For years, the expansion of data centers followed a fairly predictable script: location was chosen, the building was constructed, electrical capacity was requested… and then it was just a waiting game. That last step — connecting to the grid — has become the bottleneck that is reshaping the sector’s energy map. In the United States, an increasing number of developers are opting for an alternative approach: building self-generation \”behind the meter\” (behind-the-meter, BTM) and practically operating as if they were \”off-grid\” (although many maintain partial interconnection or backup systems).
The figures circulating in the sector are no longer anecdotal. According to Cleanview project tracking, in 2025, 48 GW of BTM projects associated with data centers were announced, with approximately 33% of the new planned capacity expected to come from this approach. By the end of 2024, the volume in this category was less than 2 GW.
What does “behind-the-meter” mean and why is it booming now
BTM, in plain language, means that the data center doesn’t wait for the electricity grid to make space: it secures its power on-site — through gas, renewables, batteries, or combinations — and operates with less (or zero) reliance on the traditional supply.
The main trigger is time. Connecting a large data center to the grid can take years due to interconnection queues, substation upgrades, permits, and transmission expansions. In that context, BTM presents itself as a shortcut: if the business needs to come online quicker than the competition, developing its own energy infrastructure could mean the difference between opening in less than 2 years or being stuck with timelines close to 7.
Simultaneously, electricity demand driven by AI has changed the scale of the issue. It’s no longer just about growing “a little” each year: we’re talking about massive campuses, continuous loads, peaks, and rapid expansion. With that pace, the grid — and especially its bureaucracy — ceases to be just a provider and becomes a strategic constraint.
The uncomfortable truth: the quick fix is natural gas
In theory, the BTM model could promote renewables + storage and speed up decarbonization. In reality, urgency is pushing toward dispatchable and available sources: natural gas. Sector tracking indicates that 72% of announced BTM projects rely on gas.
The reason is pragmatic: a data center doesn’t just buy “kilowatt-hours,” it buys continuity. And gas, with turbines or generators, offers an immediate response to a 24/7 demand without depending on weather conditions. The paradox is clear: it “shortens” deployment times, but opens an environmental and reputational front that can haunt the project for decades.
Homer City: from coal to gas fueling the AI era
The most alarming example is Homer City in Pennsylvania: an old coal plant that has been retired and will be converted into a data center campus powered by a large gas plant. Publicized plans estimate a capacity around 4.5 GW, with an investment of approximately $10 billion, and construction starting in 2025 with commissioning in 2027.
This type of development has an immediate effect: unlocking capacity in a timeframe that the grid cannot match. But it also concentrates emissions and turns the energy debate into a local issue: permits, air quality, water, noise, community acceptance… and above all, whether digital infrastructure should grow “through a shortcut” or through planning.
A recurring pattern: microgrids to accelerate AI campuses
The phenomenon isn’t limited to isolated cases. In Texas, for example, the approach of powering new AI complexes with microgrids of gas generators to avoid connection delays has become apparent, aiming to speed up commissioning.
The logic remains the same: the grid isn’t arriving in time, the project can’t wait. As the industry begins to normalize this path, what once seemed exceptional — “a data center with its own power plant” — is turning into a design option.
What happens when AI infrastructure is filled with generation… and generation is filled with on-site power
If BTM gains traction, there will be consequences extending beyond climate debates:
- More complex grid planning: When large loads partially disconnect from the grid, utilities find it harder to forecast future demand and justify investments in transmission and substations. The system enters a vicious cycle: the grid is late, more actors disconnect, and the cycle worsens.
- Risks of \”energy archipelagos\”: campuses with self-capacity, governed by private logic, prioritizing availability over system efficiency. This can improve local resilience of the data center but may not necessarily enhance the overall electricity grid’s resilience.
- New regulatory tensions: How are emissions accounted for if consumption doesn’t pass through the grid? How is air quality monitored? What if a microgrid wants to sell excess power? Each question opens a different legal framework.
- Competition for equipment: turbines, generators, transformers, switches, batteries… The electrification of AI is also an industrial race, which could strain supply chains.
The solution: lowering storage costs and accelerating grid development (simultaneously)
The emerging consensus in the sector is quite pragmatic: connection speed is now more influential than the energy mix origin. This elevates storage (BESS) to a central role: if batteries and advanced management allow renewables to operate most of the time, natural gas could become a backup, not the primary source.
But equally important is a different perspective: the rise of BTM indicates that the grid needs urgent reforms — permitting, interconnection, investments in transmission, and mechanisms for deploying capacity within technological cycle timelines. Without these reforms, the “solution” will be each developer building their own system.
Ultimately, the question isn’t whether more data centers will be built: they will. The real issue is what kind of energy system will surround them: an integrated, planned one with shared metrics, or a fragmented mosaic of accelerated islands where gas becomes the wildcard of haste.