Technical Deep Dive
The storage requirements for training large language models have evolved from a secondary concern to a primary bottleneck. When training a model like GPT-4-class (estimated 1.8 trillion parameters), the storage subsystem must handle three critical workloads simultaneously:
1. Checkpointing: Every few hours, the entire model state (parameters, optimizer states, gradients) must be written to persistent storage. For a trillion-parameter model, this is 2-4 TB of data per checkpoint. The write must complete within minutes to avoid stalling the training pipeline. Latency spikes during checkpointing can cause cascading failures in distributed training.
2. Data Loading: Training data must be streamed to GPUs at rates exceeding 100 GB/s. Unlike traditional HPC workloads, LLM training uses random access patterns across massive datasets (often 10-100 TB). This requires high IOPS and low queue depth latency.
3. Gradient Synchronization: In data-parallel training, gradients are exchanged between nodes. While this is primarily a network concern, the storage system often acts as a synchronization barrier, especially in asynchronous training setups.
Huawei's OceanStor Dorado all-flash arrays deployed in Norway use a custom ASIC-based controller architecture that provides deterministic sub-millisecond latency regardless of load. This is achieved through a proprietary NVMe-over-Fabric implementation that bypasses traditional TCP/IP stacks. The 2PB configuration likely uses 30TB NVMe SSDs in a 3:1 capacity efficiency ratio, yielding approximately 6PB raw capacity.
Performance Comparison:
| Storage Solution | Latency (99th percentile) | Sustained Write Throughput | Max IOPS | Power per TB |
|---|---|---|---|---|
| Huawei OceanStor Dorado 8000 | 0.5ms | 40 GB/s | 10M | 0.8W/TB |
| Dell PowerMax | 0.8ms | 25 GB/s | 7M | 1.2W/TB |
| Pure Storage FlashArray//X | 0.7ms | 30 GB/s | 8M | 1.0W/TB |
| NetApp AFF A-Series | 1.0ms | 20 GB/s | 5M | 1.1W/TB |
Data Takeaway: Huawei's latency advantage of 0.5ms vs. 0.7-1.0ms for Western competitors may seem marginal, but in distributed training, every millisecond of latency multiplies across thousands of GPUs. A 0.2ms improvement in checkpoint write latency can reduce total training time by 3-5% for a 30-day training run, translating to hundreds of thousands of dollars in GPU compute savings.
The open-source community has also taken notice. The GitHub repository [nvme-cli](https://github.com/linux-nvme/nvme-cli) (12k+ stars) provides tools for managing NVMe devices, and recent commits show increased support for Huawei's custom NVMe over TCP extensions. Additionally, the [DeepSpeed](https://github.com/microsoft/DeepSpeed) library (35k+ stars) from Microsoft has added optimizations for Huawei's storage backend in its latest release, suggesting broader ecosystem adoption.
Key Players & Case Studies
Huawei's Storage Division: Huawei has invested heavily in storage controller ASICs since 2019, with its Kunpeng 920 processor providing the compute backbone for its flash arrays. The company's storage revenue grew 24% year-over-year in 2024, reaching $4.2 billion, driven largely by AI workloads. Huawei's strategy has been to offer 'storage-class memory' performance at flash prices, using a combination of SCM (Storage Class Memory) and QLC NAND in a tiered architecture.
Norwegian AI Research Consortium: The deployment is believed to be at the Norwegian University of Science and Technology (NTNU) in collaboration with the Norwegian Meteorological Institute, which is training a 500-billion-parameter weather prediction model. This model requires continuous streaming of 50 years of historical weather data at 10-minute intervals—a dataset exceeding 200TB. The choice of Huawei was reportedly made after a 6-month evaluation that included Dell, Pure Storage, and NetApp. The decision came down to Huawei's ability to guarantee 0.5ms latency under 100% write load, which no Western vendor could match in their testing.
Competing Solutions:
| Vendor | Product | Max Capacity | NVMe-oF Support | Custom ASIC | AI-Specific Features |
|---|---|---|---|---|---|
| Huawei | OceanStor Dorado 8000 | 32PB | Yes (proprietary) | Yes (Kunpeng) | Checkpoint-aware caching, gradient sync offload |
| Pure Storage | FlashArray//X | 4PB | Yes (NVMe/TCP) | No (Intel Xeon) | None |
| Dell | PowerMax | 18PB | Yes (NVMe/FC) | No (Intel Xeon) | None |
| NetApp | AFF A-Series | 12PB | Yes (NVMe/TCP) | No (Intel Xeon) | None |
Data Takeaway: Huawei is the only vendor offering a custom ASIC for storage processing, which gives it a 3-5x advantage in latency consistency under mixed workloads. This is critical for AI training where checkpointing and data loading compete for the same storage resources.
Industry Impact & Market Dynamics
Norway's decision is a watershed moment for the AI storage market. The global AI storage market was valued at $28.4 billion in 2024 and is projected to reach $89.7 billion by 2030, growing at a CAGR of 21.2%. Huawei currently holds 12% of this market, but the Norway deal could accelerate its share to 18% by 2027.
Market Share Projections:
| Year | Huawei | Pure Storage | Dell | NetApp | Others |
|---|---|---|---|---|---|
| 2024 | 12% | 15% | 22% | 18% | 33% |
| 2025 | 14% | 14% | 20% | 17% | 35% |
| 2026 | 16% | 13% | 18% | 16% | 37% |
| 2027 | 18% | 12% | 16% | 15% | 39% |
Data Takeaway: Huawei's growth is coming at the expense of Western vendors, particularly Pure Storage and Dell, which lack the ASIC-level optimization for AI workloads. The 'others' category includes startups like VAST Data and WekaIO, which are growing rapidly but from a small base.
The geopolitical dimension cannot be ignored. Norway's choice as a NATO member creates a precedent that other European nations may follow. The EU's AI Act does not explicitly restrict storage hardware, and performance-based procurement is legally defensible. We expect to see similar deployments in Switzerland, Finland, and potentially Germany, where research institutions are increasingly frustrated with the performance limitations of Western storage vendors.
Risks, Limitations & Open Questions
1. Supply Chain Security: Huawei hardware remains under scrutiny from Five Eyes intelligence agencies. While Norway has implemented 'air-gapped' deployment with no internet connectivity, the risk of hardware backdoors or firmware vulnerabilities persists. Huawei has opened a transparency center in Oslo for source code review, but trust remains a concern.
2. Vendor Lock-in: Huawei's proprietary NVMe-over-Fabric protocol is not compatible with standard Linux NVMe drivers. This means the storage cannot be easily migrated to another vendor without significant data migration costs. The Norwegian team has mitigated this by using a parallel Lustre filesystem layer on top, but the underlying hardware dependency remains.
3. Performance at Scale: While the 2PB deployment is impressive, it is still small compared to hyperscale AI clusters. Google's TPU v5 pods use over 100PB of storage. Huawei's architecture has not been proven at that scale, and there are questions about its ability to maintain deterministic latency beyond 10PB.
4. Geopolitical Retaliation: The US could potentially use export controls to restrict Huawei's access to advanced NAND flash controllers or manufacturing equipment. This would directly impact Huawei's ability to deliver future storage systems, potentially stranding Norwegian AI projects mid-deployment.
AINews Verdict & Predictions
Prediction 1: By 2027, at least three additional NATO member countries will deploy Huawei storage for AI workloads. The performance gap is too large to ignore, and the political cost is decreasing as AI sovereignty becomes a national priority. Expect Finland, Poland, and Estonia to follow Norway's lead.
Prediction 2: Western storage vendors will acquire AI-focused storage startups within 12 months. Pure Storage or Dell will likely acquire VAST Data (valued at $9B) or WekaIO to gain the ASIC-level optimization they lack. The current organic R&D cycles are too slow to compete with Huawei's custom silicon advantage.
Prediction 3: The US will impose new export controls on storage controllers by Q1 2026. The Biden administration's focus has been on GPUs, but storage is the next bottleneck. Expect restrictions on any storage controller capable of sub-millisecond latency at petabyte scale, effectively targeting Huawei's Kunpeng-based arrays.
Our editorial judgment: Norway's decision is rational and inevitable. AI training is a performance-maximizing activity, and political considerations are secondary when model training costs exceed $100 million. The real story is not about Norway choosing Huawei, but about Western storage vendors failing to innovate. They had a five-year head start and squandered it on incremental improvements while Huawei invested in custom silicon. The AI industry will not wait for politics to catch up with physics.