Overview
- Columbia University researchers reported on Monday that optimized Nvidia A40 GPUs in HIVE’s Paraguay cluster delivered throughput and latency similar to H100 systems for pretraining models up to about 1.4 billion parameters, and the team submitted the work to NeurIPS.
- HIVE’s shares rose more than 22% after the university’s validation was reported, with company leaders characterizing the work as support for HIVE’s shift from crypto mining toward AI infrastructure.
- The research team spent roughly two months tuning code for A40 hardware and measured throughput, latency, and bandwidth to compare results after normalizing for raw hardware differences.
- Tests were run remotely from New York on GPUs physically located in Asunción, Paraguay, a distance of over 5,000 miles, demonstrating cross‑continent control of iterative training jobs.
- Key caveats remain: the parity was shown only at modest model scale, external peer review is pending, and HIVE’s planned capacity expansions — a 100 MW substation targeted for energization in September 2026 and a Tier‑III data center in the second half of 2027 — are still upcoming and crucial to realizing broader impact.