Can South America Compete for the Next Generation of
AI Data Centers?

This article is deliberately different from Econosur’s earlier analysis of AI cost and data-center investment. That piece examined the economics of falling AI unit costs and rising aggregate infrastructure demand. The question here is site competitiveness: how South American locations compare with the United States, China and Europe when the next generation of AI capacity is placed.

The comparison starts with a large global scale gap

The United States remains the largest national data-center market. S&P Global reported 62,242 MW of US capacity in March 2026, compared with 34,015 MW in China. China’s own National Development and Reform Commission reported 13.73 million standard racks in operation by the end of 2025 and 196 TWh of electricity consumption across computing infrastructure.

Europe is smaller than the US or China in national-scale terms but already has a dense network of mature data-center markets. CBRE expects the total European market to reach around 13 GW by the end of 2026. JLL reports roughly 3.8 GW live across the five established FLAP-D markets — Frankfurt, London, Amsterdam, Paris and Dublin — with another 1.4 GW under construction and 2 GW planned.

The gap is still clear even without forcing unlike datasets into one ratio: South America begins this investment cycle from a much smaller installed base. That creates a disadvantage in ecosystem depth, suppliers, interconnection density and customer concentration. It also means that a handful of large projects can change a national market much more quickly than in the United States or Europe.

AI is changing site selection more than it is eliminating geography

Europe provides a useful comparison. JLL found that hyperscale campuses planned for 2026–2028 are located an average of 175 kilometres from major EMEA hub cities, compared with 46 kilometres for projects delivered between 2022 and 2025. Greenfield sites account for a much larger share of the pipeline. The reason is straightforward: AI-scale campuses require more land and much larger power blocks than conventional enterprise facilities, while grid connection in established hubs can take years.

China has institutionalised a related logic through its “East Data, West Computing” strategy: demand remains concentrated in the populous east, while more computing capacity is directed toward western regions with land and energy advantages. The model is not directly transferable to South America, but it demonstrates the core principle that some compute can be separated from the place where the final user sits.

PwC’s 2026 global outlook reduces the location decision to five broad forces: power; latency and connectivity; security and trusted-region hosting; GPU access and ecosystem depth; and policy certainty and community consent. For South America, those forces need to be translated into a more operational site-readiness stack.

Electricity price belongs inside the first category, but it should not be reduced to a public tariff ranking. Hyperscale sites negotiate power contracts, self-generation structures, network charges, capacity arrangements and delivery schedules that can differ materially from standard industrial tariffs. A location with nominally cheap electricity can be uncompetitive if the required power is not deliverable on time.

South America’s structural case: strong inputs, uneven conversion

The region has several characteristics that matter more in the AI era than they did in the first cloud build-out. Brazil has a large power system, a large domestic digital economy and a predominantly renewable electricity matrix. Chile combines existing hyperscaler infrastructure with extensive fiber and submarine connectivity. Uruguay operates an unusually renewable electricity system. Paraguay has hydroelectric surplus relative to domestic demand. Argentina has land, gas and large renewable resources. Colombia and Peru add large urban demand centres and new high-density facilities.

Connectivity is also improving on both coasts. Existing and new submarine systems connect Brazil, Argentina, Uruguay and Chile with North America and, increasingly, Asia-Pacific routes. This matters because low-cost power alone cannot compensate for weak network redundancy or poor latency for many workloads.

The difficulty is conversion. Power may exist in one region while transmission is constrained in another. Fiber may land on the coast while the most attractive energy site sits far inland. Water availability can limit cooling design. Environmental review can delay otherwise well-financed projects. Import duties, access to advanced GPUs and the depth of local contractors can change total project cost. The competitive unit is therefore not the country in the abstract. It is the specific site plus its power, grid, fiber, cooling, permitting and customer structure.

Brazil: scale is the advantage — and grid delivery is the test

Brazil is the only South American market that already combines large domestic cloud demand, a deep colocation ecosystem, multiple hyperscalers, substantial renewable generation and a pipeline large enough to influence national transmission planning. The Ministry of Mines and Energy reported 205 operating data centers in June 2026 and 38 GW of requests for grid-access opinions. Of those, 7.1 GW were associated with an estimated R$159 billion in investments.

A separate EPE dataset cited by the U.S. International Trade Administration places data-center demand represented in grid-connection processes through 2038 at 26.3 GW, within 54.2 GW of combined requests from data centers and hydrogen-related industrial loads. These numbers are important signals, but they are not a construction pipeline. EPE itself stresses the uncertainty around how many requested electro-intensive projects will become real connections.

The infrastructure response is already visible. EPE is studying and planning transmission reinforcement for large loads in São Paulo and the Northeast. In Ceará and Piauí, a 2026 study proposes a scalable network solution capable of supporting up to 4 GW of new electro-intensive demand from 2032.

Brazil has also moved from policy proposal to enacted law. Lei 15.504 of 15 September 2026 established REDATA, including qualifying treatment for cloud, high-performance computing and AI training and inference infrastructure. The law links incentives to conditions including renewable or low-emission electricity sourcing, a WUE ceiling of 0.05 L/kWh and domestic R&D commitments. The tax window, however, is narrower than a simple “five-year incentive” description suggests: Article 11-J gives the Article 11-C regime a five-year horizon, while the PIS/Pasep, Cofins and IPI benefits listed in items I–III produce effects only through 31 December 2026. The Import Tax suspension in item IV is not subject to that end-2026 cut-off in Article 11-J. Effective use of REDATA also depends on habilitation, co-habilitation and other procedures to be set by federal regulation. Econosur’s earlier REDATA analysis covers the legislative design; the relevant update here is that the regime is now law but still requires implementation rules.

Large projects show why maturity labels matter. Scala’s AI City concept in Rio Grande do Sul has an initial announced phase of 54 MW and US$500 million, while the site has long-term expansion potential of up to 4.75 GW. Those are not the same thing. The 54 MW phase is a project; 4.75 GW is a potential campus build-out. A separate grid milestone is now more concrete: in September 2026 Aneel authorised direct access to Brazil’s high-voltage transmission network, with the project planning for 1.8 GW of demand by 2033. The Ministry of Mines and Energy has referred to potential expansion toward roughly 5 GW in a later horizon, subject to additional system studies.

Pecém demonstrates the other side of site selection. A 300 MW data-center project is under construction, but in September 2026 Brazil’s Federal Public Prosecutor and Public Defender went to court seeking to prevent operations before additional environmental assessment and consultation with the Anacé Indigenous people. The case includes questions around water, electricity, noise and local impacts. Cheap or renewable power does not remove permitting risk.

Chile: connectivity and institutional planning, constrained by grid and environment

Chile’s competitive case is different. The Ministry of Science reports around 62,000 kilometres of fiber, connections into a 69,000-kilometre submarine-cable network and data-center capacity that grew from 35 MW in 2013 to 198 MW in 2023. The National Data Center Plan is explicitly designed to accelerate investment while managing sustainability and territorial impacts.

That planning matters because Chile already hosts major international infrastructure and has additional cloud expansion underway. The country’s value proposition is not simply renewable electricity. It is the combination of connectivity, cloud presence, institutional planning and a mature digital economy.

The constraint is physical. Chile’s renewable generation is geographically concentrated, while demand and data-center activity are centered farther south. Transmission and storage therefore matter as much as nameplate generation. Econosur covers that problem separately in its Chile power-grid analysis and battery-storage analysis.

Water and environmental review are also part of site economics. In 2024 Chile’s Second Environmental Tribunal partially annulled the environmental approval associated with Google’s original Cerrillos project over groundwater and climate-impact assessment. Google later renounced the rights and obligations attached to that environmental approval. On 17 June 2026, Inversiones y Servicios Dataluna submitted a new consultation for the same site. The redesigned proposal uses air cooling rather than the earlier groundwater-based cooling concept and includes 44 emergency generator sets with 120.25 MW of combined installed backup capacity. Environmental organisations OLCA and MOSACAT are challenging the attempt to process the redesigned proposal through a pertinence consultation rather than a full environmental assessment. Separately, Chile’s Environmental Evaluation Service published a dedicated data-center evaluation criterion on 20 January 2026; it entered into force on 25 February 2026 and covers project components, impacts, contingencies and emergency planning. Clearer rules can improve predictability, but they also raise the evidence standard that projects must meet.

Uruguay: small demand base, unusually clean power and credible infrastructure

Uruguay illustrates how a small market can attract infrastructure disproportionate to domestic scale. The country generated 98% of its electricity from renewable sources in 2025: 46% hydro, 34% wind, 14% biomass and 4% solar. Eight percent of generated electricity was exported.

Google began construction of its second owned Latin American data center in Canelones with an announced investment of more than US$850 million. In August 2026, state telecom operator Antel separately announced a US$70 million AI-focused data center in Canelones, citing connectivity and access to strategic infrastructure in the site decision.

The limitation is obvious: Uruguay does not have Brazil’s domestic cloud market or São Paulo’s interconnection density. But AI site selection can reward different attributes for different workloads. A stable, highly renewable power system, international connectivity and a credible operating environment can support a niche role that would be invisible in a simple population or GDP comparison.

For the broader telecom and cloud foundation, see Econosur’s Uruguay digital infrastructure analysis.

Paraguay: can hydroelectric surplus become compute?

Paraguay represents the clearest energy-surplus experiment in the region. MITIC and Taiwan are studying a sovereign AI computing center under a 50/50 cooperation model. MITIC describes a 10 MW first phase with market-estimated investment of roughly US$300–500 million. As of 25 August 2026, the project remained in technical feasibility study.

The state electricity company ANDE has also introduced provisions for supplying so-called convergent industries, explicitly including data centers and artificial intelligence. The policy logic is to convert abundant electricity into higher-value digital activity rather than treating power only as an export commodity or input for electro-intensive industry.

The model still has to prove its commercial depth. Ten megawatts is meaningful for a sovereign AI installation but small relative to frontier hyperscale campuses. Paraguay also needs transmission, fiber redundancy, operators, specialized construction capacity and durable demand. The project is therefore best read as a capability-building step, not evidence that Paraguay has already become a major AI data-center market.

Econosur’s Argentina–Paraguay AI infrastructure comparison examines this energy-to-compute question in more detail, while the Paraguay grid and regulation analysis covers the electricity-system side.

Argentina: high optionality, low current scale

Argentina is the most striking gap between resource potential and current installed scale. Cushman & Wakefield estimated around 73 MW of operational data-center capacity in Argentina in August 2026, compared with 2,100 MW in Brazil and 624 MW in Chile under its own methodology. Those figures should not be mixed directly with CBRE’s metro data, but they show the relative position inside one provider’s dataset.

The country has potential power and land options ranging from Patagonia renewables to natural gas around Vaca Muerta. It also has a large domestic digital economy and submarine connectivity through the Atlantic coast. But site readiness depends on the transmission system, connection timelines, equipment economics, regulation and whether projects secure real anchor demand.

The proposed Stargate Argentina project illustrates the distinction between potential and capacity. OpenAI and Sur Energy signed a letter of intent in October 2025 to explore a large-scale data-center project, with Sur Energy leading energy and infrastructure development and OpenAI considering a possible offtake role. It is not an operating facility and should not be counted as installed capacity.

Colombia and Peru: smaller markets building AI-ready nodes

Colombia demonstrates that the regional map is broader than the Southern Cone. ODATA announced US$1.3 billion of investment for BG02 and BG03 in Cundinamarca. BG03 is designed for up to 120 MW of IT capacity, while BG02 adds 24 MW. On 16 September 2026, Colombia’s environmental authority ANLA licensed the associated BG03 Navarra–Bacatá 220 kV transmission project: a 1,037-metre line, four towers, two gantries and the Navarra 230/34.5 kV substation. ANLA states that the data center currently has 50 MW of assigned connection capacity. The environmental licence covers that transmission and substation infrastructure, not the data-center building itself. The gap between 120 MW of announced IT capacity and 50 MW of currently assigned connection capacity is exactly why announced capacity and electrically deliverable capacity should be tracked separately.

Peru is earlier in scale but already has infrastructure designed for higher-density workloads. Cirion’s LIM2 facility in Lurín has 20 MW, 12,000 square metres, redundant fiber to the Lima ecosystem and the capability to deploy liquid cooling for AI environments. It is evidence of technical readiness, not evidence that Peru has reached hyperscale market depth.

The competitive map is not a country ranking

Different workload classes create different location requirements. A consumer inference service may prioritize latency and proximity to users. Sovereign or regulated workloads may prioritize trusted domestic hosting. Frontier training may prioritize very large power blocks and GPU supply. Disaster recovery may prioritize geographic separation. The same country can therefore be attractive for one workload and weak for another.

This produces a more useful conclusion than a winner-and-loser list. Brazil has the strongest existing scale. Chile has a particularly developed connectivity and planning case. Uruguay and Paraguay test smaller energy-led models. Argentina has large optionality but must prove execution. Colombia and Peru show that new AI-ready nodes can emerge outside the largest regional hubs.

The supplier opportunity sits behind the megawatts

For B2B suppliers, the relevant market is not limited to operators and servers. AI data centers pull investment through the entire physical stack. That includes high-voltage equipment, substations, transformers, switchgear, backup generation, batteries, power-quality systems, cooling, heat rejection, pumps, water treatment, prefabricated modules, racks, cabling, fiber, network equipment, physical security, fire systems, monitoring, EPC services, environmental consulting and specialized maintenance.

The buying center can also be fragmented. A hyperscaler may define technical standards while a data-center developer controls the site, a utility controls connection works, an EPC contractor controls packages and a renewable-energy partner controls the power contract. Identifying the project owner is therefore not the same as identifying the procurement route.

What must happen for South America to capture a larger AI infrastructure share?

First, requested megawatts must become committed grid capacity. Brazil’s connection queues demonstrate demand, but they also show the risk of counting applications as projects. Similar discipline is required across the region.

Second, countries need repeatable site-readiness rather than one-off flagship projects. A single Google, Stargate or sovereign AI announcement can change perceptions, but a competitive market requires multiple developers, utilities, contractors, carriers and customers that can reproduce delivery.

Third, connectivity has to expand with power. New submarine systems and inland fiber can make energy-rich regions more useful for compute, but redundancy and latency remain workload-specific constraints.

Fourth, permitting and community acceptance need predictable pathways. Brazil’s Pecém case and Chile’s Cerrillos history show that environmental and local impacts can become material project risks. Predictable standards are more valuable than nominally easy approvals that are later challenged.

Fifth, GPU access and customer demand remain strategic. PwC identifies access to advanced chips and ecosystem depth as a major location factor. Energy can attract infrastructure, but high-end AI capacity still requires hardware, cloud operators, financing and customers able to pay for the compute.

Marcus A. Volz perspective