Hive stock jumps 25% after Columbia trains AI on Paraguay GPUs

Hive Digital shares climbed about 25% on Monday after Columbia University researchers ran neural network training jobs on the company's GPU cluster in Asunción, Paraguay and reported that Nvidia A40 cards delivered performance comparable to newer systems on certain workloads. The research focuses on methods to improve the efficiency of large language models and has been submitted for review at the NeurIPS conference in December.

Researchers based in New York executed training jobs on Hive's Paraguay infrastructure to measure latency, throughput and distributed training performance for large models. The company reported the tests assessed practical training metrics rather than theoretical benchmarks and that results supported intercontinental training over long distances for specific tasks.

Hive opened the Paraguay cluster in March as an early testbed for running AI training jobs across continents. Executive Chairman Frank Holmes described the Columbia results as a ‘proof of concept' for intercontinental AI training and highlighted the company's capacity to host remote, production-scale machine learning workloads from outside Paraguay.

The Paraguay facility runs on hydro-powered infrastructure that currently totals about 300 megawatts, with another 100 megawatts under development. Hive has also outlined plans for a 320-megawatt AI facility in the Toronto area that the company says could eventually support more than 100,000 GPUs.

The Columbia study and its planned publication come days after Hive announced a three-year, $220 million AI cloud contract with Bell Canada and Cohere. That agreement, which involves more than 2,300 Nvidia Grace Blackwell GPUs, is expected by the company to add roughly $70 million in annual recurring revenue later this year.

Hive's stock reached about $5.20, its highest level in over seven months, following the release of the research results and the recent contract announcement. The NeurIPS submission will make the researchers' methodology and detailed results available to the academic community if and when it is accepted and published.