The rapid expansion of artificial intelligence has run headlong into a very unglamorous limit: the electricity needed to power the data centersIt is not the lack of chips or capital that is currently holding back many projects, but the inability to obtain a powerful network connection in time.
Faced with this reality, major digital infrastructure operators are resorting to solutions that would not have been considered extreme not long ago. AI data centers already use turbines based on airplane engines...as well as diesel and gas generators, not just as backup power, but as the primary energy source for months or even years. The sector has gone from discussing algorithms and models to an almost desperate race for power outlets and megawatts.
AI's energy demands are overwhelming electrical grids
The rise of generative AI models has driven the electricity consumption of data centers to levels that many networks cannot absorb. In the United States, The time required to obtain a high-power connection is extended by between five and seven years.And there's even talk of waits approaching a decade in certain areas. For an industry that operates on a timescale of months, waiting that long is simply unfeasible.
In that context, generate energy on site It has become the preferred shortcut for major tech companies. Energy companies and specialized providers are building their own facilities next to data centers to bypass the network access queue. What was once an emergency measure is becoming a structural part of cloud infrastructure.
This tension is not limited to the United States. Although the text relies primarily on American examples, Europe is beginning to see similar signs: large data center projects that clash with saturated networks, slow procedures, local opposition and regulatory doubts about how to fit this explosive electricity demand into national energy plans.
The problem is the speed of AI deployment It doesn't fit with the planning timelines for networks, renewables, and storageWhile new high-voltage power lines, offshore wind farms, or modular reactors are being debated, data centers need to be operational today, not in five or ten years.
In practice, the digital infrastructure of artificial intelligence is being built with short-term decisions: First, make sure the system works; there will be time to think about efficiency later.That approach, several analysts warn, is precisely what is driving up costs, emissions, and financial risks.
Aeroderivative turbines: from the 747 to the AI server
The image is striking: aircraft engines converted into stationary generators to power server farms. So-called aeroderivative turbines, based on commercial aircraft engine cores, are being installed next to state-of-the-art data centers to produce tens or hundreds of megawatts almost immediately.
Companies like GE Vernova already supply these types of units to megaprojects linked to large AI consortiumssuch as the Stargate data center powered by OpenAI and Microsoft in Texas. The logic is simple: these turbines can be commissioned within months, distributed in modules, and located close to where the power is needed.
There have also been specialized actors who They reuse iconic commercial aircraft enginesProEnergy, for example, is buying CF6-80C2 engine cores, known for their use in Boeing 747s, to rebuild them as ground-based units capable of generating around 48 megawatts each, enough energy to supply tens of thousands of homes.
The phenomenon is not only technological, but also financial. The case of Boom Supersonic illustrates this well: its CEO, Blake Scholl, planned first to develop its supersonic aircraft and, later, to explore energy uses for its engines. However, a call from Sam Altman, CEO of OpenAIThis altered that order. The message was direct: the priority was not the planes, but obtaining additional electrical power for AI projects as soon as possible.
As a result of this urgency, Scholl has confirmed that it will sell to the company Crusoe turbines virtually identical to those of their future supersonic aircraftWith a clear objective: to convert engines designed for flight into a stable source of electricity for the cloud. This conversion demonstrates the extent to which the boundary between aeronautics, energy, and digital technology is blurring.
From emergency backup to being the main source: diesel and gas are making a strong comeback
Alongside aeroderivative turbines, Diesel and gas generators are no longer just insurance against power outagesTraditionally, these systems were only switched on occasionally, when the network failed or demand spiked. Today, in many AI data centers, they are being used as the primary power supply for extended periods.
Manufacturers like Cummins have seen this demand skyrocket. The company claims to have already sold around 39 gigawatts of capacity to data centersdoubling its volume in just one year. What's significant is not only the number, but the change in use: these generators were designed as backup power and now function as the continuous energy heart of the facility.
The US government itself is beginning to consider scenarios bordering on a war economy. Energy Secretary Chris Wright has publicly raised the possibility of temporarily requisition backup generators from data centers and large facilities (such as hypermarket chains) to support the electrical system during times of extreme stress. A clear sign that the grid is being overwhelmed by this new demand.
This shift has serious implications for global energy planning. Plants that were scheduled to close or be repurposed They have reconsidered their decisions in light of the opportunity to supply the new AI boom. This is the case with highly polluting facilities, such as the historic Fisk plant in Chicago, whose closure has been halted to meet the needs of these centers.
In day-to-day operations, all of this translates into something as prosaic as fuel tankers entering and leaving server complexes. The cloud, which was always described as something ethereal and clean, now depends on combustion engines running nonstop. to keep running AI services that millions of users perceive as almost magical.
An expensive model: energy at double the price and soaring emissions
The cost of this strategy of generating one's own energy is not exactly low. According to calculations by analysts at BNP Paribas, electricity from a gas-fired power plant built for a large technology client in Ohio It costs around $175 per megawatt hour, roughly double the average cost faced by a standard industrial consumer in the area.
Compared to other conventional sources, the difference is also significant. Nuclear, wind, solar, or even modern coal-fired power plants are usually considerably cheaper when integrated into the network mixThe problem is that these projects require years of processing, investment, and civil works, something that clashes with the urgency of AI deployments.
In addition to this economic cost, there are significant environmental impacts. Experts like Mark Dyson of the Rocky Mountain Institute warn that The emissions from these isolated plants and generators are clearly worse than those of the general grid, which combines relatively efficient gas with a growing share of renewables. Concentrating less optimized fossil fuel generation near data centers worsens the overall climate balance.
Noise, fuel-related traffic, and potential leaks or incidents They add friction with local communities and regulatorsHowever, from the operators' perspective, it's still more cost-effective to absorb those costs than to delay an AI project valued at billions of euros or dollars. Their priority is to avoid falling behind other tech giants in the competitive race.
In industry discussions, the word is starting to be used openly. “Desperation” to describe these types of decisionsIt's not so much chaotic improvisation as a forced logic: nobody chooses jet engines or diesel generators because they're clean or cheap; they're chosen because they're the only things that can be deployed in months, not years. Temporary solutions that, little by little, become almost permanent parts of the infrastructure.
Financial bubble and risk of energy bottleneck
The energy dimension of AI cannot be separated from its financial architectureLarge technology companies and businesses linked to the ecosystem are channeling billions of dollars in data center investment into special purpose vehicles (SPVs), financed by investment banks and private credit markets. It is estimated that more than $120.000 billion has been moved off-balance sheet to make official debt figures appear healthier.
This massive use of accounting engineering is a classic sign of bubble under constructionWhen an industry must continue to grow rapidly to maintain its future narrative and, at the same time, needs its financial indicators to remain stable, the incentive to conceal risks skyrockets. The danger doesn't disappear; it simply shifts to less transparent corners of the system.
Meanwhile, the race for secure dedicated energy sources It has become a geopolitical element. Large companies are buying or acquiring stakes in electric utilities and energy project developers to secure their supply. Alphabet, Google's parent company, has acquired Intersect Power for several billion dollars with the aim of guaranteeing energy, ideally clean energy, for its AI expansion.
Meanwhile, the return on investment (ROI) period for many AI projects is lengthening. Some analysts place the recovery of the huge capital investment beyond 2030This fuels a growing debate about whether we are facing yet another tech bubble. Skeptical and optimistic investors are exchanging arguments as money continues to flow into new infrastructure.
At this point, The energy bottleneck threatens to become the factor that sets the real ceiling for AI growth.If networks aren't reinforced quickly enough and temporary fossil fuel solutions begin to run up against regulatory, social, or climate constraints, simply adding more servers and chips won't suffice. The risk is that multi-billion-dollar projects will be underutilized due to a lack of affordable and socially acceptable electricity.
Renewables and microgrids: the alternative that is gaining traction
Faced with this scenario of jet and diesel engines running non-stop, a growing group of experts and companies is proposing another path: dedicated renewable microgrids, largely disconnected from the main gridThe idea is to build large-scale solar plants combined with storage and, if necessary, some fossil fuel backup, but with a much smaller global footprint.
A joint study by researchers from Stripe, Paces, and Scale Microgrids suggests that solar systems with around 44% renewable energy They would already be cost-competitive with the gas-fired power plants being built for data centers. And, according to their calculations, configurations with 90% renewable energy would outperform nuclear projects in profitability, especially if installed in areas with high solar radiation and abundant available land, such as certain southern states in the United States.
The great advantage of these microgrids is time. Large-scale solar parks can be built in less than two yearsThis is significantly shorter than the typical timelines for a traditional gas-fired power plant, a nuclear plant, or a large electrical interconnection. This timeframe aligns better with the speed at which new AI data centers are being deployed.
Some tech giants are starting to make moves in this direction. Google, for example, has strengthened its commitment to long-term renewable energy purchase agreements and strategic acquisitions for to ensure that, at least in part, the growth of AI is fueled by low-carbon sourcesHowever, the bulk of the sector still prefers diesel and gas due to technological inertia and a very earthly fear: that the cloud will "go out" if the sun doesn't shine or the wind doesn't blow.
For Europe and Spain, where renewable energy penetration is high and climate planning is stricter, The challenge lies in integrating these data centers into a system that is already in full transition.Opportunities are opening up to link new AI complexes to wind projects and dedicated solar power, or even future modular reactors, but that requires regulatory anticipation and coordination between the electricity and digital sectors.
A very physical AI infrastructure, with real costs
The public narrative surrounding artificial intelligence often focuses on spectacular demos, conversational assistants, and promises of productivity. However, the reality emerging behind the scenes is much more down-to-earth: an industrial infrastructure hungry for electricity, land, water and capital, largely sustained by 20th-century energy technologies.
While the "innovation" of AI is praised, in practice Direct pressure is being exerted on electricity grids and local communities that host large data centers. Environmental and social costs, from emissions to noise or water use for cooling, are frequently outsourced and left out of the spreadsheets shown to investors.
The paradox is clear: to power the most advanced software on the planet, technology companies They are reviving combustion engines and fossil fuel generators which many considered to have paid for themselves. These “bridging turbines” allow AI to continue growing in the short term, but leave open the question of how long a model that doubles energy costs and clashes with climate goals can be sustained.
In this context, the discussion about AI is increasingly shifting from the purely technological realm to that of the political economy of its infrastructureWho actually pays for the investments, how the risks are shared, where the energy comes from, and what footprint it leaves? The decisions made now, both in the United States and in Europe, will determine whether the next wave of data centers is built on jet engines and diesel or on a more sustainable foundation.
What we are seeing right now suggests that the artificial intelligence revolution depends not only on more sophisticated models, but on something much more prosaic: to have sufficient affordable and clean energyIf that part fails, the jet engines running at the foot of the data center will not only be a striking anecdote, but the symbol of a technological gamble that wanted to move too fast for what the physical and energy reality of the planet allowed.
