HIVE Digital Technologies Ltd. (TSX: HIVE) (NASDAQ: HIVE) said it has completed its inaugural research project using its HIVE GPUs for AI research purposes in Asunción, Paraguay, working with the Department of Industrial Engineering and Operations Research at Columbia University in New York. The company noted that the results have been submitted for publication at NeurIPS, one of the leading global machine learning and computational neuroscience conferences typically held in December, as HIVE seeks to demonstrate that high-performance AI training can be executed across continents using distributed compute.
At the center of the project is a proof of concept for intercontinental AI training: researchers in New York City reportedly ran iterative training runs on GPUs located in Asunción. HIVE said the research established a reference point for the performance of AI workloads using GPUs in Paraguay, leveraging code optimizations developed by the Columbia team. According to HIVE, the company’s A40 GPUs matched the performance of newer-generation H100 GPUs once measured results were normalized for hardware differences.
NeurIPS submission and intercontinental training test
HIVE said the research work—conducted in collaboration with Columbia University’s IEOR department—has been submitted to The Conference on Neural Information Processing Systems (NeurIPS). HIVE also described NeurIPS as one of three primary high-impact conferences in machine learning and artificial intelligence research globally, alongside ICLR and ICML.
Beyond simply demonstrating remote usage of compute, HIVE framed the project as performance validation for AI training workloads on its Paraguay-based GPU infrastructure. The company said that, using measured token-per-second, latency, and bandwidth data as a baseline, it has created a foundation for the performance benchmarking needed for future high-performance computing and AI deployment plans.
The collaboration focused on neural network pretraining using optimization theory, according to a summary provided by Columbia researchers. HIVE said Columbia’s approach involved designing and analyzing an accelerated algorithm that aims to match the performance of Muon, a method described by the company as a leading current technique, both in theory and in practice.
“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.”
Performance parity on A40 versus H100, including LLaMA throughput tests
HIVE said the study’s findings support performance parity between its A40 GPUs in Asunción and H100 systems when the comparison accounts for each platform’s raw capabilities. In particular, Columbia researchers reported optimization of their code for HIVE’s A40 GPU nodes over the prior two months and said throughput and latency testing was carried out for both Muon and variants tied to the project’s optimization work.
HIVE said the experiments included an LLM pretraining scenario involving models “up to 1.4B parameters,” with results described as matching H100 observations after normalization. The research also extended to serving behavior: the company said Columbia tested the serving throughput and latency of the 1.4B model, noting that this model “has not yet been heavily optimized.” Standard throughput and latency tests for LLaMA models were also included, according to the description provided in the release.
HIVE further highlighted the role of software optimization in the outcome. It said code optimizations developed by the Columbia team were used to run the iterative training runs on GPUs located in Asunción. The company indicated that these improvements served as the mechanism to reach the performance levels observed on newer-generation hardware after accounting for baseline differences.
Yguazú plans: 100MW substation, Tier-III data center timeline
With the performance measurements serving as a baseline, HIVE said it has established groundwork for an “HPC/AI Gigafactory” in Yguazú, Paraguay. The project is tied to the company’s ongoing infrastructure build-out, including a 100 megawatt (“MW”) substation under construction.
HIVE said civil works for the substation are complete, with commissioning expected “this summer.” The release stated that the substation is expected to be energized in September 2026. It also outlined further expansion plans: construction on a new Tier-III data center in Yguazú is expected to begin in Fall 2026, with a ready-for-service date in H2 2027.
For HIVE, the company’s stated objective is to use the research results as a reference point for AI workload performance on its Paraguay-based GPUs, while tying those benchmarks to the scaling requirements of future high-performance computing infrastructure. The company pointed to the remote collaboration itself as an operational demonstration, suggesting that geography need not be a barrier for distributed AI training runs.
“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. It shows that high-performance computing does not need to be limited by geography.”
Executives emphasize distributed computing and commercialization validation
In leadership commentary, Frank Holmes, executive chairman of HIVE, described the project as a milestone toward the company’s goal of supporting advanced AI computing infrastructure in Paraguay. Holmes said the ability for Columbia researchers to remotely use HIVE’s GPU cluster in Asunción—while conducting iterative training—illustrated the operational potential of distributed AI infrastructure.
Holmes also linked the milestone to HIVE’s broader commercial ambitions, saying it validates the commercial potential of HIVE’s AI platform and its ability to deliver compute across borders. He additionally characterized the event as meaningful for Paraguay, citing the company’s view that the country has power, strategic location, and now a proof point for its participation in the global AI economy.
Aydin Kilic, HIVE’s president and CEO, said the collaboration reflects what he described as the company’s focus on engineering-driven innovation. Kilic noted that Columbia’s research is centered on next-generation neural network pretraining and described it, in practical terms, as work intended to make AI training “smarter, faster, and more efficient” by improving the mathematical engine behind neural network learning.
Kilic said the release’s results also align with his stated belief that strong engineering can unlock value. He pointed to HIVE’s history of building the BuzzMiner in collaboration with Intel Corporation and becoming one of the largest demand-response participants in Sweden—examples included in the release as evidence of prior technology and infrastructure development efforts. The company did not provide additional detail beyond those references, but said it plans to continue investing in communities while advancing its global data center technology stack.
The release also included comments from an assistant professor in Columbia’s Department of Industrial Engineering and Operations Research, who said the work advances understanding of modern neural network optimization, including matrix-aware optimizers such as Muon and related scale-invariant methods. The professor said the research clarified theoretical foundations and evaluated performance in practical neural network training settings, and added that the team looks forward to sharing the findings with the wider research community.
HIVE’s announcement comes with reference to the original report shared via Newsfile.
Background on HIVE Digital Technologies
Founded in 2017, HIVE Digital Technologies Ltd. says it was among the first publicly listed companies to focus on mining digital assets powered by green energy. The company builds and operates next-generation Tier-I and Tier-III data centers across Canada, Sweden, and Paraguay, serving both Bitcoin and high-performance computing clients. HIVE describes its platform as combining an infrastructure-driven dual-engine hash rate service with GPU-accelerated AI computing aimed at delivering scalable, environmentally responsible solutions for the digital economy.
