Summary

  • NVIDIA and its partners have announced European AI-factory capacity in Germany, the United Kingdom and the Netherlands, expressed in GPU counts, investment totals and gigawatts, with every headline figure originating from the announcing parties rather than from an operating record.
  • Public registry and announcement records reviewed for this briefing do not establish which autonomous systems, registered IP prefixes, RIR memberships or interconnection arrangements NVIDIA itself operates behind that announced capacity.
  • For buyers and investors the decision-relevant question is not how much compute has been announced but which named party operates the facility, holds the power contract and carries the network path once the capacity begins serving customers.

The European AI build-out is now routinely described in vendor programme totals: accelerators, exaflops and gigawatts. Those totals are useful, but they answer a different question from the one that determines whether a service can be bought, measured and sustained. This briefing separates what has been announced in Germany, the United Kingdom and the Netherlands from what has been shown to be operating, and it is explicit about the layer where the public record runs out.

What has been announced

In Germany, NVIDIA and Deutsche Telekom announced an industrial AI cloud to be built in a renovated Munich data centre, reported at about 10,000 NVIDIA Blackwell GPUs, with a launch target of the first quarter of 2026. Deutsche Telekom is stated to operate the physical infrastructure (NVIDIA and Deutsche Telekom announcement coverage). The GPU figure is accelerator inventory that the two companies say will be installed. It is not a statement about how many of those processors are in commercial service on any given day, nor about which customers have contracted for them.

In the United Kingdom, NVIDIA and partners including Nscale, CoreWeave and Microsoft announced AI-factory investment intended for completion by the end of 2026, with Nscale-linked capacity described as NVIDIA Grace Blackwell systems (UK AI-factory investment coverage). The announcement names suppliers and an intended completion window. It does not identify which party operates the underlying network.

In the Netherlands, NVIDIA agreed to invest USD 2bn in Netherlands-headquartered Nebius Group to support a full-stack AI cloud, with a stated ambition of up to 5GW of NVIDIA-based computing by 2030 (Nebius investment coverage). The gigawatt figure is an infrastructure ambition attached to a partnership announcement. It is not a metered grid connection, a contracted power purchase, or a routed capacity number.

Around those country-level announcements sits a continental frame: NVIDIA describes a programme of about 20 AI factories, five of them gigafactory-scale, delivering more than 3,000 exaflops of Blackwell compute (programme framing coverage). These are vendor programme totals that aggregate announced capacity across partners and sites. They are not a register of commissioned, customer-served capacity, and they should not be read as one.

Three units that do not measure the same thing

The most common analytical error in this material is treating GPU counts, exaflops and gigawatts as interchangeable measures of the same thing.

An accelerator count describes hardware that a supplier and a partner say will be installed at a described site. It is a procurement and delivery statement, and it can be true while the site is empty.

Exaflops describe arithmetic throughput at a stated numeric precision under ideal conditions. Vendor throughput figures are useful for comparing generations of the same product family; they say very little about delivered performance on a mixed customer workload, where memory bandwidth, interconnect topology, storage throughput and scheduling behaviour usually dominate.

Gigawatts describe an intended power envelope. That envelope depends on grid connection, permitting, cooling and land — the slowest-moving inputs in any data-centre programme, and the ones least likely to be resolved at announcement.

A programme total that mixes all three units is a marketing aggregate, not a capacity statement. The three numbers can all be revised in the same direction — or in different directions — without any of them being false at the moment it was published.

Where the layers separate

An AI factory is not one asset. It is at least five layers, usually owned and operated by different parties with different incentives:

  • Accelerator supply. NVIDIA designs and sells the processors and the rack-scale systems; in Munich and in the UK capacity, NVIDIA is the named platform supplier.
  • System integration and delivery. Server and rack integration, firmware, cluster bring-up and software stack.
  • Facility, cooling and power. In the Munich project, the physical infrastructure is stated to sit with Deutsche Telekom, a national operator with existing grid relationships. In the UK and Dutch announcements, the facility layer belongs to partner or host operators rather than to NVIDIA.
  • Cloud service operation. Whoever runs the commercial service — telekom cloud operations in Germany, Nebius for its own capacity, and the named partners in the UK — sets pricing, service levels, region definitions and customer terms.
  • The network layer. Autonomous systems, registered IP address prefixes, RIR membership, transit purchase, peering and the interconnection path into each site.

NVIDIA's named role in the European announcements sits mainly at the first layer, with investment and platform commitments reaching into the second and, via Nebius, the fourth. The fifth layer is the least documented of the set.

The network layer is the least documented part of the programme

Public registry and announcement records reviewed for this briefing do not identify which autonomous systems, registered IP prefixes, RIR memberships or interconnection arrangements NVIDIA itself operates behind the announced European AI-factory capacity (NVIDIA European programme material). The country-level announcements describe GPU counts, partners and timings; they do not describe who originates the routes, contracts the transit and peers into the sites that will serve European customers.

That is a gap in the record, not a claim about the technology. It matters because the network layer is where several commercially decisive properties are actually fixed. Latency between a customer and a training or inference cluster is set by routing and interconnection, not by the accelerator count. Data-residency and sovereignty claims are only as strong as the physical and logical path the traffic takes, including any transit that leaves the claimed jurisdiction.

Subnet control, address-space ownership and route-origin assurance determine how a service behaves under attack, under a misconfiguration, or under a dispute with a hosting partner. And the failure domains that a customer's own resilience plan depends on are usually drawn at the network layer, not in the rack.

A company that buys compute from a partner without knowing who operates that layer is buying an outcome it cannot independently verify. The public record at present supports the facility and platform attributions in Munich and the partner attributions in the UK and the Netherlands. It does not support a claim that NVIDIA operates the European network path underneath them, and it does not support the opposite claim either.

What would settle it

This question is checkable, and the checks are public or contractual rather than speculative:

  • Registry records. Allocations and registrations in the relevant RIR database for European address space and autonomous system numbers attributed to NVIDIA entities would establish owned address and routing resources. Absence of such records is itself informative.
  • Routing data. Announced-prefix and origin data for any autonomous systems attributed to NVIDIA, together with their upstreams, would show whether traffic to announced sites is carried on NVIDIA-originated routes or on partner networks.
  • Interconnection records. Peering records and internet-exchange presences for the operating parties would show where European customer traffic enters and leaves the compute.
  • Facility and power evidence. Executed grid connection and power agreements, planning consent, and commissioning milestones would move a site from announced to operating.
  • Commercial service evidence. Published regions, service levels, pricing and named customer workloads would show capacity actually being sold rather than capacity being promised.

The one-line version of the test: announced capacity becomes operated capacity when a named party publishes regions, prices and service levels for it, and when the routes carrying that traffic can be traced to a registry holder. Until then, the programme totals describe intentions.

NVIDIA maintains an entity entry in the BTW directory (entity:nvidia), which records the company as the platform supplier at the centre of these announcements.