HIVE validates Paraguay AI infrastructure via Columbia study

HIVE Digital Technologies Ltd. has announced the successful completion of its inaugural AI research project, conducted in collaboration with the Department of Industrial Engineering and Operations Research at Columbia University in New York. The project utilized HIVE GPUs located in Asunción, Paraguay, and the resulting research has been submitted to The Conference on Neural Information Processing Systems (NeurIPS), one of the three primary high-impact conferences in machine learning and artificial intelligence globally, held annually in December.

Intercontinental AI Training: A Proof of Concept

The research establishes a proof of concept for intercontinental AI training, with researchers based in New York City successfully executing iterative training runs on GPUs physically located in Asunción, Paraguay — more than 5,000 miles away. A key finding of the study was that, following code optimizations developed by the Columbia team, HIVE's A40 GPUs matched the performance of newer-generation H100 GPUs after normalizing for each hardware platform's raw performance.

The Columbia research team focused on neural network pretraining using optimization theory over general geometry and under large noise, designing and analyzing an accelerated algorithm that matches the performance of the current leading method, Muon, in both theory and practice. The team tested their approach on large language model (LLM) pretraining workloads of up to 1.4B parameters.

Key aspects of the research and performance benchmarking are summarized below:

Parameter:	Details
GPU Hardware Used:	A40 GPUs (HIVE, Asunción, Paraguay)
Benchmark Comparison:	H100 GPUs
LLM Pretraining Scale:	Up to 1.4B parameters
Optimization Method Evaluated:	Muon and related scale-invariant variants
Additional Tests:	Serving throughput and latency for 1.4B model; LLaMA model throughput and latency
Research Submission:	NeurIPS (annual conference, held in December)
Collaboration Partner:	Department of IEOR, Columbia University, New York

Foundation for Paraguay HPC/AI Gigafactory

Using the measured token-per-second, latency, and bandwidth data as a baseline, HIVE has established a foundation for an HPC/AI Gigafactory in Yguazú, Paraguay. The Company has a 100 megawatt (MW) substation under construction at the site, with civil works already complete.

The key infrastructure milestones for the Yguazú development are outlined below:

Milestone:	Timeline
Civil Works:	Complete
Substation Commissioning:	Expected this summer
Substation Energization:	September 2026
Tier-III Data Center Construction Start:	Fall 2026
Ready-for-Service Date:	H2 2027

Leadership and Academic Commentary

Frank Holmes, Executive Chairman of HIVE, described the milestone as an important step in the Company's mission to bring advanced AI computing infrastructure to Paraguay. He highlighted that the collaboration demonstrates high-performance computing need not be limited by geography, and noted that Paraguay possesses the power, strategic location, and now a validated proof point for participation in the global AI economy.

Aydin Kilic, President and CEO of HIVE, emphasized the significance of the engineering outcome, stating that the results on HIVE's A40 nodes — after normalizing for each hardware platform's raw performance — matched those observed on H100 systems. He also referenced HIVE's broader engineering history, including building the BuzzMiner in collaboration with Intel Corporation and becoming one of the largest demand-response participants in Sweden.

An Assistant Professor in the Department of Industrial Engineering and Operations Research at Columbia University noted that the work advances understanding of modern neural network optimization, including matrix-aware optimizers such as Muon and related scale-invariant methods, and highlighted the potential relevance of these findings for future LLM pretraining.

HIVE Digital Technologies Ltd. announced plans to acquire the 32-megawatt Big Boden data center in Boden, Sweden, after receiving approval from the Boden Municipal Council. This transition from tenant to owner secures the company's authority over the site's future and supports its strategy to build sovereign AI compute infrastructure. The stock climbed nearly 10% in premarket trading following the announcement.

Upon closing the transaction, HIVE intends to advance the facility toward Tier III infrastructure standards. These benchmarks for security, redundancy, and uptime are required to support enterprise-scale AI and high-performance computing workloads, including NVIDIA's latest GPU architectures. The acquisition remains subject to the completion of customary closing conditions.

Strategic Partnerships

Separately, BCE Inc., Cohere, Hypertec, and BUZZ High Performance Computing (HPC), a subsidiary of HIVE, announced a partnership to expand sovereign AI infrastructure in Canada. Bell will provide data center capacity from its Merritt, British Columbia, facility, while BUZZ HPC will deliver AI cloud infrastructure using Hypertec's Canadian-built hardware and NVIDIA's DSX AI factory platform. Cohere will deploy its enterprise AI models on the platform to support secure, production-scale AI workloads for government and enterprise customers.

Financial and Operational Impact

HIVE has invested more than 960 million SEK (approximately $100 million USD) in the Boden region over the past eight years through local contractors and renewable energy procurement. The company has contributed over 575 million SEK (more than $60 million USD) in taxes to the Swedish Tax Authority during this period.

Financial Metric	Amount
Local Investment	> 960 million SEK
Taxes Paid	> 575 million SEK
Data Center Capacity	32 MW

Market Outlook

Wall Street expects HIVE to report fiscal first-quarter results around Aug. 13, 2026. Analysts anticipate a loss of 19 cents per share on revenue of $82.53 million, compared with revenue of $45.61 million in the year-ago quarter. The stock carries a consensus Buy rating with an average price forecast of $6.68.