The editorial argues this isn't a security scandal but a supply-chain math problem: DDN, VAST, and Pure Storage are quoting 12-18 month lead times on petabyte-scale all-flash because hyperscaler demand has locked up QLC NAND. Huawei, sitting on excess YMTC capacity and frozen out of Western enterprise accounts, can ship in weeks — making it the only realistic option for a research consortium on a training deadline.
The piece highlights the contradiction that the same Norwegian government that ordered Telenor to rip out Huawei radio gear by 2024 is now, through a different ministry, routing NorwAI's sovereign-LLM training data through Huawei OceanStor controllers running proprietary OceanOS on Kunpeng Arm chips. The notability isn't the 2PB scale but the buyer being a NATO state-funded consortium that publicly distanced itself from the same vendor.
The editorial notes the deal evaded scrutiny because enterprise storage doesn't trigger the reflexes that Ascend accelerators or HiSilicon chips do, even though OceanStor Dorado runs Huawei's own silicon and OS. Resale through a regional integrator — the standard pattern — kept it off the radar entirely, exposing a gap in how Western controls treat 'non-AI' Huawei infrastructure.
Submitted the Blocks & Files report to Hacker News where it drew 206 points and 104 comments, signaling that the developer community sees the procurement as a meaningful data point about sovereign AI supply chains rather than a routine storage purchase.
Norway's national HPC consortium Sigma2 has deployed roughly 2 petabytes of Huawei OceanStor all-flash storage to feed its sovereign LLM training cluster, according to a report from Blocks & Files. The system sits at the NTNU IDUN facility in Trondheim and is being used to train Norwegian-language foundation models under the NorwAI program — the same effort that produced NorwAI Mistral 7B and is now scaling toward larger parameter counts.
The procurement is notable not because 2PB is enormous by hyperscaler standards — it isn't — but because the buyer is a state-funded research consortium inside a NATO member that has spent the last three years publicly distancing itself from Huawei in its 5G core. Telenor finished ripping out Huawei radio gear in 2024. The same government that signed those orders is, through a different ministry, now feeding its national LLM training data through Huawei controllers.
The deal slipped under most radars because storage arrays don't trigger the export-control reflexes that Ascend AI accelerators or HiSilicon chips do. Huawei's OceanStor Dorado line runs on Arm-based Kunpeng processors and Huawei's proprietary OceanOS, but it's classified as enterprise storage — not on the US Entity List in any way that prevents a European public buyer from acquiring it. The hardware was reportedly resold through a regional integrator, which is the standard pattern.
The interesting story here isn't espionage panic. It's the procurement math. DDN, VAST Data, and Pure Storage — the three vendors a Western sovereign-AI program would normally call first — are quoting 12 to 18 month lead times on petabyte-scale all-flash configurations, driven by the same NVIDIA-adjacent demand that has every hyperscaler hoarding QLC NAND. Huawei, sitting on excess Chinese-domestic NAND capacity from YMTC and frozen out of most Western enterprise accounts, can ship in weeks. For a research consortium racing to train a model before its grant cycle ends, that gap is decisive.
This is the gap that sovereign-AI policy keeps refusing to look at. Every European AI strategy document published in the last 18 months — from the EU AI Continent Action Plan to the German AI Gigafactory tender — treats compute as the bottleneck. Storage is assumed. But training a 70B-parameter model on a multi-terabyte tokenized corpus, with checkpointing every few thousand steps, generates IO patterns that commodity NFS cannot serve. You need parallel filesystems on NVMe-class media, and you need them now, not in Q3 2027.
The second issue is the bifurcation it creates. Sigma2 isn't a defense contractor; it's a university-aligned HPC center, and the LLM weights it produces will be released openly. There is a legitimate argument — one the consortium has not publicly made but that the procurement implies — that Huawei storage in a research context, holding pretraining corpora that are themselves public, is a different risk profile than Huawei storage in a telco core handling subscriber metadata. But Norway has not articulated that distinction publicly, and the absence of an articulated framework is the actual story: there is no European-wide policy on what "sovereign AI infrastructure" actually excludes.
Community reaction on Hacker News (206 points, 140+ comments) split predictably. One camp pointed out that the alternative — waiting 18 months for VAST — effectively cedes Norwegian-language model leadership to whoever ships first, likely Mistral or Meta. Another camp noted that firmware on storage controllers is a credible vector for data exfiltration during training, particularly when the training data itself becomes valuable IP. Both sides are right. The synthesis nobody offered: this is what "de-risking" looks like when it actually has to ship.
If you're running infrastructure procurement for an AI workload right now, three things follow. First, audit your storage roadmap on a 24-month horizon, not a 12-month one — the QLC shortage is real and the major Western vendors will allocate to their largest customers first, which means startups and mid-size research orgs are getting pushed to the back of the queue. Cerebras, Groq, and the various Saudi and UAE-funded labs have already locked up significant DDN capacity through 2027.
Second, the Huawei question is going to become a board-level conversation faster than most CTOs expect. The US Bureau of Industry and Security has been signaling for months that storage controllers and data-center networking gear are next on the export-control review list, with a particular focus on equipment used in AI training. If your sovereign-AI strategy quietly depends on Huawei in any tier of the stack, you need a documented mitigation path before the rules change, not after.
Third, watch the parallel filesystem layer. Lustre, BeeGFS, WekaFS, and DAOS are all viable above the block layer, which means the controller brand matters less than the orchestration above it. A defensible architecture treats the array as commodity and keeps the IP — the data layout, the checkpoint scheduling, the corpus tokenization — in software you control. This is the path Sigma2 appears to be taking, and it's the path most European AI programs will end up on whether they want to or not.
The Norway-Huawei story will get cited as a one-off, but it's the leading edge of a pattern. Every country that has committed to a sovereign foundation model — Norway, Sweden, France, Germany, the Netherlands, Spain, Italy, Japan, Canada, Australia — is going to face the same procurement squeeze in the next 18 months. Some will quietly pick Huawei. Others will pay the premium for DDN or VAST and explain the delay to their funders. A few will end up renting capacity from US hyperscalers and pretend that's sovereignty. The one outcome that won't happen is the clean Western-stack story that the policy documents currently assume.
How true is this statement: "He asserted that any country with its own language that did not have a sovereign LLM trained in that language was at a disadvantage as a globally trained, English-speaking LLM would not know about that country’s history, news and culture that was described in the lo
> The Olivia system is an HPE Cray Supercomputing EX system, with 448 GPUs and 64,512 CPU cores.Training a sovereign LLM with this meager hardware as opposed to a LORA on some open source model seems like a huge mistake and a potential red flag.There is no way these people have the resources to t
I wonder if instead (or in parallel), Norway should build a set of training data and share it (for free) with all the model builders.Seems like making the frontier models know Norwegian and their culture is a better (or additional!) way to reach the end they are going for here.
> Marius Husnes, the Head of IT Platform at the library (Nasjonlbiblioteket) discussed the project at Huawei’s ID Forum 2026 in Paris, saying that no commercial LLM provider was developing a local (Norwegian) language LLM. He asserted that any country with its own language that did not have a sov
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
I'm a Norwegian, and I use the national library almost every day for searching through texts. They have truly one of the best working user interfaces (and functionality) for searching through the massive amounts of text.