Roland Berger published its findings on July 23, in a study titled "AI Infrastructure: Mapping the Global Data Center Race," the first of two planned reports on the AI data center ecosystem, authored by Edeltraud Leibrock, Nikolaus Lehmann, Urs Neumair, Wim D'Hondt, and Siyi Hao. The findings draw on Roland Berger's own Data Center Model, tracking capacity growth, energy demand, and investment activity, supplemented by a survey of 70 data center industry decision-makers across Europe.
The study projects that global installed data center capacity will nearly quadruple by 2035, from 88 gigawatts in 2025 to as much as 340 gigawatts. Over that same period, AI's share of total data center load is expected to rise from about 20% today to nearly 60%. Europe's slice of that growing pie is shrinking, not expanding: the report projects Europe's share of global installed capacity falling from roughly 13% today to about 10% by 2030.
A boom that may be outrunning its own revenue
The scale of current investment is itself part of the story. Hyperscaler capital expenditure is running at roughly $750 billion in 2026, mostly directed at data centers, and is straining global supply chains for transformers, turbines, cooling systems, and chips, according to Nikolaus Lehmann, a Partner at Roland Berger's Munich office. That spending is arguably ahead of what it's generating in return. "AI capex is racing ahead of revenue," said Wim D'Hondt, a Director at the firm's Montreal office, noting that cashflow from core business earnings and external financing are bridging the gap for now, but that momentum will only hold if revenues and costs move toward breakeven. The report flags a genuine near-term correction risk: hyperscaler capex is broadly approaching, and for some players already exceeding, operating cash flow, and supply chain dynamics mean even a modest slowdown in demand could trigger a sharp upstream contraction. Public opposition is already influencing permitting decisions in some regions, the report notes, on top of hard constraints ranging from grid access to memory supply.
Four barriers, and Portugal has already documented all four itself
The shift toward AI has changed the underlying economics of a data center. Modern AI campuses now draw power on a scale once associated with heavy industry, with the largest planned facilities consuming as much electricity as a major offshore wind farm can generate. That's why Roland Berger's Urs Neumair frames power, not chip supply, as the actual binding constraint on new AI infrastructure right now. Data centers themselves aren't the end product; they're what makes AI model training, inference, cloud computing, and a wide range of other high-value digital industries possible at all, which is why governments care as much about where they get built as private investors do.
Against that backdrop, Roland Berger names four specific obstacles slowing Europe's build-out: grid connections, energy costs and access, regulatory complexity, and limited site availability. No single reform closes the gap. Fixing permitting speed, for instance, does nothing about electricity prices or grid capacity on its own. Connecting a new data center to the electrical grid can take up to seven years in Europe, the report finds. Neumair frames the resulting US-Europe divide starkly: American developers increasingly build dedicated gas-fired generation directly alongside their data centers to route around a strained grid entirely, while European developers instead retrofit former power-plant sites specifically because those sites already have grid connections in place. The European approach avoids the need for entirely new grid connections, but it doesn't address the underlying delays in expanding grid capacity itself. Among the 70 industry experts Roland Berger surveyed, 69% expect energy availability to become an even bigger challenge by 2035 than it already is; electricity already accounts for roughly 40% of a data center's operating costs, and European energy prices remain significantly higher than in competing regions.
None of this is abstract for Portugal. Previous reporting by this publication found that data center developers have filed grid connection requests totalling roughly 26.5 gigawatts nationally. Those are requests, not approved projects or committed demand. But the figure already comfortably exceeds Portugal's entire installed generation capacity of about 23 gigawatts. The Plano Nacional de Centros de Dados, the government's own national plan approved in March, exists specifically to address the regulatory half of Roland Berger's diagnosis: a single point of contact through AICEP instead of a maze of agencies, pre-zoned and pre-licensed land, and fixed maximum timeframes for approvals, for the first time. Whether that plan can outrun a seven-year grid-connection reality it doesn't directly control is a separate question the national plan doesn't really answer.
Why Roland Berger frames this as a sovereignty problem, not just an infrastructure one
"Domestic compute underpins digital sovereignty and the productivity-driven GDP growth advanced economies need. Without it, a country depends on crucial infrastructure controlled elsewhere," said Edeltraud Leibrock, the report's senior author. The study points to a specific, recent example of what that dependency can mean in practice: in June 2026, a US export-control directive forced Anthropic to suspend foreign access to its two most advanced AI models overnight, an event Roland Berger calls "a vivid reminder that reliance on foreign compute and foreign models is a dependency that can be revoked without warning." Anthropic has itself confirmed that suspension took effect June 12 and was reversed on July 1, after the underlying export controls were lifted. Roland Berger's argument is that without sufficient domestic compute capacity, Europe risks losing twice over: the productivity gains AI brings to the wider economy, and any real stake in the industry driving that growth.
Independent numbers point the same way
Roland Berger's warning isn't the only recent data point on this gap. Stanford's AI Index 2025, an independent academic benchmark with no commercial stake in the outcome, found that the US produced 40 notable AI models in 2024 against 15 from China and just 3 from Europe, and that private AI investment in the US, at $109.1 billion, ran roughly 12 times the United Kingdom's and 24 times Germany's. That's a different methodology and a different institution reaching a broadly similar conclusion to Roland Berger's own commercial survey.
Europe's own answer, and its own delays
Europe isn't standing still on this. The European Commission's AI Gigafactories initiative, backed by a €20 billion InvestAI facility, aims to build up to five large-scale European facilities, each equipped with more than 100,000 advanced AI processors. An informal call for expressions of interest drew more than 75 proposals from 16 member states across 60 sites, a response the Commission itself has called far exceeding expectations. But the formal call for proposals has slipped repeatedly: originally targeted for December 2025, then pushed to early 2026, then to the second quarter of 2026, with no confirmed timeline as of this writing. The repeated delays are broadly consistent with the implementation and regulatory challenges Roland Berger identifies, even if the Commission has not attributed the timetable changes to those specific factors.
The picture inside Europe isn't uniformly bleak, either. Stargate Norway, a 100,000-GPU facility under construction in Narvik backed by OpenAI, Nscale, and Aker, represents a $1 billion initial phase targeting 230 megawatts of capacity, rising toward 520 megawatts, running on Norwegian hydropower with direct-to-chip liquid cooling and waste heat recycled to local industry. First GPUs are expected online by the end of 2026. Finland and Sweden separately offer electricity prices closer to US levels than the European average, a real structural advantage some Nordic sites can offer that most of the rest of the continent can't. The pattern that emerges isn't "Europe is failing," but something narrower: parts of Europe with cheap, abundant power and faster permitting are keeping pace, while the EU's larger, more bureaucratic markets are the ones falling behind.
Portugal has spent the past year positioning itself as an exception to that broader European struggle: aggressive national targets, a joint bid with Spain for one of the EU's AI gigafactories, and a permitting overhaul built explicitly to move faster than the rest of the continent. Roland Berger's report doesn't single Portugal out either way, and its own second installment, still to come, may say more about how individual countries are actually doing. What the report does establish, and what Stanford's independent numbers and the EU's own stalled gigafactory timeline both reinforce, are the core constraints: grid connections, energy costs, permitting speed, and site availability, the same ones earlier reporting by this publication already found sitting underneath Portugal's own data center ambitions. Whether Portugal's national plan is fast enough to actually outrun that bottleneck remains an open question. Portugal's answer will ultimately be measured by projects connected to the grid, not plans approved on paper.