The United States is discovering that building data centers for AI isn’t just about buying more GPUs. According to Bloomberg, nearly half of the planned data center projects for 2026 in the country are at risk of being delayed or canceled due to shortages of key electrical components such as transformers, medium-voltage cells, and batteries, many of which are made in China or heavily reliant on global supply chains where China continues to play a decisive role. Bloomberg highlights the issue as stemming from electric infrastructure components, not a lack of capital or chip demand.
This nuance matters. It’s not exactly that “half have already been canceled,” but that a significant part of the pipeline for 2026 is vulnerable due to physical bottlenecks. Sightline Climate, cited by various outlets and in its own analysis, estimates that in the U.S. there are about 16 GW of announced capacity for large data centers and AI factories scheduled to come online in 2026, but only about 5 GW are currently under construction. The remaining 11 GW remain in announced status without visible progress, despite typical construction timelines of 12 to 18 months.
At the same time, investment continues to flow into AI infrastructure. Bloomberg reports more than $650 billion in spending this year by Alphabet, Amazon, Meta, and Microsoft to expand capacity—a figure that other analyses place within a much broader boom in CAPEX for data centers. The problem is that money doesn’t produce transformers faster, doesn’t speed up switchgear delivery, nor resolve network congestion on its own.
The bottleneck is no longer just chips
For months, the debate centered around GPUs, HBM, and advanced packaging. Now, the focus is shifting to electrical infrastructure. Bloomberg indicates that waiting times for certain transformers can stretch up to five years, even though many data center projects are planned with roughly three-year construction horizons. This directly contradicts the financial and operational logic of many developments: if substations, transformers, or batteries are delayed, buildings may be ready, but the installations won’t be operational.
Dependence on China is also a significant factor. Although the U.S. has attempted to reduce exposure to advanced semiconductors and servers, it still relies heavily on an industrial base vital for electrifying new campuses. Bloomberg emphasizes that China remains the world’s largest producer of electrical equipment necessary both for internal data center infrastructure and for external systems that supply these centers.
This explains why the bottleneck affects not just a single supplier or hyperscaler but the entire ecosystem: operators, developers, utilities, engineering firms, and manufacturers are all competing for the same critical assets. Ultimately, the bottleneck has shifted from the rack to the grid.
And in Europe? Yes, the problem exists, but with nuances
Europe isn’t immune to this trend. Several indicators point to the continent facing a similar challenge. The European Data Centre Association (EUDCA) warned in its State of European Data Centres 2026 report that energy availability is already the main bottleneck to sector growth, citing network congestion, long connection times, and increased demands from AI workloads.
The International Energy Agency (IEA) echoed this at the end of 2025: within the European Union, waiting times for grid connections can range from two to ten years, depending on the country. Such delays alone threaten numerous projects, even before considering transformers or batteries.
Adding to this, Eurelectric highlighted that in major European digital hubs, connection delays can reach up to 13 years. Additionally, grid operators face challenges in sourcing components, with transformer delivery times around 2.5 to 4 years. While Europe isn’t exactly replicating the same industrial dependency pattern as the U.S., it shares the core issue: electrical infrastructure is lagging far behind the demand for AI capacity.
Meanwhile, the European Commission has already acknowledged that data centers are becoming a central energy and regulatory challenge for the EU. Brussels is preparing additional measures within its energy efficiency package for data centers, recognizing that the growth of these facilities is inseparable from pressures on the grid, energy consumption, and industrial planning.
It’s not just China — scale, networks, and timelines are the real issues
It would be overly simplistic to say “China has the parts.” China’s influence is significant—it plays a major role in manufacturing transformers, batteries, and electrical equipment. But the real issue is that demand for AI data centers is growing at a much faster rate than electrical supply chains and national grids. This creates a perfect storm: more projects, larger capacities per campus, increased competition for transformers and switchgear, and connection times that were long even before the AI boom.
Therefore, in Europe, while an exact replica of U.S. grid lock probably won’t happen, similar pressures will likely impact project feasibility and schedules. The risk isn’t just component shortages but that many developments may have land, permits, investors, or even advanced construction, yet still await years for electrical connections or critical equipment to energize their campuses.
The clear consequence: the primary limit for AI is no longer just chips. Increasingly, electricity and everything related to it is becoming the decisive factor.
FAQs
Has half of the AI data centers in the U.S. already been canceled?
Not exactly. Bloomberg suggested that almost half of the projects planned for 2026 may be delayed or canceled due to shortages of key electrical equipment. It’s a risk to the pipeline, not an extensive list of already executed cancellations.
What is the main bottleneck right now?
Transformers, switchgear, batteries, grid connections, and, broadly, electrical infrastructure. The challenge isn’t just acquiring GPUs but powering and energizing data centers on schedule.
Does Europe face the same problem as the U.S.?
Yes, but with nuances. Europe mainly suffers from network congestion, long connection times, and a global scarcity of electrical components. It’s not an exact copy of the U.S. case, but the practical effect—delays of several years in data center projects—is similar.
What does this mean for AI projects in 2026 and 2027?
Many plans can still move forward financially, but physical and electrical infrastructure may limit progress. In other words, money and chips are no longer enough if energy, grid, and critical equipment don’t arrive on time.
via: tomshardware