Europe · GPU clouds
European neoclouds: signed capacity runs far ahead of live
Nscale, the UK neocloud that filed for a US listing in September 2026, had about 25,000 GPUs active against 461,000 active and contracted, and $2.6bn of active contract value against $103.4bn signed, as of 31 Aug 2026 (S-1, SEC EDGAR). Nscale and Nebius, the two European-headquartered neoclouds that publish both, have 251 MW live against 3,370 MW contracted.
Updated 11 October 2026 · data build 2026-10-10 · how we know · corrections
Nscale: 5% of its signed GPUs are live, and 6% of its signed megawatts
From Nscale's S-1 registration statement, as of 31 Aug 2026. 'Active' means online and earning revenue; 'contracted' means under signed customer contracts but not yet built. Measured
The 81 MW is our sum of the five sites the S-1 lists as active (Portugal 30, Norway 6, 5 and 25, Iceland 15 IT MW); the filing gives no single active-MW figure. Total contract value (TCV) is the company's own measure over full contract terms; the audited remaining performance obligation was $56.4bn at 30 Jun 2026. The largest contracts are with Microsoft (up to $43.8bn, through 2033) and Anthropic (up to $44.6bn, at a US site, financing not yet secured). The S-1 also records OpenAI withdrawing from Stargate UK and Stargate Norway in April 2026, with Microsoft taking the Norway capacity, and says the company's reliance on uncommitted financing initially raised substantial doubt about its ability to continue as a going concern, which management judged its plan to cut or defer capital spending would resolve.
Source: Nscale Ltd, Form S-1, SEC EDGAR, filed 18 Sep 2026 · as of 2026-10-11 · Compiled by StackPath
Europe's two big neoclouds have 7% of contracted power live; US peers have 17% and 41%
Contracted MW against live (active) MW, global, as each company reports them. US-headquartered CoreWeave and Crusoe are shown for comparison. Calculated
| Company | HQ | Contracted MW | Live MW | Live share | As of |
|---|---|---|---|---|---|
| Crusoe | US | 6,000 | 1,000 | 17% | Sep 2026 (gross MW) |
| CoreWeave | US | 3,700 | 1,500 | 41% | Q2 2026 |
| Nebius | NL | 2,000 | 170 | 9% | live Dec 2025; contracted Feb 2026 |
| Nscale | GB | 1,370 | 81 | 6% | 31 Aug 2026 |
Live MW as a share of contracted MW. Definitions differ: Nscale reports IT MW (S-1, 31 Aug 2026); Nebius reports active power at 31 Dec 2025 against contracted power at Feb 2026, so its ratio mixes two dates; CoreWeave's figures are from its Q2 2026 filing and Crusoe's are gross MW from a company release. Treat the ratios as indicative.
Source: Nscale S-1 · Nebius 20-F and 6-K (SEC EDGAR) · CoreWeave 10-Q · Crusoe release, 17 Sep 2026 · as of 2026-10-11 · live share is our arithmetic · Compiled by StackPath
18 GPU clouds with European capacity, on identical fields
Headquarters, European sites, announced and live capacity, backlog, largest contracts, funding and status. US-headquartered neoclouds appear for their European sites only. Reported
Download CSV ↓Source: company filings, releases and press (see CSV for every link) · as of 2026-10-11 · Compiled by StackPath
Contracts, funding and sources by company
Nscale (GB, S-1 filed (IPO pending)). Filed S-1 with SEC 18 Sep 2026 (no S-1/A on EDGAR as of 11 Oct 2026); 25k active GPUs vs 461k active+contracted; H1 2026 net loss $1,020.1m Live MW = sum of IT capacity contracted at the five 'active' sites in the S-1 Facilities table (all in Europe: PT 30, NO 6+5+25, IS 15 = 81 IT MW); the S-1 does not state a single active-capacity MW figure. Contracted 1.37 GW includes active. Europe total 576 IT MW (active+contracted); remainder US 770 MW and Indonesia 24 MW. 10 GW is 'potential power capacity' (line of sight), mostly Monarch, West Virginia. Site names are not given in the table; mappings to Glomfjord/Kvandal/Sines are inferences.
Contracts: Microsoft $43.8bn (Statements of work Sep 2025-Apr 2026, payments up to ~$43.8bn through Dec 2033, subject to delivery/availability); Anthropic $44.6bn (Anthropic Services Agreements 25 Aug 2026, up to ~$44.6bn, Vera Rubin NVL72 at Monarch Compute Campus (West Virginia, US); no binding financing commitments yet); NVIDIA $1.2bn (Master cloud services agreement Jul 2025: reserved GPU capacity in Iceland, UK and Norway, 12-48 month terms, $1.2bn aggregate); Figure (Figure Services Agreement 24 Aug 2026: potential up to 100,000 Vera Rubin GPUs from H2 2027; value not stated); OpenAI (Stargate UK) (Stargate UK announced 16 Sep 2025 with NVIDIA, Microsoft, OpenAI; OpenAI withdrew Apr 2026 (per S-1))
Funding: Not stated cumulatively in S-1. Equity: Series C $1.64bn (27 Mar 2026), earlier Series B and SAFEs. Debt facilities (commitments): GPU Financing $1.4bn (Feb 2026), revolver $900m, $725m senior term loan (Jul 2026), Macquarie Iceland $331.9m, Ward County GPU $1.85bn and North Carolina GPU $1.2bn (Aug 2026), Dell Financial Services leases ~$2.54bn, Sep 2026 subscription agreement min $3.1bn. Cash $1.5bn at 30 Jun 2026; debt carrying value $137.4m.
Sources: Form S-1: TCV, RPO, GPUs, sites table, contracts, financials, going concern · ClusterMAX 2.0 review: Unavailable · ClusterMAX 3.0 free text
Nebius (NL, NBIS (Nasdaq)). Listed AI cloud; 170 MW active power end-2025, >2 GW contracted Feb 2026, targeting 5 GW contracted by YE 2026 and >1 GW/year deployment from 2027 Contracted power = secured by land and power commitments (20-F definition); active power = consumed by IT equipment and revenue-generating. Latest published active-power figure found is 170 MW at 31 Dec 2025; no Q2 2026 active MW found. Lappeenranta (310 MW), Bethune (240 MW) and Huuru (EE) site names/MW from the brief were NOT found in the 20-F or Q2 letter and are unverified. Mantsala expansion MW not verified. 5,000 MW is a YE2026 contracted target, not achieved.
Contracts: Microsoft $17.4bn (5-year dedicated GPU capacity at Vineland, New Jersey (US); TCV up to ~$17.4bn through 2031; all tranches delivered per Q2 2026 letter); Meta (2025) $2.9bn (Cloud Infrastructure Services Agreement 1 Nov 2025, 5-year, two tranches delivered Nov 2025 and Feb 2026); Meta (2026) $27bn (Infrastructure Services Agreement 13 Mar 2026, up to ~$27bn incl. $12bn of dedicated-cluster orders (deploy H1 2027) plus an unsold-capacity order); Reflection, Cohere and others (Q2 2026) (Four Q2 deals averaging >$1bn TCV each at $20-25m ACV per MW)
Funding: Not stated cumulatively. Convertible notes: $4.16bn (Jun+Sep 2025), $4.34bn (Mar 2026, plus greenshoe), $5.0bn (priced 19 Aug 2026); public equity offering Sep 2025; NVIDIA pre-funded warrant ~$2bn (11 Mar 2026); ATM sales May-Jun 2026; first secured debt $775m (Jul 2026). Cash $8.04bn at 30 Jun 2026.
Sources: 20-F FY2025: sites, contracted/active power definitions, Microsoft/Meta contracts · Q2 2026 shareholder letter: revenue, ARR, 5 GW target, new sites · Q2 2026 financial statements: RPO $37.49bn, convertibles · Q2 2026 OFR: cash, liquidity sources · $5.0bn convertible pricing · UK 65 MW by 2027 · ClusterMAX 2.0: Gold · ClusterMAX 3.0: Platinum · Nebius newsroom: 310 MW Lappeenranta, 240 MW near Lille, Mäntsälä 75 MW
CoreWeave (European sites) (US, CRWV (Nasdaq)). Largest US neocloud; 1.5 GW active, ~3.7 GW contracted (Q2 2026); European footprint in UK, Spain, Sweden, Norway with no per-site MW disclosed Global MW figures are company totals, not Europe. Europe MW not disclosed in filings. UK GBP 2.5bn commitment, Crawley/London Docklands, Barcelona and specific Nordic partners from the brief are unverified here. Major contracts (OpenAI, Meta, Microsoft) not re-extracted in this pass.
Funding: Q2 2026: $3.1bn delayed-draw term loan; $1bn Jane Street strategic investment; >$10bn unsecured debt and convertibles incl. inaugural Eurobond.
Sources: Q2 2026 results: revenue, backlog, active/contracted power · 10-Q: RPO $103.7bn · 10-K: countries of data centres; 850+ MW active end-2025, 3.1 GW contracted · ClusterMAX 2.0: Platinum · ClusterMAX 3.0: Platinum
Fluidstack (US). Moved global HQ from London to New York (4 Dec 2025); $830m Series A at $7.5bn valuation (Jan 2026); building custom data centres for Anthropic in New York and Texas Founded in London; company now describes New York as its global HQ, so hq set to US. No live European capacity verified. No MW, GPU, backlog figures found in public company releases.
Funding: Series A $830m, $7.5bn valuation, led by Situational Awareness (Jan 2026). Debt/backstops not verified.
Sources: HQ relocation to NYC · Series A $830m / $7.5bn (round dated Jan 2026) · ClusterMAX 2.0: Gold · ClusterMAX 3.0: Unavailable · Structure Research: Fluidstack withdraws from France
Mistral Compute (Mistral AI) (FR). Compute platform launched Jun 2025 ('tens of thousands of GPUs'); Aug 2026 plan of up to 1 GW by 2030 via enterprise coalition (European Compute Units); still onboarding first customers per SemiAnalysis 1 GW is a 2030 ambition stated in the page summary, not in the article body. Essonne/Bruyeres-le-Chatel site, 40 MW and 18,000 Grace Blackwell figures NOT found on Mistral pages fetched; unverified.
Funding: Equity: EUR 1.7bn (Sep 2025), EUR 3bn Series D at >EUR 21bn post-money led by Samsung (8 Sep 2026). Data-centre debt not verified. $830m of debt from a bank consortium for the Bruyères-le-Châtel site (Maddyness, 30 Mar 2026).
Sources: Mistral Compute launch · Up to 1 GW by 2030 (summary line) · EUR 3bn Series D · Maddyness: $830m debt, 13,800 GB300, 44 MW
Sesterce (FR). Announced campuses (Bosquel 1 GW with SoftBank, Rovaltain 40 MW, Jamsa 200 MW phase 1); no live MW or GPU count published All MW are planned; Bosquel is a JV figure (Sesterce share not stated). Jamsa image text mentions 600 MW vs 200 MW phase 1 - 200 used. EUR 1.8bn plan and 40,000 GPU target from the brief unverified.
Sources: Bosquel 1 GW JV · Rovaltain 40 MW · Jamsa 200 MW / EUR 10bn · ClusterMAX 3.0
Evroc (SE). Sovereign cloud live with customers (Swedish Armed Forces 'Combat Cloud', KPMG Sweden, Jun 2026); own data centres planned; partner sites in Paris, Stockholm, Frankfurt 10,000 = planned GB300 GPUs per SemiAnalysis page (Nov 2025). Mosas/Orebro and MW figures unverified; funding not found.
Sources: ClusterMAX 2.0 review: sites, up to 10,000 GB300 plan · Newsroom: customers 2026 · evroc flagship release · evroc Mougins release
Polarise (DE). German sovereign AI cloud; partner on Telekom's Munich Industrial AI Cloud; joining an EU AI Gigafactory bid consortium (Aug 2026) No MW or GPU counts for live sites published. Not mentioned in ClusterMAX free text.
Funding: 'High double-digit million-euro' debt financing from SWI Stoneweg Icona (18 Aug 2026); amount undisclosed.
Nebul (NL). Dutch sovereign AI cloud (NVIDIA Elite Partner); 2025 Gartner Cool Vendor; no capacity figures published No sites, MW, GPUs or funding found in public pages fetched. Not in ClusterMAX free text.
Sources: Company site
Northern Data / Taiga Cloud (now Quake AI, RUM Group) (DE, RUM (Nasdaq, parent RUM Group Inc.)). Acquired by Rumble 17 Jun 2026 (85.2% of shares); parent renamed RUM Group; GPU business now Quake AI with ~22,000 H100/H200 at ~85% utilisation and ~250 MW unmonetised capacity targeted for 2027 Company HQ kept as DE (Northern Data AG); parent RUM Group is US-listed. 250 MW is 'targeted for 2027', location not specified. Per-site MW not disclosed. Not in ClusterMAX free text.
Contracts: Together AI (Multi-year agreement to deploy HGX B300 GPU cloud capacity (Q2 2026))
Funding: Acquisition by share exchange; Tether holds pre-funded warrants and a EUR 317.5m note; up to $200m Tether funding arrangement for a VAT matter.
Sources: 8-K: completion of Northern Data acquisition · Q2 2026 results: 22k GPUs, 250 MW, revenue · 10-Q: subsidiaries, Tether arrangements
Verda (formerly DataCrunch) (FI). Raised $189m (EUR 163m) Series B led by Emergence Capital (22 Sep 2026); total funding >$450m; offers GB300 NVL72, B300, B200, H200 Bronze tier is company-stated; not confirmed in SemiAnalysis free text. Sites (Helsinki, Iceland), MW and GPU counts not found.
Funding: $450m Company-stated total funding 'over $450M (EUR 390M)' incl. Series B $189m (22 Sep 2026); mix of equity and debt not split.
Sources: Company site: funding, ClusterMAX 3.0 Bronze claim (25 Sep 2026) · ClusterMAX 3.0 free text
Gcore (LU). Luxembourg edge/cloud provider with GPU cloud (Hopper GPU VMs Mar 2026); 50+ cloud locations, 210+ PoPs No GPU counts, MW, Sines site or funding figures verified.
Sources: HQ Luxembourg, locations · ClusterMAX 2.0: Silver · ClusterMAX 3.0 free text
Hyperstack (NexGen Cloud) (GB). London-based GPU cloud with infrastructure in 'Europe and Canada'; offers H100/H200, B300, GB300 NVL72 No sites, MW, GPU counts or funding verified.
Sources: Company site · ClusterMAX 3.0 free text
Scaleway (Iliad) (FR). Iliad's cloud arm; Milan region (Mar 2026); EU Cloud III award (Apr 2026); AI Gigafactory bid consortium (May 2026); acquired Qarnot (Oct 2026) GPU count, MW, OpCore pipeline and Scaleway revenue not verified.
Sources: News list 2026 · ClusterMAX 2.0: Silver · ClusterMAX 3.0
OVHcloud (GPU/AI part) (FR, OVH (Euronext Paris)). General cloud with GPU offers; Q3 FY2026 revenue EUR 289.6m (+6.9% LFL), Public Cloud +20%; FY2026 results due 20 Oct 2026; GPU revenue not disclosed Revenue in EUR (EUR 289.6m), not converted; whole-company, not GPU. No GPU MW disclosed.
Sources: Q3 FY2026 revenue · ClusterMAX 2.0: Underperforming · ClusterMAX 3.0
Crusoe (European sites) (US). US AI infrastructure developer/cloud; 6 GW+ gross contracted, 1 GW gross operational (Sep 2026); European (Iceland/Norway) sites not mentioned in 2026 releases MW are gross (not IT) and global. Iceland and Norway sites from the brief NOT verified in sources fetched; europeSites left empty.
Funding: $3.9bn Series F initial close at $30.9bn post-money (17 Sep 2026); cumulative not stated.
Sources: Series F, contracted/operational GW, TCV · ClusterMAX 2.0: Gold · ClusterMAX 3.0: Bronze
Cerebras (Europe) (US, CBRS (Nasdaq)). Listed chipmaker with inference cloud; no European data-centre site found in Q2 2026 10-Q Europe 200 MW target from the brief NOT verified; only a France subsidiary is mentioned (functional currency note). Include only if a European site is confirmed.
Contracts: OpenAI (MRA Dec 2025: 750 MW inference capacity committed, deployed 2026-2028; option for 1.25 GW more by 2030)
Sources: 10-Q: OpenAI 750 MW, RPO $25.4bn · W.Media: Cerebras targets 200 MW in Europe
CloudCIX (IE). Cork-based sovereign cloud platform (registered at Hollyhill, Cork; c/o Cork Internet eXchange); GPUs supported NVIDIA B200 deployment from the brief NOT verified; no GPU count or MW published on site.
Sources: Company site
42 European neocloud sites, 156 MW of them live
Each site as announced or reported; MW as the company states it. Reported
Solid bars are live; light bars are under construction, contracted or planned. Not shown: Fluidstack's 1,000 MW at Bosquel, Somme (dropped Mar 2026; site now a SoftBank–Sesterce JV).
Download CSV ↓Source: company releases and press (one link per row) · as of 2026-10-11 · Compiled by StackPath
How SemiAnalysis rates them: Nebius at the top, others criticised over managed-cluster claims
SemiAnalysis's ClusterMAX rates GPU clouds on reliability, software and support from hands-on testing and customer interviews. It does not measure capacity. Paraphrased from the free sections, with links. Reported
| Company | ClusterMAX | What SemiAnalysis says (paraphrased) |
|---|---|---|
| Nscale | Unavailable (not yet tested) (ClusterMAX 3.0) | Not yet tested in 3.0 (also unavailable in 2.0); SemiAnalysis says Nscale's marketing of managed clusters and inference endpoints does not match what it offers. link |
| Nebius | Platinum (ClusterMAX 3.0) | Nebius moved up from Gold (2.0/2.1) to Platinum in 3.0, alongside CoreWeave. link |
| CoreWeave (European sites) | Platinum (ClusterMAX 3.0) | CoreWeave remains Platinum; 3.0 notes its focus on long-term bare-metal contracts that still include networking, burn-in and day-two operations. link |
| Fluidstack | Unavailable (exited managed clusters) (ClusterMAX 3.0) | Fell from Gold in 2.0 to unavailable in 3.0 after moving to very large bare-metal TPU deployments. link |
| Mistral Compute (Mistral AI) | Unavailable (not yet tested) (ClusterMAX 3.0) | Not yet tested; Mistral told SemiAnalysis it was still onboarding first customers in August 2026. link |
| Sesterce | Unavailable (not yet tested) (ClusterMAX 3.0) | Not tested; a trial of its on-demand H100 offer turned out to be resold from a third-party marketplace. link |
| Evroc | Unavailable (ClusterMAX 2.0) | SemiAnalysis lists planned sites but has not been able to test evroc's services. link |
| Verda (formerly DataCrunch) | Bronze (company-stated) (ClusterMAX 3.0) | Company states it was rated Bronze; SemiAnalysis free text says its Instant Clusters changes appear meaningful but notes shared-filesystem locking problems. link |
| Gcore | Silver (ClusterMAX 2.0) | Rated Silver in 2.0; 3.0 free text praises its attentiveness but states no new tier. link |
| Hyperstack (NexGen Cloud) | ClusterMAX 3.0 | Test called a success but reviewers worried about the human support required; the section precedes the Participation Ribbon heading, so the tier is inferred. link |
| Scaleway (Iliad) | Unavailable (no capacity or declined testing) (ClusterMAX 3.0) | Silver in 2.0; in 3.0 it had no capacity to test or declined, and was reported to be shifting from Nvidia to AMD. link |
| OVHcloud (GPU/AI part) | Underperforming (ClusterMAX 3.0) | Rated underperforming in both 2.0 and 3.0; SemiAnalysis sees its general-purpose cloud as a large but slow-moving player in AI. link |
| Crusoe (European sites) | Bronze (ClusterMAX 3.0) | Dropped from Gold in 2.0 to the bottom of Bronze in 3.0, citing unreliable clusters. link |
Source: SemiAnalysis, ClusterMAX 3.0 (23 Sep 2026) · ClusterMAX 2.0 (6 Nov 2025) · clustermax.semianalysis.com · paraphrased; ratings belong to SemiAnalysis · Compiled by StackPath
Our reading. The useful number for a neocloud is not the headline backlog but how much of it is built. Nscale's filing is the first time a European neocloud has had to state both under securities law, and it shows about one GPU live for every eighteen signed. Contracted capacity can still slip or be cancelled, as OpenAI's withdrawal from Stargate UK shows.
'Signed' and 'live' use each company's own definitions: Nscale's active vs contracted IT MW (S-1), Nebius's active vs contracted power (20-F; the two dates differ), and gross MW for Crusoe. CoreWeave and Crusoe are US-headquartered and their figures are global; they are shown as comparators. Most European GPU clouds publish no MW or GPU counts. A blank is a gap in disclosure, not zero. Nscale's S-1 facilities table gives country, type and IT MW but no site names; matching rows to Glomfjord, Kvandal or Sines is not attempted here. ClusterMAX ratings are SemiAnalysis's, paraphrased from the free sections of its posts with links; Verda's Bronze tier is as stated by the company.
Questions people ask
- What is a neocloud?
- A cloud company that rents out GPUs for AI training and inference, usually as dedicated clusters on multi-year contracts, rather than offering the full range of services a hyperscaler like AWS, Azure or Google Cloud does. Nebius, Nscale, Fluidstack and CoreWeave are examples.
- How much capacity does Nscale actually have running?
- Its S-1 (filed 18 Sep 2026) says about 25,000 active GPUs and 461,000 active and contracted GPUs as of 31 Aug 2026, five active and twelve contracted sites, and about 1.37 GW of active and contracted capacity. Active contract value was $2.6bn against $103.4bn active and contracted. H1 2026 revenue was $140.6m with a net loss of $1,020.1m.
- Which European neocloud is rated highest by SemiAnalysis?
- Nebius. SemiAnalysis's ClusterMAX 3.0 rating (23 Sep 2026) put it in the top Platinum tier alongside CoreWeave, up from Gold in ClusterMAX 2.0 (Nov 2025). ClusterMAX rates service quality, not capacity.
- Are there neoclouds in Ireland?
- Only small ones. CloudCIX in Cork runs a sovereign cloud with GPUs but publishes no capacity, and TensorX, a Dublin inference cloud, raised an €8m seed round in June 2026 (Silicon Republic). None of the large European neocloud sites is in Ireland; they cluster in the Nordics, Iceland, Portugal, the UK and France.