Summary

• Blackfuel has selected Digital Realty’s BCN1 facility for a multi-megawatt AI inference deployment using liquid-cooled Dell infrastructure and AMD GPUs.
• The service connects directly to Digital Realty’s ServiceFabric network so customers can reach inference capacity privately from applications, clouds, networks, and enterprise data environments.
• BCN1 opened this year with 14MW of planned capacity, placing the inference deployment inside Barcelona’s expanding Mediterranean connectivity hub.

Enterprise AI infrastructure is beginning to move closer to the applications and corporate data that actually use it, with a new multi-megawatt Barcelona deployment designed around inference rather than the vast training clusters that have dominated much of the data-centre investment cycle.

Digital Realty says Blackfuel has selected its BCN1 data centre as the first location for a dedicated AI inference platform built with liquid-cooled Dell Technologies infrastructure and AMD GPUs. Neither company has disclosed the precise megawatt figure beyond describing the deployment as multi-megawatt.

The project is being connected directly to Digital Realty’s ServiceFabric interconnection platform, allowing Blackfuel to deliver inference services over private network paths to customers whose applications, data, clouds, and service providers already connect into the wider Digital Realty ecosystem.

That architecture reflects a different infrastructure problem from training a frontier model. Large training runs can consume enormous amounts of compute for sustained periods, but production inference is repeated continuously as users, applications, and automated agents send requests to models. Latency, network consistency, data movement, and proximity to existing systems can therefore become as important as the raw performance of the accelerator.

Inference is becoming a network problem

Digital Realty argues that connecting the inference platform privately can reduce the delay before a model begins returning an answer, particularly when an application has to make repeated model calls rather than a single request. The company gives the example of an agentic workflow making between 20 and 50 sequential inference calls, where cutting 50 milliseconds from time-to-first-token on each request could reduce an overall task by between one and 2.5 seconds.

Those figures are illustrative calculations supplied by Digital Realty rather than independently measured performance results from the Barcelona system, but they show why network design receives more attention once AI becomes part of an operational workflow. A small delay repeated through a long chain of automated calls can accumulate into a noticeable slowdown, particularly when software agents wait for the result of one inference before deciding what to do next.

Private connectivity also gives organisations another way to control how data moves between internal systems and model infrastructure. It does not automatically solve questions around model governance, privacy, data residency, or security, but it can reduce reliance on sending every request across ordinary public internet paths.

ServiceFabric currently spans more than 800 Digital Realty and third-party data centres, while Digital Realty says the service also connects customers with hundreds of cloud on-ramps. For Blackfuel, that network provides a route for taking a Barcelona deployment beyond customers physically colocated in the same building.

Barcelona’s connectivity is part of the proposition

BCN1 opened in May and has 14MW of planned capacity across roughly 15,000 square metres of technical space. Located in the Sant Adrià de Besòs area, the facility forms part of Digital Realty’s Iberian network alongside its established Madrid operations and a newer presence in Lisbon.

Barcelona’s role in that network is influenced as much by submarine and terrestrial fibre as by local computing demand. Mediterranean cable routes now connect the city into systems spanning Europe, Africa, the Middle East, and beyond, while overland links provide onward routes towards other European data-centre markets.

That connectivity has been developing alongside new routes carrying data between Barcelona, northern Spain, and Atlantic cable systems, giving the city a stronger role as an exchange point rather than merely an endpoint for Spanish workloads.

For inference providers, dense connectivity can affect where capacity is commercially useful. A GPU cluster with poor access to customers, cloud platforms, enterprise networks, or data sources may be technically powerful but operationally awkward, whereas a smaller pool of compute in a well-connected location can serve distributed workloads with more predictable network behaviour.

Dedicated inference adds another layer to AI

Blackfuel’s infrastructure also points towards a more specialised market between hyperscale clouds and companies purchasing their own GPU systems. Dedicated inference providers can offer accelerator capacity and model-serving infrastructure without customers needing to operate the underlying hardware themselves, while connectivity platforms allow that compute to be reached from existing hybrid environments.

The commercial question is whether organisations running sustained AI workloads will value that separation enough to support a distinct infrastructure category. Hyperscalers already offer inference services, GPU instances, private connectivity, and managed AI platforms, while enterprises can also deploy their own systems where utilisation, governance, or latency justifies the capital expense.

A provider such as Blackfuel therefore has to compete on more than accelerator access. Availability, predictable performance, network proximity, private connectivity, pricing, operational support, and the ability to serve several model stacks without forcing customers into a single cloud ecosystem all become part of the offer.

Digital Realty has an interest in that model because it makes the data centre and its interconnection fabric more than a property layer beneath AI services. Instead of simply leasing power and space to a GPU operator, it can connect that capacity directly to clouds, carriers, enterprise systems, and other facilities across its network.

The Barcelona installation remains the first step in Blackfuel’s wider plan, so its eventual significance will depend on customer adoption and whether the company expands beyond BCN1. Even so, the deployment captures a change already visible in enterprise AI: once models move from experiments into workflows, the location of inference, the path to corporate data, and the milliseconds between sequential requests start becoming infrastructure decisions rather than software details.