Summary
- Nebius will deploy an initial 22MW at Greenergy Data Centers' Hüüru site near Tallinn, with the first phase targeted for operation later in 2026.
- A second Mäntsälä facility of up to 70MW is planned from 2027, taking Nebius's stated Finnish footprint to 455MW across three sites.
- The programme spreads AI capacity across several Northern European power markets rather than concentrating expansion behind a single campus.
Nebius is extending its Northern European AI infrastructure footprint with a 22MW deployment at Greenergy Data Centers in Estonia and plans for a second facility of up to 70MW at Mäntsälä in Finland.
The Estonian capacity will be installed at Greenergy’s Hüüru data centre west of Tallinn, with the initial 22MW phase scheduled to enter operation later in 2026. Nebius will become the site’s first international AI cloud customer, while Estonia’s investment agency expects the deployment to become the largest AI infrastructure installation in the Baltic states at that initial scale.
In Finland, another Mäntsälä facility is planned with up to 70MW from 2027. Nebius says its combined Finnish estate would then reach 455MW across three sites, adding to an increasingly distributed European portfolio of owned facilities and third-party capacity.
Hüüru moves into a denser class of load
Greenergy Data Centers already operates the Hüüru facility, but a 22MW AI deployment changes the demands placed on the site. High-density GPU infrastructure brings electrical distribution, cooling, controls, networking, and commissioning into much closer alignment than a conventional enterprise colocation installation.
The facility is being expanded to support that load, with nearly €200m of wider investment previously reported around the site. Tensor Estate became a shareholder in Greenergy in July, while the Nebius commitment supplies a large anchor customer for the next phase of physical capacity.
Bringing 22MW into service before the end of the year leaves a relatively short delivery window. Utility capacity has to be available when the computing equipment arrives, but energisation alone does not make the installation ready. Switchgear, busway, cooling plant, controls, network connectivity, fire protection, and the GPU systems themselves all need to pass through commissioning before the load can generate revenue.
Neither the Estonia material nor the available detail on the second Mäntsälä facility specifies rack densities or the final cooling architecture. Those omissions limit any assessment of the mechanical system, particularly when modern AI racks can demand liquid cooling and place much larger heat loads on facility water loops than older air-cooled halls.
The same applies to power procurement. Estonia and Finland both offer access to increasingly low-carbon electricity systems, but annual energy sourcing and physical connection capacity remain separate questions. A data centre may procure renewable electricity while still requiring major substations, network reinforcement, and firm backup systems to support continuous operation.
A distributed European capacity model
Nebius is assembling capacity across several countries rather than waiting for one giant campus to carry its European growth. That can reduce exposure to a single connection queue or planning process, and it gives the company more options for placing compute near customers, fibre routes, and particular jurisdictions.
The operating complexity rises in parallel. Several countries mean more utility relationships, different construction supply chains, multiple maintenance organisations, and potentially varying electrical and cooling designs. Standardising hardware can help, but the facility layer still has to accommodate local infrastructure and regulation.
In the UK, Nebius has already taken 22MW at Kao Data’s Harlow campus, while Ark is expanding Longcross Park around another Nebius deployment. Further projects extend the same model into France, Spain, Iceland, and other markets.
Rapid expansion also exposes the distinction between facility efficiency and total resource use. Nebius has reported strong efficiency metrics and heat recovery at some operations, but its rising infrastructure footprint has also increased absolute electricity demand and market-based Scope 1 and 2 emissions. More capacity can improve the efficiency of individual computing workloads while still increasing the company’s total environmental load.
Capital follows the same physical expansion. GPU infrastructure is expensive long before it produces revenue, and Nebius has increasingly used financing structures tied to deployed computing assets. Connection delays or commissioning problems therefore affect both capacity delivery and the economics of hardware already procured for a site.
Finland and Estonia remain attractive locations precisely because power, climate, and digital connectivity can support this type of development, but neither is insulated from infrastructure pressure. AI demand is moving quickly enough that power systems which once looked comfortably supplied can face concentrated connection requests around a small number of suitable industrial locations.
The Hüüru phase now has the closest deadline: 22MW is expected online later this year. Its delivery will provide a clearer measure of how quickly an established Baltic colocation site can be converted into infrastructure capable of carrying dense AI load, while the second Mäntsälä facility extends that test into 2027.
