Nvidia chips in the spotlight as HIVE reports Paraguay GPU cluster matches H100 performance for AI pretraining

Nvidia’s data-center GPUs were back in focus after a report said HIVE, a company operating computing infrastructure, achieved performance in AI research using its Nvidia A40 GPUs in Paraguay that matched performance researchers had observed on Nvidia H100 systems. The account, published by Yahoo Finance and carried by CCN, framed the result as a meaningful benchmark for large-language-model (LLM) pretraining work, which is the compute-heavy phase where models learn from large volumes of text before fine-tuning.

According to the report, the Paraguay-based GPU setup used Nvidia A40 accelerators and delivered performance that researchers associated with H100 systems. While the article did not spell out the specific training settings, model sizes, or evaluation methodology in the information provided here, it did link the comparison to LLM pretraining research output, suggesting the performance parity claim was grounded in a testing context rather than marketing language.

The report also connected the announcement to market reaction for HIVE, describing a rise in the company’s share price following the research-related performance claim. For Nvidia, the practical takeaway is not that the company has changed its products or roadmaps, but that an installed base of earlier-generation accelerators can still be deployed to produce competitive AI research results when configured and operated effectively.

Nvidia’s A40 is a data-center GPU designed for enterprise and cloud workloads, including AI training and inference. In contrast, Nvidia’s H100 is part of the company’s newer data-center generation, marketed for high-performance AI workloads, including large-scale model training. The report’s core message, as described in the available text, is that HIVE’s A40 deployments in Paraguay were able to track performance expectations associated with H100 systems for a specific pretraining use case.

This kind of result matters to AI infrastructure operators because the economics of building and running training environments are shaped by more than peak specs. Data center power, cooling, chip utilization, scheduling, and the ability to iterate on training runs can all influence outcomes. Even when newer hardware has advantages on paper, operators may achieve practical performance targets through careful system design and workload optimization, depending on what is being measured.

There is also a broader market implication. Nvidia’s ecosystem is tightly coupled to how customers deploy GPUs, whether in owned data centers, colocation facilities, or hybrid models. When reports circulate about performance parity across different GPU generations, it can reinforce confidence among buyers that they can meet research and deployment needs without always starting with the newest hardware, particularly if their primary objective is to advance experimentation and pretraining throughput.

Still, important details were not included in the information available for this write-up. The report did not provide the precise benchmark numbers, the LLM pretraining configuration, the batch sizes, the number of GPUs used in the Paraguay cluster, or whether the comparison to H100 was done under identical conditions. Without those specifics, it is difficult to translate the claim into a general performance expectation for all AI workloads.

What to watch next is whether HIVE and its researchers publish more technical documentation around the pretraining tests, including the measurable outcomes and how “matching” was determined. Nvidia will also remain a central point of reference for the industry, since the credibility of these kinds of comparisons can affect how operators evaluate the trade-offs between older accelerators such as the A40 and newer systems such as the H100 for future training and research cycles.