Energy Companies as Foundations of AI Infrastructure

The chief constraint on AI is no longer silicon, capital or ambition. It's electricity. The International Energy Agency projects that global data center power consumption will more than double by 2030 to roughly 945 terawatt-hours, about what Japan consumes in a single year. In the United States, a Lawrence Berkeley National Laboratory report for the Department of Energy estimates data centers could draw as much as 12% of the nation's electricity by 2028, up from 4.4% in 2023.

The grid can't deliver that power on the timelines AI demands. Interconnection for large new loads now runs four to seven years in the regions where demand is heaviest. Nobody racing to stand up AI capacity has that kind of time.

And the line keeps getting longer. A Reuters review found that requests from very large power users, mostly data centers, now exceed 700 gigawatts across the Midwest, Mid-Atlantic and the South. In Texas, large-load requests have climbed from about 48 gigawatts in 2023 to more than 474 gigawatts, over five times the record peak demand on the Texas grid. The surge prompted Gov. Greg Abbott in August to pause new data center connections pending an audit that asks, among other things, which projects are real and which are bringing their own power. BloombergNEF (BNEF) has put a number on what the bottleneck costs. AI chip shipments imply 207 gigawatts of new U.S. data center capacity by 2033. BNEF expects 63 gigawatts of that will never get built because the electricity won't be there.

The mismatch has redefined who matters in AI. Energy companies can step into a role few of them planned for: laying the foundation on which AI infrastructure gets built. They own natural gas at its point of origin, they control the land above it and they've run critical infrastructure around the clock for a century.

Our conviction is simple. You cannot call yourself an AI-first company if you're not an energy-first company, because AI runs on enormous amounts of electricity. The companies that are already energy-first hold the cards.

Why the action has moved "behind the meter"

A behind-the-meter power project generates electricity alongside the load it serves and never touches the utility grid. No interconnection queue. No waiting on a utility to weigh an enormous new AI load against its obligations to residential and commercial customers. A grid-connected data center can wait most of a decade for approvals. An energy company can bring dedicated generation online in roughly 18 to 24 months.

Energy companies did not stumble into this advantage. They own the inputs: natural gas at the point of origination, inexpensive land far from congested metros, produced water for cooling, plus the turbines and reciprocating engines that convert gas to electricity without a utility in the loop. And they already run hazardous, always-on infrastructure at industrial scale, exactly the operating culture a gigawatt AI campus requires.

The proof is already under construction. In June, Chevron signed a 20-year agreement to deliver 2.67 gigawatts of behind-the-meter power to a Microsoft data center in West Texas, a project estimated at $7 billion to $9 billion, with power generation expected as early as 2028. ExxonMobil and NextEra Energy are partnering on a 1.2-gigawatt gas-fired plant dedicated to data centers and designed to capture more than 90% of its emissions. Williams has secured $5.34 billion from Blackstone, Apollo and KKR to fund five behind-the-meter power projects, including two off-grid facilities in Ohio to serve a hyperscale campus. North of the border, Pembina Pipeline and its partners took a final investment decision in July on the Greenlight Electricity Centre, a C$4.6 billion, 932-megawatt behind-the-meter gas plant that will power Meta's first Canadian data center in Sturgeon County, Alberta.

These are not one-off deals. BNEF counts 124 gigawatts of announced on-site gas capacity being built for U.S. data centers, and by S&P Global's assessment, gas accounts for more than 80% of the announced behind-the-meter pipeline. BNEF's own math explains why. The greatest new data center load the U.S. grid has ever connected in a single year is about 10 gigawatts. If that pace holds, at least 48 gigawatts of on-site gas has to be built by 2035 just to meet BNEF's base-case forecast. The market is choosing gas because it can be permitted, delivered and financed within the window AI demands.

There is an economic story underneath. In gas-rich basins, production has outrun pipeline capacity so badly that prices at the Waha Hub in West Texas' Permian Basin settled negative a record 49 times in 2024 and stayed negative for 25 consecutive days in March of this year. Producers are paying to give gas away. Midstream operators are now evaluating data centers sited directly at compressor stations, converting stranded molecules into compute revenue without waiting for new pipelines. A takeaway problem becomes a new business.

From selling molecules to selling compute

Two years ago, most energy operators saw their role in this story ending at power origination. Today, the ambition runs much further up the stack: land and site development, shell construction, cooling, connectivity and, for the boldest, compute tenancy and GPU capacity sold as a service. Every step up the stack means billions in capital and billions in potential revenue.

The buyers are multiplying, too. BNEF expects non-hyperscaler capacity to reach 138 gigawatts by 2035, more than double the 57 gigawatts it projects for Microsoft, Meta, Alphabet and Amazon combined, as neoclouds, enterprises and new entrants build their own giga-campuses. Of the 100 largest announced U.S. projects, 52 come from companies that have never built a data center (which BNEF flags as execution risk). Together, they account for 87 gigawatts. For operators whose core business is already building and running large industrial assets on schedule, this should be read as an opening.

The defining strategic question in the industry right now is: How far can we climb?

Consider what happened in West Virginia. Fidelis New Energy set out to build the Monarch campus, a site with a power runway beyond 8 gigawatts and the country's first state-certified AI microgrid. In March, AI hyperscaler Nscale acquired the company behind the project and signed a letter of intent with Microsoft for 1.35 gigawatts of compute there. The compute buyer decided it needed to own the energy asset.

In Alberta, the climb is running the other way. Tourmaline, Canada's largest natural gas producer, is proposing to build Malachite One beside its Banshee gas plant west of Edson: a 930-megawatt power plant and up to 7.1 million square feet of data center space, with construction starting in 2027 and first operations targeted for 2028. The producer plans to own the data center itself, no hyperscaler required. Whichever direction the deal runs, the conclusion is the same: controlling AI economics means controlling both the power and the compute.

Energy and AI infrastructure are becoming the same business.

Oilfield services companies are making the same pivot. SLB has shipped more than 1.3 gigawatts of prefabricated modular data center infrastructure since April 2024 and expects to exceed 2 gigawatts by the end of this year. Halliburton's joint venture with VoltaGrid has secured manufacturing capacity for 400 megawatts of modular power systems. These are drilling companies selling data center capacity, something unthinkable five years ago.

The winners will be the operators who can answer one question with a plan, not a vision: How do we bring power to market and stand up usable GPU capacity first? This market rewards an 18-month execution window. It won't wait four to seven years, and it won't reward companies still deciding what they want to be when they grow up.

The same convergence is happening inside the fence line

Gigawatt campuses get the headlines. A quieter shift is underway inside energy operations themselves, and it explains why IT and OT can no longer live in separate worlds.

A decade of digital transformation put sensors on nearly everything: compressors, pumps, wellheads, process units, substations. A single operation can now generate terabytes of machine data every day. The original plan, pushing it all to the cloud for analysis, broke down on economic grounds. Moving and processing that much data can cost tens of millions of dollars a year, and most of it is never used. The data is also worthless without context. A pressure reading means nothing until it is tied to an asset, a facility and an operating state, which is why so many AI initiatives quietly turn into data cleanup projects. It's why energy companies are increasingly hiring chief AI officers: to direct traffic among hundreds of proposed use cases and stop the ones the data cannot yet support.

The answer? Bring AI to the data instead. The edge in oil and gas lives wherever data is born, usually at a sensor on a compressor far outside any IT network and beyond reliable cloud reach. Ruggedized edge systems with onboard GPUs, hardened for extreme heat and hazardous locations, became economical only in the past year. Models triage the terabytes locally, flag the vibration signature that precedes a failure and send only the valuable megabytes upstream. Predictive maintenance, computer vision, digital twins and AI copilots for drilling engineers all follow from that architecture.

Security is the underlay, not a line item

Both halves of this story expand the attack surface. Every new behind-the-meter campus is new critical infrastructure that must be defended. And IT/OT convergence, by design, opens doors into operational environments that were built as sealed bubbles precisely so nothing could reach them.

OT security is not IT security with different hardware. The priorities invert. In operations, availability and safety come first. The patch-and-scan reflexes of enterprise IT can halt a process, stop revenue and put people at risk. Control systems are certified as integrated stacks, so they often run for 20 years. Upgrades happen only during turnarounds planned three years in advance, and each one takes a facility offline for roughly a month. A refinery runs on the order of a hundred systems operating in tandem. Take one down and the plant stops, at a cost measured in millions of dollars per hour. And failure in these environments is measured in injuries, not tickets.

The industry has already seen how this plays out. The Colonial Pipeline shutdown in 2021 began as ransomware on the IT side. Operators ended up halting fuel to much of the East Coast because they couldn't be certain the operational systems were safe. Halliburton's 2024 ransomware incident cost the company roughly $35 million. As AI opens more doors between enterprise and operational systems, the cost of not knowing grows.

This is why we push our clients toward what we call "operational trust," a chain of custody from sensor to decision. Is the asset validated? Is the data traceable? Can every action be audited? Does what the model reports make physical sense? AI in operations should begin as an observer, with humans on the loop (not just in the loop), acting as a physical check. AI should never hold decision rights over safety instrumented systems. Trust is earned in increments. 

Companies that deliberately build this discipline will scale AI faster than those that blindly trust or refuse to trust at all.

Where to start

Two steps, one tactical and one strategic. The first makes AI safe and usable inside your operations. The second decides what role you will play in AI infrastructure. Start the first now, and don't let it delay the second.

Step 1 (Tactical): Make AI safe and operational inside the fence line

Begin with visibility. Asset discovery, segmentation and secure remote access are proven, fast to deploy and a prerequisite to everything else, because every downstream move gets safer and cheaper once you can see and control what is connected. Then build the data foundation: contextualize operational data and push compute to the edge so AI can scale past pilots. Let AI start as an observer, with humans on the loop, and extend its role only as each link in the chain of custody is proven.

Step 2 (Strategic): Choose how far up the stack to climb

Whether you stop at behind-the-meter generation or climb toward GPU capacity, make it an explicit capital allocation decision backed by a first-to-market plan, not a reaction to whoever calls next. This is where you answer the question that opened this piece with a plan rather than a vision: how do we bring power to market and stand up usable GPU capacity first? The operating discipline from Step 1 is what makes that plan credible to a tenant who is trusting you with critical infrastructure.

Energy has powered every industrial revolution. This time, energy companies are doing more than fueling the machines. They're becoming the platform the machines stand on. WWT works with 12 of the 20 largest energy organizations in the U.S., helping operators secure OT environments, build edge and data architectures and turn AI infrastructure strategy into something they can test, validate and execute. The foundation is yours to build.