HIVE Digital Technologies just demonstrated a potential challenge to Nvidia's pricing dominance. The Bitcoin mining company trained large language models on its older Nvidia A40 GPUs in Paraguay and achieved performance matching Nvidia's flagship H100 chips. The market responded, with HIVE's stock climbing as much as 22%.
HIVE partnered with Columbia University's Department of Industrial Engineering and Operations Research for the project. Researchers in New York remotely operated HIVE's GPU infrastructure in Asunción, Paraguay, roughly 5,000 miles away. Over two months, they trained models with up to 1.4 billion parameters on the A40 GPUs. After optimization, the older A40 systems matched the H100's performance for these workloads. The team submitted the findings to NeurIPS, a top machine learning conference.
There is a significant caveat. The performance parity was demonstrated at a scale of 1.4 billion parameters. Frontier models today use hundreds of billions or trillions of parameters. The A40s will likely not match H100 performance at that massive scale. The project does, however, prove the viability of distributed, intercontinental AI training workflows using the cheapest available compute.
This work supports HIVE's strategic shift from pure crypto mining toward high-performance computing and AI data centers. Its Paraguay facility is central to this plan, offering abundant, cheap, renewable hydroelectric power. The successful demonstration provides concrete performance data for HIVE to pitch future AI compute deals.