In the collective imagination, artificial intelligence is an ethereal cloud of algorithms. The reality is much more complex and what we know for sure is that an energy eater that needs to “eat” constantly. Satya Nadella, CEO of Microsoft, has summarized with unusual crudeness: “The problem is no longer that it is missing Nvidia chips, but that there are not enough plugs.”
And so that these plugs have power 24 hours a day with the 99.999% reliability that the sector demands, Big Tech has ended up looking where no one expected: thousands of meters below the ground, towards the salt caverns.
When the bits hit the underground. The AI race has entered a “slow start” phase in the construction of these underground caverns, which could hinder the rollout of data centers. According to Fortunethe reason is mathematical since these digital infrastructures do not tolerate interruptions and require extreme reliability.
To guarantee this constant flow, natural gas has become the indispensable backup. However, as they explain, it is not enough to produce gas; you have to save it. Industry projections indicate that only about half of the storage that will be needed to meet future demand has been planned. Without these artificial caves dug thousands of meters below the surface, hyperscalers (Google, Amazon, Meta) are left at the mercy of gas pipelines, vulnerable to corrosion, landslides or extreme weather events.
But why salt caverns? The technical answer lies in flexibility. As detailed by experts in Fortunethere are two ways to store gas: in depleted oil fields or in salt caverns.
The former are cheaper, but structurally slow. The gas is injected in summer and extracted in winter, following a classic seasonal cycle. AI, on the other hand, does not understand seasons. Their demand peaks are constant, sudden and difficult to predict. The salt caverns, created by injecting water to leach the mineral, act as a high-pressure lung: they allow gas to be injected and extracted with a much higher frequency, adapting to the volatility of the electrical grid that powers the servers.
The “supercycle 2.0”. Given this scenario, companies like Enbridge they have taken the lead. Greg Ebel, CEO of the company, has confirmed that they are expanding their facilities in Egan (Louisiana) and Moss Bluff (Texas). “This demand dramatically changes the economics of supply,” he said.
But it is not enough. Jack Weixel East Daley Analytics analystwarns that double the capacity currently planned is needed. Projects such as the Freeport Energy Storage Hub (FRESH), in Houston, They seek to connect up to 17 gas pipelines to a new salt dome by 2028, but construction times—often exceeding four years—clash with the urgency of AI.
For his part, Jim Goetz, CEO of Trinity Gas Storage, defines it as the “storage supercycle 2.0”. His company has just reached the final investment decision (FID) to expand its capacity in East Texas, seeking to support critical infrastructures such as Stargate, the titanic $500 billion project from OpenAI and Microsoft.
The shadow of a doubt. The underlying question is not only whether the salt caverns work—they work—but what type of energy system they are consolidating. Natural gas is fast, flexible and reliable, but it also introduces new dependencies and risks. According to analystsgas infrastructure on the Gulf Coast is especially vulnerable to extreme weather events. A direct hurricane over Texas or Louisiana can disrupt production, exports and transportation at the same time. In that scenario, even with gas available in other regionsthe lack of nearby storage can leave data centers without electrical backup.
Added to this is the question of price. The sustained increase in demand to fuel data centers, LNG exports and reindustrialization is already pushing up gas and electricity bills. Without enough storage capacity, that volatility is amplified. As the sector points out, storage acts as a buffer; when it is missing, the peaks transferred directly to the consumer. Furthermore, the criticism is more structural since AI is pushing to prolong dependence on fossil fuels just when governments and companies were committed to reducing it.
Look beyond the gas. Aware of this physical limit, large technology companies are no longer looking only at salt caverns and gas pipelines. They look for any firm source of electricity that does not depend exclusively on the traditional energy market.
An example is Fervo Energy, a geothermal startup that has just closed one of the largest financing rounds in the sector, with Google as an investor and client. His commitment to advanced geothermal —constant electricity 24 hours a day—reflects the extent to which AI is redrawing the energy map. This is not an immediate or universal solution, but it is a clear signal: the problem is no longer technological, but energy-based.
A problem only in the United States? The United States is the epicenter, but not the only scenario. The clash between AI and energy is global, although responses vary. In Europe, the rise of AI is leading to rethinking the closure of gas and coal plants. Some electricity companies are negotiating to convert old plants into data centers, taking advantage of their access to the network, water and already depreciated infrastructure. The logic is the same: firm, immediate and available energy.
China, for its part, has chosen another path. Beijing not only promotes underwater data centers either large energy clusters in interior provinces, but directly subsidizes the electricity that powers its AI. The objective is to reduce the “fuel” of digital models and compensate for the lower energy efficiency of national chips compared to those from Nvidia.
The return to the underground. In all cases, the pattern repeats itself. Renewables are growing, but not fast enough or with the stability necessary to sustain the demand for AI in the short term. Gas – with salt caverns, temporary turbines or recycled plants – becomes the inevitable crutch.
In our race to create an intelligence that lives on the plane of ideas, we have ended up returning to mining, drilling, and the depths of the Earth. The future of AI may not be decided just in laboratories or data centers, but in something much less visible: who controls the underground that keeps your plugs on.
GIPHY App Key not set. Please check settings