX’s S-1. On 25 September Elon Musk put it higher on X — 110,000 GB200 and 440,000 GB300 — a claim rather than a filing. Even on Epoch’s June count, one American campus holds roughly 1,250 chips for every one the two Brazilian claimants disclose.

Two scoping notes, because this is the number most likely to be misread. It counts disclosed local operators only — the hyperscalers plainly hold far more silicon inside Brazil than this, since AWS could not offer P6-B300 in São Paulo otherwise, but none of the three discloses an in-region chip count. And it counts what has been disclosed; an undisclosed fleet is not a fleet of zero.

Claro’s claim deserves the closest reading, because it is the one with NVIDIA’s name on it. The February release carries an endorsement from Marcio Aguiar, NVIDIA’s Enterprise director for the region. Claro says the base could enable it to build Brazilian LLMs and SLMs, fine-tune for Brazilian Portuguese, and sell LLM tokens at competitive prices as Brazil’s first telco sovereign-AI provider.

Read the mechanism rather than the announcement. Claro acquired the AI infrastructure and it is “operated with Oracle”; NVIDIA’s own quote describes Claro using “an accelerated computing platform with Oracle.” In Claro’s product navigation, Claro GPUaaS sits directly beside Claro Cloud by Oracle. Claro bought the chips; Oracle provides the building. In April, Claro CEO Rodrigo Marques told Convergência Digital that the platform runs from an Oracle data center outside Brazil because the tax burden makes importing the GPUs uneconomic: “Sem o Redata ou benefício fiscal não vai ter data center para IA aqui” — without Redata or a tax benefit, there will be no AI data center here. Hold that sentence; the tax law it names comes later in this piece.

Now read what the product page actually sells. It promises pay-per-use capacity, LGPD alignment, data processed in Brazil, and — in the page’s own title — infrastructure “em Reais”. It names no GPU model, no cluster size and no price.

Then read what Claro’s own executive says about where the chips are. In an interview with the Brazilian telecoms publication Convergência Digital, Mário Rachid — the same Claro empresas digital-solutions director quoted in the February announcement — states it plainly: “Today our GPUs are outside Brazil, for cost reasons, but we have invested in upgrading the data center for power and cooling, and the 256 GPUs purchased will come to Brazil to meet sovereignty and data-protection demands.” The offer runs 100% in Claro’s São Paulo data center from January 2027.

So the fleet is 256 GPUs — now on the record from Claro rather than inferred — and for the moment they are not in Brazil. Latin America’s first NVIDIA Cloud Partner is selling Brazilian data sovereignty on hardware that sits in another country until January. TELETIME reported as much at the June launch: the GPUs behind the offer were hosted outside Brazil. The product is also sold by the fraction: per the same report, the smallest unit sold is one MIG — the smallest sliver NVIDIA lets you partition a GPU into — with a one-hour minimum. In August, a Claro empresas executive told TELETIME that customers could save up to 50% on acquisition costs.

TIVIT’s competing Brazilian offer, GPU for AI, does name the silicon — NVIDIA H100 on dedicated physical machines in its private cloud — and pitches stable pricing against the public cloud, but likewise publishes no capacity and no rate.

Claro’s silence on price is not a regional trait: two Brazilian vendors covered below, GPUBrasil and EVEO, both publish rates against named chips, so the reticence is Claro’s and TIVIT’s rather than the region’s. And Claro is not merely reselling — Rachid says the 256 GPUs were purchased. It owns the silicon and rents the platform and the building around it, which is a chip owner and a facilities tenant at once. Its stated differentiators are a currency and a data-protection statute; until January, the jurisdiction is the part it does not yet have. By its CEO’s account, what keeps the chips abroad is not engineering. It is tax.

The hyperscaler grid is not the whole menu. Local operators fill some of the gaps, at far smaller scale: in Mexico, Bleeding Edge declared a Blackwell B300 AI Factory in Querétaro fully operational in June and says its entire initial capacity is already contracted; BITS advertises 64 NVIDIA GPUs in operation in Chihuahua, billed in pesos; AI-GDC in Nuevo León says on its own site that it runs “thousands” of Blackwell GPUs across 10 MW and targets 5,000 in the second half of 2026 — a claim no independent source has yet confirmed. In Brazil, EVEO lists H200 and HGX B300 servers from its São Paulo SP2 facility.

Brazil’s bench is deeper than the three claimants suggest, though none of these publishes a fleet size. Magalu Cloud, the public cloud of retailer Magazine Luiza, said in September and November 2025 that it offers NVIDIA L40 VMs in one- and two-GPU configurations, priced in reais; its regions are in Brazil, but it does not say which region hosts the GPUs, and its product page lists dedicated-GPU instances without a model, count or rate. Under Labs sells bare-metal H200, RTX PRO 6000 and L4 servers from two facilities in Greater São Paulo, Cotia and Barueri, on contracts in reais, and markets the product as sovereign bare-metal GPU; no price is published. Audaks, a São Paulo IaaS host, says it offers A4000, A6000, L4, L40S and H100 capacity from a Tier III data center in São Paulo, quote-only, according to a company blog post. Cirion advertises GPU-optimized infrastructure for AI training and inference in its Rio de Janeiro data center, naming no model, count or price. Regionally, EdgeUno and AtlasCloud announced in November 2025 a partnership to deploy H100 and H200 clusters in São Paulo, Bogotá, Santiago, Buenos Aires and Mexico City; EdgeUno’s cloud pricing page, checked on 28 September, lists VPS, VPC and bare-metal products and no GPU.

Two Brazilian vendors publish prices, and the prices say something the fleet sizes do not. GPUBrasil quotes an hourly rate card in reais — as listed on 28 September, H200 at R$25.83 an hour, H100 SXM at R$14.40, A100 80GB at R$14.94 — and sells Claude, GPT and Gemini tokens from the same balance through an OpenAI-compatible endpoint. EVEO quotes monthly: H200 at R$27,930, an RTX PRO 6000 SE at R$13,150, a 2×T10 box at R$1,690.

Now read what those two pages actually promise, because it is not what a neocloud promises. GPUBrasil’s stated advantages are payment advantages: Pix, no subscription, no IOF, no card spread. Its availability listing names North America, Europe, Asia-Pacific and a “Global” option, and no Brazilian location, and the site names no operating company and no CNPJ. It is a Brazilian payments front-end on foreign compute, not a Brazilian neocloud. EVEO does name its facilities — and offers every tier from both São Paulo SP2 and Miami FL1, billed in reais either way. Its pitch is that the card is already racked and energised; its own specification table gives lead times of up to three business days on the entry box, up to 15 on the H200 and RTX PRO tiers and up to 45 on the HGX B300, which is quote-only. Neither is an on-demand cluster.

So one pattern in this market is narrower and more revealing than a shortage of chips. Some Latin American GPU vendors compete on currency and payment rails, and treat the silicon’s location as secondary. Claro’s page promises data processed in Brazil while its director says the chips are abroad until January. GPUBrasil bills in reais for GPUs it lists abroad. EVEO bills in reais whether the server sits in São Paulo or Miami. The Dominican Republic’s Nebula describes an NVIDIA-only facility in Santo Domingo with an HGX H100 node marked online — and on the same page lists cluster nodes in Virginia, Dallas, Miami, Panama City, São Paulo and Curitiba, without saying which GPU runs where. Not every vendor works this way — BITS promises “100% Datos en territorio MX,” AI-GDC says “100% en México,” Uruguay’s DigitalIA names its Montevideo rack, and in Brazil Magalu Cloud and Under Labs both describe Brazil-based GPU offers — but in the one respect that decides whether a vendor is any use for a sovereignty requirement, the invoice currency is often clearer than the address. That is a billing layer maturing faster than a compute layer.

What no operator in the region has independently verified is a cluster — thousands of accelerators on one fabric, under one scheduler, for months, which is the only product a frontier model is trained on. The qualifiers matter. AI-GDC says on its own site that it runs thousands of Blackwell GPUs in racks of 72; that is a disclosure, not yet a confirmation. Bleeding Edge’s Querétaro AI Factory is Blackwell B300 and fully contracted, but publishes no GPU count. The biggest single coherent machine any regional operator has put its name to is one rack: on 21 September Ecuador’s Telconet announced it had ordered a Vera Rubin NVL72 — 72 Rubin GPUs and 36 Vera CPUs on one NVLink domain, due at the end of 2026 — which it says is the first such order in Latin America. That is a serious machine and another telco taking the chip risk. It is still a rack, not a campus. The hyperscalers disclose no in-region chip counts, and Oracle, whose seven Latin American regions are covered below, publishes no per-region GPU availability at all. Any of them could already be running a Latin American cluster. On the verifiable record, none is, and the gap between a rack and a campus is the whole question.

Latin America is not four countries, and the smaller markets tell the same story in sharper type. This is the best locally hosted GPU offer we could verify in each, from the operators’ own pages and filings as of 28 September.

Read down the status column. The one serious machine on this list, Telconet’s NVL72, is on order and is a single rack. The biggest number, X8 Cloud’s 4,000-plus H100s for Paraguay, belongs to an announcement of up to US$50 billion over 30 years; ObrasParaguay, a Paraguayan public-works tracker, listed it in May as in planning, subject to approvals, location to be confirmed, and a former senator, Desirée Masi, has questioned the company’s limited operating history. What is listed today is a desk-side developer box in Costa Rica, a rack of 2017–18-generation accelerators in Uruguay, gaming-class cards in Argentina, and an H100 node in the Dominican Republic whose siblings are listed in Virginia and São Paulo.

Peru and Colombia are the two rows where the only local evidence is an executive’s word or a marketing line. Cirion says it has GPU-optimized capacity in Lima but publishes nothing a buyer could price. In Colombia, the state’s own BioNube sovereign-cloud project, three campuses planned for Santa Marta with the UAE’s G42, is still in development, and the Procuraduría has opened a preventive review of related contracts.

And the location pattern repeats. DonWeb’s Peruvian storefront prices GPU servers in soles; the servers are in Rosario, Argentina. Telecom Argentina is converting old telephone exchanges into edge data centers of 10 to 50 MW “for present and future AI demand,” but we found no GPU service attached to them. For Bolivia, Guatemala, Honduras, El Salvador and the Caribbean beyond the Dominican Republic, we found no locally hosted GPU offer we could verify. El Salvador is the exception that proves the distinction: its government ordered NVIDIA B300s in 2025 and, per Bloomberg Línea, took delivery at a geothermal site in November, but that is state-owned compute, not a rental offer. That is a statement about the public record, not proof that none exists.

The local operators are not competing in a vacuum. If a Brazilian bank or a Mexican retailer wants a frontier GPU inside its own jurisdiction tomorrow, the first calls go to the hyperscalers, and all three publish machine-readable availability by region. The answer is not what the megawatt figures imply.

Read the columns, not the rows. Brazil can rent a current-generation training GPU from two of the three, and Blackwell from both: AWS lists P6-B300, and Azure’s price list carries an ND GB200 v6 meter in Brazil South alongside the ND H200 v5. AWS’s São Paulo region carries P6-B300, P5e and P5, the HGX-class Blackwell and Hopper instances; P6-B300 arrived there only on 28 August 2026. Google’s São Paulo region carries no A3, A4 or A4X instance at all, only the RTX PRO 6000-based G4, switched on on 24 September, and an eight-year-old T4. Azure’s retail price list prices the ND H200 v5 in Brazil South at US$169.60 an hour, against US$93.28 in Mexico Central, an 82% premium for the same node. A price listing is not proof of capacity, but it is the same evidence this piece accepts for Mexico. We found no GPU SKU in Brazil Southeast.

Mexico can rent an H200 cluster from exactly one of the three: Azure. Azure’s Mexico Central lists ND H200 v5 — no A100 and no T4 — but it is not training-only: the price list also carries A10 v5 and NVas_v4 SKUs. AWS’s Mexico region publishes general purpose, compute, memory and storage instance families and no accelerated computing section. Google’s GPU regions-and-zones table does not list Mexico.

Chile can rent nothing from the big three. Google’s GPU zone table lists none in Santiago, Azure’s price list carries only four NVas_v4 SKUs in Chile Central and no training-class GPU, and AWS’s Chilean region, promised for the end of 2026, is not yet listed. The local alternative is Gtd, whose GPUaaS page promises dedicated GPU capacity “ubicada en territorio nacional” and fixed prices in pesos — without naming a GPU model. This is the country that leads Latam-GPT, the region’s flagship open language model.

Two caveats on that table, and the second one is the more important.

First, it is availability, not intent. Amazon has announced a Chilean region, backed by more than US$4 billion, for the end of 2026; Google’s first Blackwell-generation part in Brazil, the RTX PRO 6000 in G4, only landed on 24 September, and it is not an HGX-class training part. Neither announcement puts a training cluster in the table yet. Announced, contracted, under construction and rentable-today are four different columns, and only the last one is above.

Second, and this is a hole rather than a caveat: the table covers the three clouds that publish GPU availability by region. Oracle does not. OCI lists seven Latin American regions — São Paulo, Vinhedo, Santiago, Valparaíso, Bogotá, Querétaro and Monterrey, more than any of the other three — and its shape catalogue runs from H100 and H200 to B200, B300, GB200 and GB300, plus AMD’s MI300X and MI355X. The catalogue does not say which shapes exist in which region; it points instead to per-tenancy service limits. So published documentation supports no claim either way about Oracle GPU capacity in Brazil or Mexico. The one in-region statement we found is 18 months old and from Colombia: Oracle’s country manager told Forbes Colombia in March 2025 that the company had brought NVIDIA GPUs into the country, without naming a model, a count or a product, and we found nothing published since that confirms a rentable offer. What is on the record for Brazil is narrower and more pointed: Oracle is the platform Claro’s sovereign-AI offer runs on, and Claro’s own CEO has said the GPUs behind it sit in an Oracle data center outside Brazil.

Latitude.sh is a Brazilian-founded bare-metal provider and the closest thing Latin America has to a homegrown bare-metal cloud. Its public catalogue, read from its own pricing data, lists 17 bare-metal plans across 26 locations. Six of those locations are Latin American: São Paulo and São Paulo 2, Buenos Aires, Santiago 3, Bogotá and Mexico City 2. Brazilian, Argentine, Chilean, Mexican and Colombian availability appears on twelve of its fourteen non-GPU plans.

Three plans carry GPUs: one H100 80GB, eight HGX B300, eight RTX PRO 6000. All three are offered in the United States. None of the three is offered in any of the company’s six Latin American locations. For a São Paulo customer, the nearest GPU in the catalogue is in Miami.

The same catalogue prices the identical eight-way RTX PRO 6000 node on demand at US$48/hour in the United States, US$52 in Japan and US$66 in Australia. Geography carries a 37% premium at the top of that range. Latin America is not priced at a premium. It is not priced.

The last piece covered Law 15.504 of 15 September 2026 and the Redata regime as a cooling specification. Re-read for this question, the statute is something more deliberate.

The law works by inserting Arts. 11-A to 11-J into Law 11.196/2005. Art. 11-A §1º defines qualifying “datacenter services” to include cloud computing, high-performance processing, and the training and inference of artificial intelligence models by name. Brazil did not write a tax break for buildings. It wrote one for the neocloud business model, explicitly.

And Art. 11-B §1º I is the provision that would force a domestic market into existence: at least 10% of the effective supply installed under the benefit must be made available to the Brazilian domestic market, and it may not be diverted to export or own use even when domestic demand is lacking. The quota covers processing, storage and data-treatment capacity rather than GPUs as such, but ten per cent of a hyperscale campus’s processing would be a real compute rental business. Then come the qualifiers, all in the same article:

• §6º lets an operator delete the obligation entirely. The 10% domestic-supply commitment may be substituted by investing an additional 10% of the value of the benefited goods into R&D. An operator that would rather not sell compute in Brazil can write a cheque to a research institute instead.
• §§2º–3º let it be met without a commercial market. The supply can be transferred free of charge to research institutions or the public sector rather than sold — and free transfers count toward the 10% with a multiplier still to be set by regulation. Set that multiplier generously and a donation discharges the quota.
• §7º cuts it by a fifth for establishments in the North, Northeast and Centre-West — taking the floor to 8%.
• Art. 11-C §4º gates the whole thing: the tax suspension applies exclusively to products listed in a federal executive act. On 25 September, ten days after enactment, ConJur reported that the implementing regulation — the products list, the habilitation procedure, the sustainability indicators, the proof deadlines and the free-cession multiplier — had not been issued, and we found none by 27 September. Habilitation, under Art. 11-A §6º, sits with the Receita Federal. Until that list exists, nothing is suspended.

Then the timing. Art. 11-J gives the incentives a five-year term — but its sole paragraph provides that the benefits on PIS/Pasep, Cofins, their import equivalents and IPI produce effects only until 31 December 2026, under the consumption-tax reform of Constitutional Amendment 132/2023 and Complementary Law 214/2025. Three of the four categories of tax the law suspends have, as drafted, roughly three months to run. The import duty, which the sole paragraph leaves out, is by implication the one that survives on the five-year clock.

The reason is the reform, not a drafting slip: PIS/Cofins give way to the new CBS and IBS from 2027. The problem is what replaces the relief. Trench Rossi Watanabe’s alert on the law notes that it provides, “at least for now,” no equivalent treatment for CBS or IBS. A 25 September analysis in ConJur adds that the reform’s full, immediate credit on capital goods does not, so far, generally cover typical data-center equipment. On current law, an operator that imports GPUs in 2027 gets the import-duty suspension and a question mark on the rest.

A regime that names AI training as a qualifying service, makes the domestic-supply quota optional, and sunsets most of its own tax relief at the end of the quarter is not a plan to build a Brazilian neocloud. It is a plan to attract capital and hope one shows up. Claro’s CEO has already said, in April, what happens without the benefit: the data center stays abroad.

Where the market has not produced a neocloud, three governments are doing the procurement themselves, with markedly different degrees of seriousness.

• Brazil is in the market right now. On 20 August 2026, President Lula announced an AI supercomputer for the Augusto Severo Science and Technology Park in Macaíba, metropolitan Natal. The technical specification is LNCC’s; the money is FNDCT’s; the instrument is FACC Public Selection Notice nº 27/2026, published the same day in the DOU (edition 157-A, extra), a technique-and-price procurement under Decree 8.241/2014. The value stated in the selection instrument is R$959,040,959.04. LNCC itself flags in its announcement that this figure must be distinguished from the publicised numbers — a federal AI-infrastructure package of roughly R$2.3 billion, within which the Rio Grande do Norte machine was announced at approximately R$1.06 billion. Three numbers, one machine, and the government’s own laboratory telling readers which one is contractual.
  
  The other half of that R$2.3 billion package matters just as much to the question of who owns the GPUs. About R$1.3 billion goes to a Rio de Janeiro supercomputing project with Huawei and iFlytek to develop large language models, with the Chinese partnership scheduled to begin in July 2027. For the Natal machine, government officials told Reuters anonymously that they expect NVIDIA to win, and science minister Luciana Santos had earlier said the same to Folha de S.Paulo. The government’s stated principle is “not to depend on a single company, technology or country.” In practice, that means one Chinese stack and, most likely, one American one. The Chinese side may extend to the chips themselves: in June, Huawei Cloud’s Latin America president, Mark Chen, said at Web Summit Rio that the company is studying running its newest Ascend AI chips in its Latin American cloud — with no timeline, and with the Ascend 950 family still early in deployment in China.
  
  The live detail matters more than the announcement. FACC’s own tender page lists this as an in-person proceeding with envelopes opened at 10:00 on 4 November 2026 at its Petrópolis headquarters, sealed bids accepted until the preceding business day. The notice published in the DOU set the session for 8 October, as August coverage reported; Errata nº 01, signed 21 September, moved it to 4 November. Anyone underwriting this should read the errata, not the August write-ups.
  
  And Errata nº 01 rewrote the acceptance criteria. Item 4.5 moved the benchmark floor from MLPerf Inference and Training v6.0 to v6.1. The named LLM inference workloads changed from Llama 2-70B, Llama 3.1-405B, Mixtral-8x7B and DeepSeek-R1 to Llama 3.1 8B, DeepSeek-R1 and GPT-OSS-120B, in the MLPerf v6.1 Datacenter Closed Division, Offline and Server scenarios, with official LoadGen, TTFT/TPOT quality targets — plus a mandatory Agentic Inference workload. All three named inference workloads appear in MLPerf Inference v6.1, the round published on 16 September 2026, which also retains Llama 2-70B; the errata moves to the current suite and chooses within it. The effect is still notable: Brazil is not buying teraflops. It is buying an audited MLPerf submission, and a Chinese open model and OpenAI’s open-weights model are now among the things the machine must demonstrably serve.
• Chile already rented. Latam-GPT, launched 10 February 2026, is the region’s flagship open model — a Llama 3.1 70B base plus a licensed regional corpus, led by CENIA with Chile’s science ministry, CAF, Data Observatory and Amazon Web Services as named partners. AWS’s Southern Cone technology lead is quoted stating that AWS optimised the infrastructure and cut training time by 64%, from 25 days to nine. The sovereignty project ran on a US hyperscaler’s cloud, and the release says so plainly.
• Argentina has a letter. OpenAI’s own 14 October 2025 statement describes a Letter of Intent with Sur Energy to explore a large-scale data center, with OpenAI saying it “welcomes the potential opportunity to become an offtaker.” The widely-quoted 500 MW and US$25 billion figures appear nowhere on OpenAI’s page. What does appear is the structural admission: Sur Energy would lead the formation of a consortium, “including partnering with a cloud infrastructure developer.” Eleven months on, the neocloud seat in Stargate Argentina is still described as a seat to be filled — and trade coverage through mid-September 2026 reports no final investment decision and no construction start, with candidate sites in Neuquén province, Senillosa and Arroyito, none confirmed.

The research side is organising too, and its diagnosis matches this one. On 22 September, at the CARLA conference in Córdoba, EuroHPC kicked off an EU-LAC Supercomputing Network for AI spanning Argentina, Brazil, Chile, Colombia, Costa Rica, Mexico and Uruguay, naming “fragmented HPC capacity, uneven access to computing resources, investment needs and skills gaps” as the problems to solve.

Power developer, offtaker, tax regime, tender, model, and a domestic rental market with a public hourly rate. Every role in the AI stack now has a named Latin American actor. None of them has scale — and scale is the only thing separating a GPU business from an AI one.

For anyone selling, financing or underwriting into this, three things follow.

Stop quoting regional megawatts to a compute buyer. The unit that matters to a customer is an instance type in a region, and it does not aggregate. A Mexican enterprise that needs an H200 cluster has one hyperscaler to call, plus a handful of small local operators. A Chilean one has no hyperscaler GPU in-country and one local provider that does not name its chips. A Brazilian one has HGX-class Blackwell on demand from a single hyperscaler, or on a quote with up to 45 business days’ lead time from a local host. Those are three unrelated sales motions inside one “Latin American AI market,” and the regional inventory figure predicts none of them.

The neocloud gap is the region’s clearest unoccupied position — and the reason it is unoccupied is financial, not technical. A neocloud is a balance sheet that buys depreciating silicon against multi-year offtake. Brazil’s answer was a tax regime whose largest components expire in December and whose domestic-supply quota can be bought out for a research grant. Claro took the chip risk on 256 GPUs but not the Brazilian tax bill, and parked them in an Oracle facility abroad until its own data center is ready. Until someone in the region is willing to own inventory at scale, and on Brazilian soil, the layer stays foreign, and Redata’s 10% is the only lever anyone is holding. Watch the implementing regulation — specifically the free-cession multiplier and the products list — more closely than any announcement.

Price the currency, and read the fine print on jurisdiction. CBRE names currency volatility as a contract risk for international operators in São Paulo, and Claro turned that risk into its headline feature: billing in reais, data in Brazil, LGPD on the page. But its own director says the GPUs sit abroad until January 2027. If you are buying Brazilian compute for a sovereignty or LGPD requirement, the question to put in writing is not whether the vendor markets local processing — it is which rack, in which building, on what date. And do not confuse the tiers: the suppliers that publish a rate card in reais sell single cards and eight-GPU boxes, some of them hosted abroad, while the three companies calling themselves neoclouds run small or undisclosed fleets and publish no rate at all. Ask for the address, the invoice currency and the interconnect in the same sentence.

The last piece ended by saying the region had until 2027 to prove the difference between a signed contract and a running rack. Add a second gap to watch. A rack that runs on chips you do not own, in a cloud you do not control, is not sovereignty. It is a lease with a flag on it.