HIVE Digital Technologies just wrapped up something that sounds like it shouldn’t work: running AI training workloads on GPUs sitting in Paraguay, controlled by researchers in New York, over 5,000 miles away.
The project, a collaboration with Columbia University’s Department of Industrial Engineering and Operations Research, used HIVE’s A40 GPUs located in Asunción to handle remote large language model pre-training and optimization workloads. It ran for approximately two months, and the company says the results were good enough to submit a research paper to NeurIPS, one of the most prestigious machine learning conferences in the world.
Older GPUs, newer tricks
HIVE claims that after optimization, its A40 GPUs delivered performance metrics that matched or exceeded those of Nvidia’s H100 GPUs when normalized for raw hardware capabilities. The A40 is a workstation-class GPU that’s considerably cheaper and older than the H100, which has been the gold standard for AI training since its launch.
That’s a meaningful finding. While 1.4 billion parameters is small compared to frontier models from OpenAI or Anthropic, it covers a huge swath of practical AI applications. Fine-tuning, domain-specific models, and research workloads all frequently operate at this scale.
