HIVE Digital Technologies Ltd. (NASDAQ: HIVE) announced the completion of a research project using its GPU infrastructure in Asunción, Paraguay, conducted in collaboration with Columbia University's Department of Industrial Engineering and Operations Research.

The research has been submitted to the Conference on Neural Information Processing Systems (NeurIPS), an annual machine learning and computational neuroscience conference held each December.

The project established what HIVE describes as a proof of concept for intercontinental AI training, with researchers in New York City running training workloads on GPUs located in Asunción, more than 5,000 miles away. Using code optimizations developed by the Columbia team, the research found that HIVE's A40 GPUs matched the performance of H100 GPUs after normalizing for each hardware platform's raw performance, specifically for large language model pretraining of up to 1.4 billion parameters.

A researcher from Columbia's IEOR department stated: "In our use case of pretraining LLMs of up to 1.4B parameters, our results match those of H100s after normalizing for each hardware's raw performance."

HIVE said it will use the token-per-second, latency, and bandwidth data gathered from the project as a baseline for a planned HPC/AI facility in Yguazú, Paraguay, where the company has a 100-megawatt substation under construction. Civil works are complete, with substation energization expected in September 2026. Construction on a new Tier-III data center is slated to begin in fall 2026, with a ready-for-service date in the second half of 2027.

The research focuses on neural network pretraining using optimization theory, including analysis of the Muon optimizer and related scale-invariant methods. An assistant professor at Columbia's IEOR department said the work "advances our understanding of modern neural network optimization" and that the team looks forward to sharing findings with the broader research community.