HIVE Digital Technologies Ltd. (TSX: HIVE) (NASDAQ: HIVE) said it has completed its first research project using its HIVE GPUs for AI research, demonstrating that iterative training runs can be executed across continents—running in Asunción, Paraguay while researchers at Columbia University in New York remotely tested and optimized the work. The company said the research has been submitted to NeurIPS, with additional benchmarking finding that HIVE’s A40 GPUs matched the performance of newer-generation H100 GPUs after performance normalization.
The milestone positions the company’s hardware and software stack as a platform for “intercontinental AI training,” and it supports HIVE’s plans for an HPC/AI “Gigafactory” in Yguazú, Paraguay. HIVE also provided a timeline for infrastructure progress at the Yguazú site, including the construction status of a 100 megawatt (MW) substation and the expected next steps for a new Tier-III data center.
Columbia University research submitted to NeurIPS
HIVE’s inaugural research effort was carried out in Asunción, Paraguay in collaboration with the Department of Industrial Engineering and Operations Research at Columbia University in New York. The company said the project’s findings were submitted to The Conference on Neural Information Processing Systems (NeurIPS), held annually in December. HIVE said NeurIPS—along with ICLR and ICML—forms one of the three primary high-impact conferences for machine learning and artificial intelligence research.
According to HIVE, the experiment served as a proof of concept for distributed training, with researchers in New York City successfully running iterative training runs on GPUs located in Paraguay. With that testing, the company said it established a reference point for the performance of AI workloads using GPUs in Asunción.
HIVE reported that, using code optimization developed by the Columbia team, its A40 GPUs achieved performance comparable to H100 GPUs. The company characterized the result as an important validation of how its optimized environment performs under AI training workloads, while emphasizing that normalization was used to compare raw hardware performance across platforms.
“Over the past two months, we optimized our code for the A40s and tested the throughput and latency of Muon and our variants. In our use case of pretraining LLMs of up to 1.4B parameters, our results match those of H100s after normalizing for each hardware’s raw performance.”
The research approach, as described by a Columbia researcher, focused on neural network pretraining using optimization theory over general geometry under large noise. HIVE said the work included designing and analyzing an accelerated algorithm that matches the performance of a leading method called Muon in both theory and practice. The company further stated that the study examined performance for pretraining algorithms and evaluated throughput and latency characteristics for the GPU nodes used in Asunción.
Benchmarks across throughput, latency, and bandwidth
HIVE said the Columbia team carried out additional measurements beyond training pretraining runs. The company reported that the research team tested serving throughput and latency for a 1.4B parameter model, noting that the model had not yet been heavily optimized. HIVE also said the researchers included standard throughput and latency testing for LLaMA models.
HIVE stated that its analysis used measured token-per-second, latency, and bandwidth data as a baseline. The company said those results now provide a foundation for its larger infrastructure planning in Paraguay, linking the research performance data to the next stages of building an HPC/AI deployment in Yguazú.
Yguazú substation and Tier-III data center timeline
Following the performance benchmarking from the Asunción research environment, HIVE said it has “established a foundation” for an HPC/AI Gigafactory in Yguazú, Paraguay. The company described the Yguazú project as anchored by a 100 MW substation currently under construction.
HIVE said civil works at the substation are complete, with commissioning expected this summer. The company added that the substation is expected to be energized in September 2026.
Looking further out, HIVE said construction on a new Tier-III data center in Yguazú is scheduled to begin in Fall 2026, with the facility expected to be ready for service in the second half of 2027.
“Seeing professors from Columbia University in New York City remotely utilize our AI-optimized GPU cluster in Asunción, more than 5,000 miles away, demonstrates the power of distributed AI infrastructure.”
HIVE’s executive chairman Frank Holmes framed the outcome as a demonstration that high-performance computing need not be constrained by geography, pointing to what he described as the role of advanced power supply, data center design, software stack, and execution in enabling Paraguay’s participation in AI computing. HIVE’s president and CEO Aydin Kilic also said the project validates engineering efforts and the company’s belief that performance can be unlocked through code and optimization work tailored to its GPU nodes.
HIVE added a note from another Columbia contributor expressing that the research advances understanding of modern neural network optimization, including matrix-aware optimizers such as Muon and related scale-invariant methods. The contributor said the work clarified theoretical foundations and evaluated them in practical neural network training settings, and they said they look forward to sharing the findings with the broader research community.
What the company says the research means for its platform
In its own commentary, HIVE emphasized that the project was conducted using its GPU cluster in Asunción and involved optimization across advanced AI workloads. The company said the Columbia team optimized code for HIVE’s A40 GPU nodes over the past two months, and in HIVE’s described use case—pretraining large language models of up to 1.4B parameters—results matched those observed on H100 systems after normalizing for hardware raw performance.
HIVE connected the milestone to its broader operating footprint, stating that it builds and operates Tier-I and Tier-III data centers across Canada, Sweden, and Paraguay, serving both Bitcoin and high-performance computing customers. The company said its dual-engine infrastructure is driven by “hashrate services” alongside GPU-accelerated AI computing.
Founded in 2017, HIVE Digital Technologies Ltd. said it was among the first publicly listed companies to prioritize mining digital assets powered by green energy. For more information, visit hivedigitaltech.com and for context on the announcement see the company’s source.
The research submission to NeurIPS places HIVE’s performance validation in the public research pipeline ahead of the conference in December, while the company’s Yguazú build-out timeline sets out a path from benchmarking in Asunción toward expansion of its HPC/AI infrastructure in Paraguay.
