Norway built an LLM training cluster on 2 PB of Huawei flash

The Norwegian University of Science and Technology deployed 2 petabytes of Huawei OceanStor Dorado flash arrays to back their AI training infrastructure. Not Dell, not Pure, not NetApp. Huawei. The storage feeds 288 Nvidia H100 GPUs running distributed LLM training jobs. The setup achieves 140 GB/s read throughput across the cluster. That is the kind of bandwidth you need when you are checkpoint-saving mid-training or shuffling massive tokenised datasets between epochs. What caught my attention is the flash-to-GPU ratio. Two petabytes for 288 GPUs works out to roughly 7 TB per GPU. Most enterprise setups I have seen budget closer to 2 to 4 TB per GPU for training clusters, relying on network-attached object storage for the bulk dataset and only staging working sets locally. Norway went the other direction: over-provision the flash, keep latency predictable, avoid the object storage tax during the training loop. The Huawei angle is political, not technical. Norway is not under the same export restrictions as the US or UK. They can buy Chinese flash arrays without needing State Department approval. The Dorado series uses the same NVMe SSDs as everyone else. The controller firmware is proprietary, but the raw performance numbers line up with western equivalents at the same price point. I wonder if this becomes a trend. If you are a European research institution building out GPU clusters for open model training, Huawei flash is 20 to 30 percent cheaper than Pure or NetApp for comparable IOPS. The trade-off is vendor lock-in to a Chinese supply chain, which matters more to some buyers than others. The throughput number is the real validation. 140 GB/s across 288 GPUs means the storage is not the bottleneck. You can train without waiting on checkpoint writes. That is what matters.

---

Source: Norway’s 2 petabytes of Huawei flash storage and LLM training