HIVE Digital Technologies’ Paraguay GPU cluster matched the performance of NVIDIA’s flagship H100 processors in a Columbia University study, validating the company’s push into AI infrastructure and establishing a proof of concept for intercontinental AI training. Researchers from Columbia’s Department of Industrial Engineering and Operations Research ran iterative training workloads on HIVE’s A40 GPUs in Asunción over a two-month period and submitted their findings to NeurIPS, one of the world’s three leading machine learning conferences.
The research focused on neural network pretraining algorithms, specifically testing optimization techniques for language models up to 1.4 billion parameters. According to the study, after the Columbia team optimized their code for HIVE’s A40 nodes and normalized performance for each hardware platform’s raw specifications, the results matched those achieved on H100 systems. The team also tested serving throughput and latency for deployed models, including standard benchmarks for LLaMA variants.
Frank Holmes, HIVE’s Executive Chairman, emphasized the geographic significance of the work. “Seeing professors from Columbia University in New York City remotely utilize our AI-optimized GPU cluster in Asunción, more than 5,000 miles away, demonstrates the power of distributed AI infrastructure,” he said. “It shows that high-performance computing does not need to be limited by geography.” This intercontinental arrangement underscores HIVE’s vision of bringing advanced AI computing to Paraguay, leveraging the country’s abundant hydroelectric power from the Itaipú Dam.
Aydin Kilic, HIVE’s President and CEO, described the result as validation of engineering excellence. “Over the past two months, the Columbia team optimized their code for HIVE’s A40 GPU nodes in Asunción across advanced AI workloads. In their specific use case, after normalizing for each hardware platform’s raw performance, the results on HIVE’s A40 nodes matched those observed on H100 systems. That is a powerful result,” Kilic stated.
The study’s acceptance at NeurIPS positions HIVE’s Paraguay infrastructure as a credible platform for AI research and production workloads. The company is now scaling this proof of concept through its planned expansion at the Yguazú site, where a 100 megawatt substation is under construction. Civil works are complete, with commissioning expected in September 2026. A new Tier-III data center is scheduled to begin construction in Fall 2026, with a ready-for-service date in the second half of 2027.
A researcher from Columbia’s Department of Industrial Engineering and Operations Research noted the broader implications. “We study neural network pretraining using optimization theory over general geometry and under large noise,” the researcher said. “By clarifying the theoretical foundations of modern optimizers such as Muon and evaluating them in practical neural network training settings, this research highlights their potential relevance for future LLM pretraining.” The work advances understanding of matrix-aware optimization methods that improve training efficiency and speed.
For HIVE, which began as a Bitcoin miner powered by green energy, the Colombia study represents a strategic shift toward diversified infrastructure. The company now operates data centers across Canada, Sweden, and Paraguay, serving both cryptocurrency mining and high-performance computing clients. The Paraguay GPU validation strengthens HIVE’s position in the growing market for cost-effective AI training infrastructure, where code optimization and hardware efficiency can rival the performance of newer, more expensive processors when properly deployed.
Sources
- TMX Newsfile / HIVE Digital Technologies Ltd. — Official press release on the Columbia University study, research submission to NeurIPS, performance validation, and Paraguay infrastructure timeline.
- Stock Titan — Reporting on HIVE’s A40 GPU performance matching H100 results after code optimization and normalization.
- The Cryptonomist — Confirmation that HIVE’s A40 GPUs in Paraguay matched H100 performance after Columbia’s code optimizations.
- StreetInsider — Reporting on the study’s findings and submission to NeurIPS conference.
