COMPUTE ATLAS Supercomputer supply-chain graph
1153 systems 482 sites Sourced data

Blog/2026-10-07

· Compute Atlas

How Nebius and Nscale build GPU clouds

Nebius owns Finland and rents in Iceland and the UK; Nscale mostly leases, with Microsoft and OpenAI as anchors. Their PUE and WUE figures are targets or company statements, and only one Nscale cluster has a benchmark.

Nebius and Nscale are the two neoclouds that sell European and Nordic GPU capacity at scale, and they have built it in opposite ways. Nebius owns one large campus in Finland, rents halls in Iceland and England, and has pinned its balance sheet to a Microsoft contract in New Jersey. Nscale owns little and signs a lot: it runs its own site in Arctic Norway, builds campuses at Narvik and Loughton, and leases in Iceland and Portugal, with Microsoft as the dominant buyer and OpenAI named at two sites. The efficiency numbers attached to these campuses are mostly targets or company statements, not measurements, and one widely repeated Nscale claim turns out to belong to its landlord. Among the clusters discussed, only Nscale’s Svartisen has a TOP500 result in our data, and it runs older AMD hardware. This article lays out the structures, then what can and cannot be checked. Our entries include nebius-mantsala, nebius-longcross-uk, nebius-verne-iceland, nscale-sines-gb300, microsoft-nscale-narvik-norway, microsoft-nscale-loughton-uk, nscale-verne-iceland and nscale-svartisen.

Three ways to build a neocloud

The first model is to own the campus. Nebius’s Mäntsälä site in Finland is the clearest case. A Nebius release of 8 October 2024 announced that capacity would triple to 75 MW, accommodating up to 60,000 GPUs, with H100 installed, H200 from November 2024 and Blackwell expected in 2025. The same company’s March 2026 announcement adds a 310 MW multi-building campus at Lappeenranta on a site of about 100 acres, with first customer capacity in 2027 and NVIDIA Blackwell and Vera Rubin hardware. That release gives no PUE, no WUE and no investment figure.

The second model is colocation, where the neocloud rents a hall and fills it with its own GPUs. Nebius took this route in Iceland: Verne’s March 2025 release describes a 10 MW cluster of H200 GPUs at its campus and calls it the largest single implementation in its Icelandic history. Nebius’s own post of the same month gives the 10 MW figure and expected full operation by the end of March 2025 and does not name the GPU model, so the H200 detail rests on Verne. In England, Techerati reported 4,000 Blackwell Ultra B300 GPUs in a first phase growing to about 9,000 at Ark Data Centres’ Longcross Park campus in Surrey, a site of 54 MW across two data centres. Nebius’s 6 November 2025 page confirms Blackwell Ultra and Quantum-X800 InfiniBand but, in the version we fetched, describes the site only as in the London area and gives no count. The 54 MW belongs to the site, not to Nebius’s cluster.

The third model is the anchor contract. Nscale’s growth is built on these. Its 15 October 2025 release announced an expanded deal with Microsoft for about 200,000 NVIDIA GB300 GPUs across Europe and the US: roughly 104,000 in a 240 MW Texas campus from Q3 2026, about 12,600 at Sines from Q1 2026, about 23,000 at Loughton from Q1 2027, and about 52,000 from the Aker-Nscale joint venture at Narvik. Nebius’s anchor is also Microsoft, though on a different structure. An SEC prospectus supplement states a five-year agreement signed 8 September 2025 for dedicated GPU clusters at a new Vineland, New Jersey facility, worth about $17.4 billion through 2031 (up to $19.4 billion with extra services), delivered in tranches during 2025 and 2026 and subject to deployment and availability. Nebius says it expects to finance the build from the contract’s cash flows and debt secured against the contract. That is the point of an anchor tenant: it turns a customer’s credit into the collateral for the equipment.

Nscale’s four sites

Glomfjord and the one benchmark. Nscale’s own site at Glomfjord, inside the Arctic Circle, was 30 MW when the InfraPartners announcement of March 2025 described an expansion to 60 MW using prefabricated modules, powered by local hydroelectric energy with liquid cooling for the latest GPUs. The cluster there, Svartisen, is built on AMD Instinct MI250X GPUs and EPYC 7713 hosts. TOP500 lists 7.40 petaflops Rmax and 12.38 Rpeak from 66,528 cores, ranked 156th in November 2024 and 239th by June 2026, with no power figure on the page. Nscale’s account of the submission says it ran Ethernet with RoCE and was standing up an AMD MI300X platform. Among the sites discussed it is the only HPL result we hold, and it is for older hardware.

Narvik. The Narvik story has changed three times. The Stargate Norway release of 31 July 2025 set a 230 MW plan with ambitions to add 290 MW, a target of 100,000 NVIDIA GPUs by the end of 2026, a 50/50 Nscale-Aker joint venture and OpenAI as an initial offtaker with an option to scale. The partners committed approximately $1 billion to the initial 20 MW phase, including upwards of $250 million in equity. Then Microsoft appeared: the October 2025 contract above puts about 52,000 GB300 GPUs at Narvik, and Microsoft’s own feature values a five-year agreement with Nscale and Aker at $6.2 billion, starting service in 2026, without giving a GPU count. On 14 April 2026 Nscale added more than 30,000 Rubin GPUs for 2027 on the 230 MW campus, now solely managed by Nscale after the joint venture was rolled up. That release does not mention OpenAI. A May 2026 financing of $790 million from Nordic and Dutch banks, with an equal uncommitted accordion for a further 115 MW, is project finance for the campus. We cannot tell from fetched pages what became of the 100,000-GPU OpenAI target, and the 230 MW figure has carried at least three different GPU totals.

Loughton and Stargate UK. Nscale’s UK announcement describes 23,040 GB300 GPUs at Loughton on 50 MW scaling to 90 MW with delivery from Q1 2027 for Microsoft Azure. The same release describes a separate Stargate UK arrangement with OpenAI and NVIDIA: up to 8,000 GPUs in Q1 2026, potentially reaching 31,000, spread across multiple sites including Cobalt Park. It states a UK target of up to 58,640 GPUs overall. Nothing on that page says any of this was installed.

Sines and Keflavik. At Start Campus in Sines, Portugal, Nscale’s May 2026 release says over 12,600 Blackwell Ultra GPUs are deployed in the first building for Microsoft, and announces 66,000 or more Vera Rubin NVL72 GPUs from late 2027, with Nscale investing 230 million euros in shared infrastructure and 465 million euros in a second 200 MW building. Start Campus is fully permitted for 1.2 GW. In Iceland, Nscale is the tenant of a 15 MW lease at Verne for about 4,600 Blackwell Ultra GPUs, 85 percent liquid cooled, per Verne’s release. BeBeez reported the first 7.5 MW phase complete in April 2026, with 128 racks including 64 liquid-cooled GB300 racks.

PUE and WUE: who is claiming what

The figures that circulate for these campuses deserve a source-by-source reading.

SiteFigureStated byNature
SinesPUE 1.1, WUE 0Start CampusTarget, per its own text
MäntsäläPUE as low as 1.1Nebius, Oct 2024Described “under high IT loads”
Nebius portfolioPUE 1.25 (2025)Nebius sustainability reportIT-load-weighted annual average
Finland siteWUE 0.018 L/kWhNebius sustainability reportReported for end of 2025

The Sines row matters most, because we had recorded it as an Nscale claim. In Nscale’s May 2026 release the sentence sits inside the “About Start Campus” paragraph, which says the campus “targets” a PUE of 1.1 and a WUE of 0 using ocean cooling. Start Campus’s own page on the GB300 deployment uses the same words, adds rack densities above 130 kW and says Portugal’s grid was 71 percent renewable in 2024, alongside its statement that the campus uses 100 percent renewable energy; the page does not say how the two are reconciled. A target of 0 water use applies to the landlord’s design; neither page gives a metered number, and both predate any year of operation with this load.

Nebius’s numbers are better documented. The October 2024 release gives a best-case PUE of 1.1 under high load, about 20,000 MWh of recoverable heat a year (heating about 2,500 Finnish homes by its estimate) and mentions a separate supercomputer, ISEG, at the facility that it credits with 35.26 GFlops per watt, without relating that machine to the cloud clusters. The 2025 sustainability report, as relayed by a wire service on 30 July 2026, reports a portfolio PUE of 1.25 against an industry average of 1.54, a Finland WUE of 0.018 L/kWh and 19.5 GWh of heat exported to district heating in 2025. The heat export is in line with the 2024 expectation. The distance between the 1.1 headline and the 1.25 portfolio mean is the usual one: the best site at full load looks better than the weighted fleet. The report is company-published and we did not see an external audit. The Lappeenranta release promises closed-loop liquid cooling that eliminates reliance on local water and says Mäntsälä’s heat recovery avoided about 4,000 tonnes of CO2e in 2025, but sets out no PUE or WUE for the new site.

What the numbers do and do not show

Lease megawatts are not IT megawatts. Nscale’s Keflavik lease is 15 MW for about 4,600 GPUs. Epoch AI estimates about 2,300 B300 chips and 5 MW of IT power now, reaching 10 MW and 4,600 chips by the fourth quarter of 2026, with 15 MW projected for year end, and counts 13 dry coolers for 8.6 MW of cooling capacity. These are modelled estimates, not operator figures, and they disagree with the lease on timing, but they bracket the same fact: the first 7.5 MW phase was a half of the lease, and the two give different densities: 4,600 GPUs on 15 MW of lease is about 307 per MW, while Epoch’s 4,600 on 10 MW of IT power is 460 per MW (our arithmetic). A lease figure therefore appears to include headroom or overhead that an IT-power figure does not, or one of the two is off.

Counts of “deployed” GPUs are also self-reported. Sines’s 12,600 figure appears in Nscale’s release and on Start Campus’s page, which is two statements but a single origin of fact. For Longcross, the count rests on trade press. For Vineland, no page we read names a GPU model or count.

What we could not confirm

We could not confirm any measured PUE or WUE for a Nscale site; the status of OpenAI’s capacity at Narvik; any installed GPU count at Loughton; the site, model and count of Nebius’s UK cluster from Nebius itself; whether Microsoft’s $6.2 billion Narvik value overlaps the roughly 52,000 plus 30,000 GPU figures; or the power draw of any of these sites in operation. We also did not fetch Nscale’s S-1, filed per a 18 September 2026 release, which should be the first audited source on its contracts and capital spending.

What to watch

Three things would move this from claims to evidence. The first is the Nscale prospectus once public: contracted revenue, tenant concentration and lease terms for Sines, Narvik and Loughton. The second is a year of operation at Sines and Keflavik, which would let the PUE and WUE targets be tested against metered data. The third is whether Nebius’s Lappeenranta campus ships in 2027 with the heat recovery and closed-loop cooling described, and whether the 2026 sustainability-report format, with a portfolio PUE and site WUE, becomes a norm that Nscale and the landlords it rents from are asked to match.

Sources

deep-diveai-clusterspower-and-coolingsovereignty