Mercado Libre is Latin America’s answer to Amazon and PayPal rolled into one—120,000 employees, 18 countries, and the sixth-largest e-commerce platform in the world by gross merchandise volume. When a company moves at that scale, search isn’t just a feature, but a critical component of how the company interacts with its core customers.
That’s why they wanted to rebuild their search engine from the ground up, using AI to stop guessing what customers want and start knowing. But moving responsibly meant starting small, validating before committing, and proving ROI before scaling.
- Access to GPUs in constrained Virginia region not available from other providers
- Observability and support infrastructure to responsibly scale an agentic search engine across 18 countries
- Full-featured POC access with no gated capabilities to evaluate real operational fit before committing
Solution
CoreWeave gave Mercado Libre something most providers couldn’t—immediate access to real GPUs, including NVIDIA B200s and R6000s. They also had full platform access from day one. No sandboxed environment, no feature tiers, no gap between what they evaluated and what they would run in production.
- CoreWeave ARENA
Mercado Libre ran a full three-week POC with unrestricted access to NVIDIA B200 and R6000 GPUs in CoreWeave ARENA, a POC environment. No other provider offered this level of openness at the evaluation stage. - CoreWeave Kubernetes Service (CKS)
The team migrated inference workloads for query expansion onto CoreWeave’s Kubernetes offering with zero disruption. Workloads that previously ran on other stacks transitioned seamlessly, as if they had never switched providers. - SUNK (Slurm on Kubernetes)
SUNK gave Mercado Libre a one-click Slurm cluster running alongside Kubernetes. This enabled GPU time-slicing and workload prioritization without the weeks-long overhead of standing up an independent Slurm environment. - Bare metal infrastructure
Running on bare metal gave Mercado Libre the raw performance and control needed for latencysensitive, real-time inference workloads—critical when query expansion must respond to live user searches at massive scale. - Transparent pricing and 24/7 support
Clear, predictable pricing was a deciding factor before the POC even began. Once in production, CoreWeave’s team delivered direct, always-on support through a dedicated Slack channel, resolving issues within hours.
Outcomes
The team went from POC to production in three weeks. Mercado Libre came in looking for a provider that could deliver at their scale. What they left with was something harder to find—a platform that performed exactly as advertised and a partner invested in their success at every step.
Instant capacity, zero compromise
When Virginia’s GPU market was locked up, CoreWeave made compute available immediately. This enabled Mercado Libre to run a full three-week POC with no gated features, giving the team a true preview of production operations before any long-term commitment.
Inference that just works
Migrating query-expansion workloads onto CKS was seamless. Workloads that previously ran on other stacks transitioned without disruption, letting Mercado Libre focus on building Search 2.0 rather than managing infrastructure.
A partner that moves with you
From SUNK’s one-click Slurm clusters to 24/7 Slack support, CoreWeave gave Mercado Libre the tooling and expertise to scale responsibly. Any issues were resolved within hours, with full transparency and documentation every step of the way.
