A Bitcoin mining company just pulled off something that should make Nvidia’s pricing team slightly uncomfortable. HIVE Digital Technologies trained large language models on older A40 GPUs in Paraguay and achieved performance that matched Nvidia’s flagship H100 chips, the same hardware that companies are currently paying a premium to get their hands on.
The trick wasn’t better hardware. It was smarter software. And the market noticed, with HIVE’s stock climbing as much as 22% on the news.
What HIVE actually proved
HIVE partnered with Columbia University’s Department of Industrial Engineering and Operations Research to run its first external academic AI project. The setup: Columbia researchers in New York remotely operated HIVE’s GPU infrastructure in Asunción, Paraguay, roughly 5,000 miles away.
Over approximately two months, the team trained large language models with up to 1.4 billion parameters on HIVE’s Nvidia A40 GPUs. After optimization and normalization, the A40 systems matched the performance of Nvidia’s far more expensive H100 chips for those workloads.
