AI Hits the Electric Wall: Three Winners Benefiting from the Crunch

For a long time, the narrative surrounding artificial intelligence was one of chips, algorithms, and venture capital. But as massive computing campuses sprout up from the ground, a different reality is setting in: the next phase will not just be played out in server rooms, but rather in substations, transmission lines, and the lead times for the most mundane heavy industrial equipment. In fact, it is already happening.

Until late 2025, the dominant AI narrative rested on a simple equation: more compute, larger models, greater value. While this logic remains, it is now colliding with a much harsher reality. Next-generation data centers are no longer mere technical buildings; they are voracious infrastructures capable of absorbing electricity volumes comparable to entire cities.

Billions on the Table, but Construction Stalls

Capital is certainly not the issue. Tech giants (Alphabet, Amazon, Meta, and Microsoft) are set to commit over $650bn this year to expanding their AI-related capacities. Yet, this abundance no longer guarantees project commissioning. Nearly half of the data centers planned in the United States for 2026 could be delayed or even abandoned.

The breaking point is not financial; it is material. Projects are being suspended due to the availability of critical electrical components: transformers, switchgear, batteries, cables, and connection equipment. Their share of a data center's total cost remains modest—less than 10% depending on the context—but their absence is enough to halt an entire site. In this new industrial hierarchy, the least spectacular link has become the most decisive.

Technological America Depends on an Industry It Does Not Control

The United States dominates the technological frontier and maintains a crucial advantage in advanced semiconductors. However, this power falters when it comes to manufacturing the electrical equipment required to actually power these data centers. Despite a decade of industrial reshoring policies, domestic capacity remains insufficient. This dependency notably points towards China, despite tariffs, strategic tensions, and national security concerns. The irony is almost perfect: Washington seeks to maintain its AI lead over Beijing, yet must continue importing the very equipment needed to plug in its own ambitions.

The Real Enemy Is Lead Time

In the old data center economy, waiting 24 to 30 months for a high-power transformer was manageable. In the AI economy, this tempo has become untenable. Operators now want to deploy new capacity in under 18 months, while lead times for certain equipment can climb as high as five years.

The entire logic of the sector is being disrupted. AI lives by the rhythm of rapid investment cycles, narrow market windows, and permanent pressure on time-to-market. The power grid, conversely, moves at a slow pace, burdened by heavy interconnection procedures, strained supply chains, and industrial capacities that cannot be expanded in a single quarter.

This mismatch creates a new form of macroeconomic friction. Innovation is accelerating, but the physical world refuses to follow at the same speed. When grid operators or utilities slow down connections, it is not out of conservatism; it is because the system is not dimensioned to absorb such a sudden concentration of demand.

To move quickly, companies are securing orders very early and sometimes resorting to retrofitting old transformers from decommissioned power plants. When a sector backed by hundreds of billions of dollars must turn to refurbished hardware to meet deadlines, it is a sign that the industrial chain was not prepared to absorb a shock of this magnitude.

The problem is further compounded by the fact that data centers are not alone in the market. The same equipment is in high demand for the electrification of transport, electric vehicles, heat pumps and general grid expansion. AI is not entering an empty highway; it is plugging into an already overstretched infrastructure.

Geopolitics Enters the Engine Room

The other shift, more discreet but equally structural, is geopolitical. Much has been said about digital sovereignty regarding chips or models. Electrical sovereignty must now be added to the list. Dependence on foreign suppliers does not just extend lead times; it also exposes projects to regulatory changes, trade barriers, diplomatic arbitration and supply chain disruptions. The technological standoff between the US and China is thus taking a nearly symmetrical form. One side needs electrical components manufactured in Asia to deploy its computing power. The other needs advanced American technologies to stay in the race for chips and systems. AI is therefore no longer just a competition between labs and cloud giants. It is also a confrontation between productive apparatuses, industrial networks, and construction capacities.

The Next Competitive Advantage Will Not Be Purely Computational

For years, comparative advantage was measured by GPU density and model sophistication. The phase that began several months ago is reshuffling the pack. The differentiator may now lie in less glamorous but far more determining factors: who can connect the fastest, secure their equipment, obtain the necessary power and build a responsive network without waiting five years? AI is thus becoming an energy issue as much as a computing one—and perhaps, above all, an issue of industrial execution. As long as bottlenecks in transformers, switchgear, and interconnections are not eased, announcements of colossal spending will not automatically translate into effective capacity.

Three Obvious Winners

The first obvious winner is GE Vernova. The reason is simple: if the crux of the problem becomes the transmission and conversion of electricity, players positioned across the heavy electrical chain gain strategic value. GE Vernova offers a comprehensive suite for data centers, ranging from grid solutions to microgrids and power conversion—precisely where the bottleneck lies.

The second name is Eaton. As AI becomes a matter of power distribution, protection, load management, and the rapid deployment of critical infrastructure, Eaton is clearly pushing its "grid-to-chip" positioning. It is developing architectures tailored to AI load peaks and working on modular solutions to accelerate data center commissioning. In other words, if the market is now paying for the ability to move electricity intelligently and quickly, Eaton is at the heart of the matter.

The third name is Bloom Energy. Bloom does not primarily sell traditional grid components but rather a workaround solution. If the grid cannot keep up, if interconnections take too long, and if hyperscalers want to secure on-site power, then decentralized production becomes invaluable. Bloom emphasizes that electricity has become the limiting factor for data center expansion, and its narrative shows that operators are increasingly seeking to reduce their grid dependence. It is a direct beneficiary of the "bring your own power" thesis.