Why AI’s Cost Problem Could Reshape
Data Center Investment in South America

Econosur’s existing AI data centers and digital infrastructure analysis establishes the regional baseline: data centers depend on power, fiber, cooling, telecom networks, operators, regulation and paying digital demand. The question in this analysis is narrower and more economic. What happens to location decisions when useful AI becomes cheaper to consume but much more infrastructure is required to serve the resulting demand?

Why cheaper AI can require more infrastructure

There is an apparent contradiction in the economics of AI. The price of a unit of model capability is falling very rapidly. Stanford HAI calculates that the cost of querying a model at GPT-3.5-level MMLU performance fell from about US$20 per million tokens in November 2022 to US$0.07 by October 2024. The same report, using Epoch AI data, notes annual inference-price declines ranging from roughly 9-fold to 900-fold depending on the benchmark.

Lower inference prices do not imply that total compute spending must fall. Cheaper intelligence can unlock more applications, longer context windows, more agent steps, larger enterprise deployments and more frequent model use. OpenAI provided a current company-level example in September 2026: according to Reuters, an 80% price reduction for its lower-cost Luna model was followed by a roughly tenfold increase in usage. That is not proof of a universal economic law, but it illustrates the mechanism.

The IEA reaches the infrastructure side of the same problem. Its April 2026 central projection sees global data-center electricity consumption rising from 485 TWh in 2025 to about 950 TWh in 2030. AI-focused data-center electricity consumption is expected to triple over that period. The power density of AI servers increased about elevenfold between 2020 and 2025 and could rise another fourfold by 2027.

The commercial consequence is important. AI providers need to lower the cost of delivering useful intelligence while simultaneously securing much more physical capacity. OpenAI states that compute is a critical input in a cycle where more compute enables better models, better models drive more usage and revenue, and revenue supports further infrastructure investment.

Anthropic shows the same tension from another angle. Reuters reported in August 2026 that the company’s valuation case depends partly on the expectation that revenue will eventually grow faster than the cost of GPUs, training, inference and infrastructure. Anthropic’s fast revenue growth therefore does not eliminate the compute problem. It raises the importance of whether compute becomes a smaller share of each dollar of revenue as the business scales.

The location decision is becoming a systems problem

For conventional software, geography can be secondary. For hyperscale AI infrastructure, geography is embedded in the cost structure. A data-center site needs power generation, firm grid access, substations, transmission, backup systems, cooling, water strategy, fiber redundancy, construction capacity, equipment access, permits, technical staff and long-term financing.

The IEA explicitly notes that data-center investment has become too large to be funded only from corporate balance sheets and will increasingly depend on capital markets. Financing conditions therefore join power and connectivity as a physical constraint on expansion.

This is why South America should not be evaluated through electricity prices alone. Argentina can offer Vaca Muerta gas, wind and cool Patagonian locations, but large projects still need fiber, substations, contracts and finance. Brazil has scale and demand, but grid-access requests are running far ahead of installed capacity. Chile has a mature digital market, but its energy regulator is warning about transmission timing. Uruguay combines renewable power and institutional stability, but its advantage is partly the ability to convert those conditions into actual projects. Paraguay has abundant hydroelectricity, but much of its current large digital load is still associated with Bitcoin mining rather than AI.

Five markets, different evidence states

Headline megawatts are not comparable unless the underlying status is visible. A grid-access request is not a financed project. A letter of intent is not a construction order. Operating Bitcoin-mining capacity is not automatically AI-compute capacity. A planned cloud region is not operational until it launches.

Argentina: energy abundance meets execution risk

Argentina is currently the clearest case of a country trying to turn an energy proposition into a new digital-infrastructure proposition. The country starts from a small operating base compared with Brazil or Chile, but its project announcements are very large.

OpenAI and Sur Energy signed a letter of intent in October 2025 to explore a large-scale data center in Argentina. Reuters reported the proposal at up to 500 MW and up to US$25 billion at full scale. The wording matters: the project is a proposal based on a letter of intent. It should not be described as a financed or under-construction 500 MW facility.

The current pipeline is broader than Stargate. Reuters reported on 7 September 2026 that Pampa Energía is evaluating a data center next to its Loma de la Lata thermal power plant in Neuquén with potential consumption of up to 500 MW. Interested investors are asking first for a 20–40 MW pilot. Pampa estimates that electrical infrastructure for the full 500 MW could cost almost US$900 million.

FlexDomes is seeking investors for a 120 MW first-stage project in Neuquén with an estimated first-stage cost of US$1.4 billion. Green Capital has described a 300 MW first stage in Chubut costing about US$3 billion, with a much larger long-term concept, but still needs investors. In Bahía Blanca, Pampa has agreed to provide 30 MW for the first stage of an IT park whose concessionaire expects final investment agreements by the end of 2026.

Argentina has also changed the regulatory perimeter. Decreto 105/2026 explicitly includes artificial intelligence among the technology activities covered by the RIGI framework. That improves the formal investment route, but incentives do not remove the need for bankable contracts, fiber, grid works, equipment and project finance.

Econosur’s earlier Argentina–Paraguay AI infrastructure analysis describes this as the energy-to-compute problem. Argentina’s current challenge is to turn resource availability into contracted, connected and financed compute capacity.

Brazil: scale, policy and a grid pipeline far larger than the operating market

Brazil is different because it already combines a large digital economy, cloud demand, financial services, telecom infrastructure and a substantial operating data-center base. The Ministry of Mines and Energy reported in June 2026 that Brazil had 205 operating data centers.

The pipeline is much larger than the installed base. The same ministry cited 38 GW of data-center grid-access requests and said requests for new projects had grown 330% between 2024 and 2025. Those requests are a demand signal for the power system, not 38 GW of committed construction.

Brazil is beginning to translate the demand into transmission planning. EPE published a June 2026 study proposing a scalable solution for up to 4 GW of electro-intensive loads in Ceará and Piauí from 2032, explicitly including data centers and hydrogen plants. That makes the bottleneck visible: the competitive question is not only whether renewable electricity exists, but when large loads can actually connect.

REDATA adds the policy layer. The Senate approved PL 278/2026 on 1 September 2026 and transmitted it for presidential sanction on 4 September. The bill provides tax benefits for qualifying data-center equipment while attaching infrastructure conditions. Participating companies must direct at least 10% of installed processing, storage and data-treatment capacity to the domestic market, cover contracted electricity demand through renewable or low-emission sources via supply contracts or self-production, and meet an annual cooling water-efficiency index of no more than 0.05 litre per kWh, alongside domestic investment obligations. The final Senate wording replaced “clean” with “low-emission”, widening the energy-policy question; whether natural gas qualifies will depend on how the regime is implemented and regulated.

The Econosur REDATA analysis is therefore directly connected to the AI cost question. Brazil is not competing only through cheap electricity. It is trying to combine tax treatment, domestic compute, sustainability rules and power-system expansion into an industrial policy for digital infrastructure.

The private sector is already building energy structures around that demand. Ascenty and Casa dos Ventos announced a renewable self-production arrangement covering 110 average MW and more than US$500 million. Ascenty’s 2025 financial statements state that the 110 MW average power-supply contract is scheduled to begin on 1 January 2028 and run for 15 years. Casa dos Ventos and Ascenty separately state that the two associated generation projects are expected to begin operations in 2027. Those dates are not contradictory: project commissioning is expected before contractual supply to Ascenty begins. The distinction between project operation, contracted supply and current data-center consumption is important.

Chile: a mature digital market meets transmission and water constraints

Chile already has a meaningful operating data-center market, international cloud demand and strong Pacific connectivity. The National Data Center Plan says capacity grew from 35 MW in 2013 to 198 MW in 2023 and projects a tripling over the following five years. It also cites about 62,000 kilometres of fiber and connections into roughly 69,000 kilometres of submarine cables.

Current market-capacity figures need careful interpretation. InvestChile reported 166 MW operating in April 2026 based on a JLL dataset, while Cushman & Wakefield reported 279 MW of operational capacity in Santiago for the second half of 2025. Those figures should not be silently combined or treated as measuring an identical universe. Different provider definitions, geographic boundaries and capacity methodologies can produce materially different totals.

The more important signal is physical. In May 2026, Chile’s National Energy Commission warned that expansion of electricity infrastructure was not advancing at the same pace as data-center demand. The regulator highlighted transmission constraints, the need for storage, interruptible loads and better medium-term planning.

This directly connects to Econosur’s Chile power-grid analysis and battery-storage analysis. A country can have abundant renewable generation and still face a location bottleneck if the required power cannot reach a specific site at the required time and reliability.

AWS adds a major demand and investment signal. AWS continues to list a South America (Chile) Region as coming in 2026, with three Availability Zones. The company says it plans to invest more than US$4 billion in Chile over 15 years to construct, connect, operate and maintain its data centers. As of 10 September 2026, that Region should still be classified as planned rather than operational.

Water adds another design constraint. Chile’s Cerrillos Data Center case became a prominent example after environmental litigation over groundwater impacts. The project developer had already proposed replacing water-based cooling towers with air-cooled chillers in 2022. In 2024, the Second Environmental Tribunal partially annulled the approval and required the environmental process to consider climate-change effects on the Santiago Central aquifer.

Uruguay: small market, real construction and a water-design lesson

Uruguay demonstrates why a smaller domestic market can still become relevant when a project has a named investor, physical site, environmental process and construction activity. Google announced more than US$850 million for a new data center in Canelones in 2024. Uruguay XXI currently describes construction as underway, and President Yamandú Orsi visited the works in April 2026.

The cooling system is particularly relevant to the regional comparison. Uruguay’s environmental project record states that normal operation will use an air-cooling system. Government reporting from the construction launch also emphasized that the project uses air rather than water for server cooling.

That design history is important because it shows that water is not merely an ESG talking point. Cooling technology can change the project configuration and the environmental acceptability of a site. Chile and Uruguay therefore provide two different examples of how water constraints can influence data-center design.

Uruguay is also building a domestic cloud and AI layer around its telecom system. Antel and Google Cloud announced in February 2026 that Google Distributed Cloud infrastructure would be hosted at Antel’s Pando data center. In August, the Uruguayan government announced a separate US$70 million Antel data center specialized in AI in Canelones.

Econosur’s Uruguay digital-infrastructure analysis therefore remains relevant: the country’s proposition combines renewable electricity, telecom infrastructure, cloud partnerships, legal stability and a small but concrete project pipeline.

Paraguay: hydroelectric abundance does not automatically equal AI capacity

Paraguay has perhaps the clearest low-cost energy proposition among the five countries. Itaipú and other hydroelectric resources have already attracted energy-intensive digital loads. The analytical challenge is to separate operating Bitcoin-mining infrastructure from dedicated AI and HPC capacity.

HIVE Digital Technologies reported 300 MW of operating digital infrastructure in Paraguay and has been expanding Yguazú toward a 400 MW power footprint. Its filings also make clear that the current large Paraguayan operating base is primarily associated with Bitcoin hashrate services.

The transition toward AI is becoming more concrete, but it remains a separate project path. HIVE reported in July 2026 that civil works were complete for a new 100 MW substation at Yguazú, with energization expected in September 2026. The company said construction of a new Tier III data center was expected to begin later in 2026, with ready-for-service targeted for the second half of 2027.

The Paraguay–Taiwan sovereign AI project is at an earlier stage. MITIC said in May that the first phase would target 10 MW of computing capacity under a 50/50 cooperation structure and estimated the first-phase investment at US$300–500 million. By 25 August 2026, MITIC still described the project as being in technical feasibility study.

This is why Econosur’s Paraguay power-grid and regulation analysis matters. Abundant generation does not eliminate questions about transmission, contract terms, allocation of electricity between domestic uses and electro-intensive industries, or whether the computing hardware and customer demand actually arrive.

The five countries are competing with different assets

The comparison shows why a simple ranking by electricity price would be misleading. The cheapest megawatt is not necessarily the most valuable megawatt. A hyperscaler or AI-infrastructure investor values electricity that can be contracted, connected, cooled and supported at the right site and timetable.

Brazil may therefore support higher-cost power than Paraguay but still win a project because it offers stronger enterprise demand, carrier networks, cloud ecosystems and a deeper construction base. Uruguay may be small but can gain credibility by moving a named project into construction. Argentina may offer extraordinary energy potential but still lose time if financing, fiber and transmission remain unresolved. Chile can have renewable supply and digital demand yet face local grid bottlenecks.

Where supplier and investment demand can emerge

If AI infrastructure continues to scale, South American opportunities will not be limited to data-center developers. Large projects create demand across energy, electrical equipment, construction, cooling, connectivity and operations.

The South America energy-infrastructure analysis is a direct neighbour to this market because data centers increasingly behave like major industrial loads. Their procurement needs resemble large industrial projects as much as conventional IT installations.

For suppliers, headline market growth is not enough. The commercial question is where a project has moved far enough to specify equipment, appoint engineering partners, secure electricity and begin procurement. A 500 MW announcement with no financing can generate less near-term supplier demand than a 30 MW first phase with a signed site, power agreement and construction schedule.

What would change the regional ranking?

The regional ranking is not fixed. A single signed hyperscale contract can move a market faster than years of policy announcements. New transmission can turn stranded renewable supply into a data-center location. A water restriction can force a redesign. A new submarine cable can change latency and redundancy. A financing shock can postpone projects that already have land and electricity.

The most useful forward indicators are therefore operational rather than promotional: signed power agreements, energised substations, permits, financing closes, construction starts, anchor leases, hardware procurement and commissioning dates.

Argentina’s next decisive signal would be a definitive contract and financed first phase for one of its large proposals. Brazil’s is the conversion of REDATA and grid-access demand into actual connected capacity. Chile’s is whether power-system expansion keeps pace with hyperscale demand. Uruguay’s is successful completion and operation of Google’s project plus development of the Antel AI layer. Paraguay’s is whether new Tier III and sovereign AI projects move from feasibility and substation works into operating AI/HPC service.

Conclusion

The central AI cost problem is not that every unit of intelligence is becoming more expensive. The opposite is happening in many benchmarks. Inference costs are falling rapidly. The problem is that lower unit costs can unlock far more usage while frontier models, cloud services and agentic workloads require increasingly dense and capital-intensive compute infrastructure.

That changes what a competitive AI location looks like. Electricity is central, but power price alone is insufficient. The viable product is a complete system that combines generation, transmission, substations, cooling, water strategy, fiber, permits, finance, operators and demand.

South America now offers several real experiments in that model. Brazil has scale and policy momentum. Chile has an operating digital market but visible grid and water constraints. Uruguay has moved a major Google project into construction. Argentina has large energy-based proposals but still needs to convert announcements into contracts and financed phases. Paraguay has proven its ability to host large digital loads but is only beginning the transition toward dedicated AI/HPC infrastructure.

The investment question is therefore not whether South America has cheap energy. It is which countries can turn energy into bankable, connected and operational compute.