The Glomfjord Lesson: Watch the Operational Megawatts
Bill Sharratt
President, Pacific Rubber & Packing
Say "Glomfjord" in a room full of data center people, they’ll know what it is. A remote Norwegian fjord town has become known for something the AI infrastructure boom avoids: the gap between the megawatts in the press release and the megawatts actually running.
The headline mentions the IPO, the financing and the customer. But the least glamorous part of a liquid-cooled AI facility turned out to be the part that took rows of servers offline, again and again.
What Happened at Glomfjord
In mid-October, Bloomberg reported that Nscale's Glomfjord facility in northern Norway was about six months behind schedule, with 6.4 MW of a planned 30 MW running. This is a former crypto mining site, converted for AI, filled with Nvidia B200 GPUs and GB200 NVL72 racks for a single anchor customer.
The reported causes read like a commissioning engineer's bad dream:
Repeated coolant leaks, attributed to faulty pipe sealant and a filter problem. Since December, staff shut down server rows several times to deal with them.
A full cooling system replacement in June, followed by another significant leak in September, this one attributed to human error.
A December transformer fire. Norway's Directorate for Civil Protection (DSB) traced it to a high-voltage cable left outdoors for six months that failed when energized. Its inspection logged ten deviations, including inadequate earthing. The case closed in September after fixes.
Diesel generators carrying a site designed for hydropower until May 2026.
Full disclosure, an anonymous source told DCD the fire caused no delay, the leak was fixed within weeks and no penalties were imposed. Bloomberg's sources say the delays exposed Nscale to penalties worth millions, and customers were offered better terms on new contracts instead. Nscale declined to comment, citing its pre-IPO quiet period.
Pick whichever version you like.
The friction pile-up
No single item on that list is screaming. A cable stored badly. A sealant that should not have been in that loop. A filter that did not do its job. A technician having a bad day. Every one of them is the kind of thing that gets waved through because it costs a few hundred dollars and nobody's bonus depends on it.
Big failures in AI infrastructure rarely arrive as one dramatic event. They arrive as a stack of small, cheap, "we'll sort it out at commissioning" decisions that compound. Each leak means a row goes dark. Each dark row means GPUs that cost more than a house sit idle. Each idle week moves a delivery date, and in 2026 delivery dates have penalty clauses attached.
Nscale's own S-1 makes the point for me: its contracts include delay penalties, and at least one major customer can terminate a tranche if capacity shows up late. The spreadsheet assumed the plumbing would just work. The plumbing had other plans.
Sealing is not a commodity anymore
For decades, anything that sealed a fluid joint in a data center was a line item. Air cooling was forgiving. A weeping fitting in a chilled water plant was a maintenance ticket, not a revenue event.
Direct-to-chip liquid cooling changed the math. Coolant now runs through cold plates with microchannels a fraction of a millimeter wide, sitting millimeters from silicon drawing well over 1,000 watts per GPU. In that environment, every wetted material is a risk vector in two directions:
It can leak. The wrong sealant, the wrong compound, or the wrong installation, and you are de-energizing a row to protect hardware worth millions.
It can shed. Seals, sealants and fittings can release particulates into the loop. Those particles do not cause a dramatic puddle. They quietly collect in cold plate channels and filters, raise pressure drop, reduce heat transfer, and show up later as thermal throttling or a clogged filter nobody can explain.
Glomfjord was reportedly hit by both families of problem: a sealant issue and a filter issue. That pairing should get every operator's attention. Filters are where contamination goes to announce itself.
The truth is that most liquid cooling procurement still treats these components the way air-cooled facilities did: buy to spec, buy to price, assume it is fine. Very few buyers ask for cleanliness data, lot traceability or particulate documentation on the parts that touch their coolant. Semiconductor fabs learned to demand that decades ago. AI data centers are about to learn it the expensive way.
Questions worth asking before your own Glomfjord
If you operate, finance, or lease liquid-cooled capacity, these belong in your next design review or diligence call:
What touches the coolant? Can your team list every wetted material in the loop, from CDU to cold plate, and who qualified each one?
Where is the cleanliness data? Not material safety or flammability ratings. Actual particulate and extractables data for seals, sealants and fittings.
Can you trace it? If a filter loads up in month three, can you trace the contamination back to a component lot?
Who signs off on commissioning fluid quality? Flushing, filtration and fluid sampling should have an owner, a spec and a record.
What does a leak cost you? Put a number on one row dark for one day, including penalties. Then compare it to the price difference on the parts that prevent it.
What do your contracts say? Tenants: ask about cooling-related downtime history. Operators: know whether delay penalties cover "our sealant failed."
Watch the operational megawatts
The AI buildout is measured in megawatts: gigawatt campuses, multi-billion-dollar contract values, GPU counts with a lot of zeros. Glomfjord is a reminder that customers pay for operational megawatts, and operational megawatts depend on things that cost less than a single GPU.
Sources
- Nscale's Glomfjord data center delayed, site suffers leaks in cooling system and a fire, DCD, Oct 16, 2026 (summarizing Bloomberg, Oct 15, 2026)
- Nscale's AI data center in Glomfjord, Norway, is six months late, Kaiku (DSB inspection, S-1 and Macquarie filing details)
- Nscale's Data Center Delays Put Delivery Risks in Focus Ahead of IPO, The Energy Mag
Continue your engineering review
Connect this topic with cooling-loop architecture, seal qualification, and coolant compatibility.
Explore liquid cooling engineering guides →