$HIVEJun 22, 2026, 1:31 PMAI processed
HIVE's Paraguay A40 GPUs matched H100s in Columbia Univ study; paper sent to NeurIPS
AI Summary
HIVE reports Columbia University's IEOR team remotely ran iterative pretraining on HIVE A40 GPU nodes in Asunción and — after code optimizations and normalizing for raw hardware performance — found A40s matched H100 performance; the research has been submitted to NeurIPS. The company says this validates intercontinental AI training, provides baseline token/sec/latency metrics for its platform, and supports its Yguazú build-out (100 MW substation civil works complete, energization expected Sep 2026 and a Tier‑III data center targeted H2 2027).
Positives
- Columbia Univ study found HIVE A40 nodes matched H100 performance after optimizations and normalization
- Research submitted to NeurIPS, a top-tier machine-learning conference (adds external credibility)
- HIVE has civil works complete for 100 MW Yguazú substation and published energization / data center timelines
Negatives
- Performance claim depends on code optimizations and normalization for hardware raw performance
- Study used pretraining on models up to 1.4B parameters — not necessarily representative of very large LLM training
- Energization and datacenter readiness are forward-looking projections and subject to execution risk