HIVE, a cryptocurrency mining and data center company, has wrapped up an artificial intelligence research project alongside Columbia University. The work was run on graphics processing units housed in a facility in Paraguay, and the results point to a cheaper way to train AI models — even with older hardware.
GPUs in Paraguay, researchers in New York
The project used GPUs that HIVE had already deployed in its Paraguay data center. Those chips weren't cutting-edge, but they got the job done. Columbia's team connected remotely to run the AI workloads, tapping into computing power located thousands of miles away.
The setup matters because it shows that geographic distance doesn't have to be a barrier for serious research. The university didn't need to buy or lease expensive local servers; instead, it used capacity that HIVE already had online in South America.
Older chips, lower costs
A key takeaway from the collaboration is that older GPU generations can still handle modern AI training tasks. Most of the hype around AI infrastructure focuses on the latest Nvidia H100 or B200 chips, but HIVE's Paraguay project deliberately used what many would consider last-generation hardware.
That choice kept costs down. Newer GPUs are expensive and hard to get. By relying on existing, older inventory, the project demonstrated a path for organizations that can't afford the latest gear — or don't want to wait for it.
Renewable energy and global teamwork
HIVE's Paraguay data center runs partly on renewable energy, though the company hasn't disclosed the exact mix. The project highlighted that pairing older GPUs with cleaner power can further reduce the environmental and financial cost of AI research.
The collaboration itself was international: a North American university and a Canada-based company working with hardware in South America. That kind of cross-border setup could become more common as companies and schools look for less expensive compute options.
The project is finished, and the results are in. For now, HIVE has shown that it's possible to do useful AI research without the newest chips or a data center in a major tech hub. Whether other firms will follow that model — and whether they can replicate it at scale — remains an open question.
