Latin America GPU Market: Brazil, Chile, and Mexico Cloud Infrastructure

Latin America's GPU infrastructure market is concentrated in three primary hubs: Sao Paulo, Brazil; Santiago, Chile; and Queretaro, Mexico. These three markets account for approximately 70 percent of the region's 450 MW of total data center capacity and an estimated 8,000-12,000 H100-equivalent GPUs. The region is growing at 20-25 percent annually, driven by nearshoring from US-based AI companies seeking lower costs, regional language AI models for Portuguese and Spanish, and fintech AI workloads serving Latin America's rapidly digitizing financial sector.

The defining characteristic of Latin American GPU infrastructure is the combination of US-tier network latency (60-120ms to US East Coast) with emerging-market power reliability. Brazil's grid has an average reliability of 99.7 percent versus 99.99 percent in US data center markets, requiring UPS and generator configurations that add 10-15 percent to facility costs. Power costs range from $0.08-0.12 per kWh in Chile (solar-backed) to $0.14-0.20 in Brazil and $0.10-0.16 in Mexico. The region's GPU pricing reflects these infrastructure costs, with H100 reserved rates of $3.00-$5.00 per GPU-hour, roughly 20-30 percent above US pricing.

Sao Paulo concentrates over half of Latin America's GPU infrastructure, anchored by the three major cloud providers. AWS operates four availability zones in the sa-east-1 region with H100 and Inferentia2 instances. Azure's Brazil South region serves GPU capacity from two data centers in Campinas and Sao Paulo. GCP's southamerica-east1 runs TPU v5p and H100 in Sumare, 90km from Sao Paulo. Ascenty, the largest independent data center operator in Latin America (acquired by Digital Realty), operates 18 facilities across Brazil including its SGR8 campus in Jundiai, which hosts a 2,000-GPU H100 cluster dedicated to AI workloads.

Brazilian GPU pricing is the highest in Latin America at $3.50-$5.00 per GPU-hour for H100 reserved, driven by high import duties on GPU hardware (approximately 35 percent combined duty and taxes for IT equipment), Brazil's complex tax structure (ICMS tax varying by state from 7-18 percent), and power costs of $0.14-0.20 per kWh. The import duty structure has created a secondary market where hardware procurement costs 40-50 percent more than US prices. GPU providers like Ascenty, ODATA, and Scala Data Centers mitigate this by importing containers of GPUs under special customs regimes, achieving effective duty rates of 12-15 percent, but this advantage is not fully passed through to end-user pricing.

Chile presents Latin America's most interesting GPU infrastructure opportunity due to its solar energy advantage. The Atacama Desert receives the highest solar irradiance of any inhabited region on Earth at 2,500-3,000 kWh/m2 per year, producing solar electricity at $0.02-0.04 per kWh during peak generation hours. Data centers in Santiago and the Alto Hospicio region are increasingly pairing solar PPAs with grid supply to achieve blended power costs of $0.08-0.10 per kWh, competitive with the Nordics. The Santiago hub hosts approximately 2,000-3,000 H100 GPUs across facilities operated by Equinix (SC1/SC2), EdgeConnex, and HostDime.

Chile's political stability, strong rule of law, and active free trade agreements make it the preferred Latin American market for international GPU providers. The Chilean government's Digital Transformation Ministry offers a Fast Track for data center construction permits, reducing build time from 24 to 12 months. Claro, Chile's data center, completed a 10 MW GPU wing in 2025 hosting 1,500 H100 GPUs. The main limitation is connectivity: latency from Santiago to US West Coast is 130-160ms, and to US East Coast 160-190ms, which restricts Chile's role to training and batch inference workloads rather than real-time serving for North American users.

Mexico's GPU infrastructure is concentrated in the Queretaro data center corridor, 200km northwest of Mexico City, which hosts 70 percent of the country's data center capacity. Queretaro offers 40-50ms latency to Dallas and 50-60ms to the US East Coast, making it the lowest-latency Latin American location for US-facing AI workloads. KIO Networks operates a 30 MW facility with H100 and L40S GPUs serving Mexican fintech and US nearshore clients. Equinix's MX1/MX2 in Queretaro host AWS, Azure, and GCP edge zones plus a 1,500-GPU cluster from a major US GPU provider that uses Mexican capacity for overflow training.

Mexico's GPU market benefits from the USMCA trade framework, which allows duty-free import of GPU hardware manufactured in the US, reducing hardware procurement costs by 25-35 percent versus Brazil. Industrial power in Queretaro averages $0.10-0.16 per kWh, with natural gas-fired plants providing baseload and wind power from Oaxaca providing renewable PPAs. The main constraint is water availability for cooling. Queretaro is a semi-arid region, and new data center builds are required to use closed-loop liquid cooling or dry coolers with zero water consumption. This has increased GPU cluster deployment costs by 10-15 percent compared to traditional evaporative cooling designs.

Latin American GPU infrastructure is fundamentally shaped by submarine cable connectivity. The region's external capacity is dominated by the AMX-1, Tannat, Monet, and Seabras-1 cables connecting Brazil to the US and Europe, the Curie cable connecting Chile and Panama to Los Angeles, and the ARBR cable from Argentina to Brazil. Total international capacity for Latin America is approximately 120 Tbps as of 2026, versus 800+ Tbps for transatlantic routes connecting Europe and North America. The bandwidth constraint means that moving petabytes of training data between Latin America and North America requires 2-5 days versus hours for intra-North America transfers.

The EllaLink cable, operational since 2022, connects Brazil directly to Europe (Fortaleza to Sines, Portugal) with 72 Tbps capacity and 56ms latency. This is the lowest-latency direct route between South America and Europe and positions Brazilian GPU infrastructure in Fortaleza as a potential near-Europe training hub. Fortaleza-based GPU clusters offer 56ms to London and 65ms to Frankfurt, compared to 110-130ms via US routing. Scala Data Centers' Fortaleza facility has capitalized on this, building a 10 MW GPU wing for European AI clients seeking lower power costs with competitive latency to European end users.