Skip to content
Industry inflectionsAESIBEFLNCPOWLWMB

Onsite Gas Power Generation Adds Storage and Distribution to Manage AI Load Swings

Editorial illustration for Onsite Gas Power Generation Adds Storage and Distribution to Manage AI Load Swings
Published 5 min read

Summary

Atlas Energy Solutions and Williams say AI load swings force onsite gas plants to add storage and power distribution, lifting Atlas's capex above $2 million per MW.

AI data center load swings are changing how onsite gas power plants are built. On their third-quarter earnings calls on November 3, 2026, Atlas Energy Solutions (AESI) and The Williams Companies (WMB) both said the hardest part of powering an AI campus directly with gas generation is keeping up with fast swings in AI electricity demand. Atlas also said capital spending for these plants now starts above $2 million per megawatt.


Why AI load swings force onsite plants to add equipment

A growing number of AI data centers are not waiting for grid upgrades. Instead, developers build gas-fired plants next to the campus and supply power directly, an arrangement known as behind-the-meter power because the electricity never passes through a utility meter. Atlas supplies frac sand (proppant) and also develops these plants. Williams operates US natural gas pipelines and has built an onsite power project for an AI customer called Socrates. These plants were first designed mainly around gas engines or turbines, but AI training and inference can push power demand up and down sharply within a short time, and generators alone struggle to follow.

Atlas management said that when the industry started using distributed power to supply data centers 12 to 18 months ago, operators quickly found the original designs "were not able to handle the load swings," so they added a significant amount of expensive load-management equipment [1]. That equipment includes battery storage and power distribution centers; at the plant Atlas is building, the battery storage pad and the power distribution center foundations are already complete [1]. Every developer supplying an AI campus faces the same demand profile, so the design change may spread to similar projects.


Onsite gas power capex per MW rises from about $1.6 million to above $2 million

Atlas is the company that put numbers on the cost. On August 4, Atlas said its Socorro project was a 120 MW behind-the-meter power contract with total project capital of about $190 million [2], which works out to roughly $1.58 million per megawatt. On November 3, Atlas's CFO said the company had previously guided to a range of $1.5 million to $2.5 million of capex per megawatt; the low end of that range now starts at $2 million and continues to climb, and generators account for only 30% to 40% of total project capex [1]. He added that the extra cost passes through into a higher price per megawatt [1].

Williams described the same challenge from the operating side. On August 4, the company said it did extensive load testing before Socrates started up to confirm the plant could follow AI load as intended [3]. On November 3, its COO said these projects are complex, some of them first of their kind, require working with customers to manage load following, and are performing as anticipated [4]. Taken together, the two companies show that following AI load has become a core design requirement for onsite plants, but only Atlas has quantified the rise in cost per megawatt.


Most onsite plant spending now goes beyond the generators

By Atlas's account, 60% to 70% of an onsite plant's capital goes to items other than generators, including storage, power distribution, controls and spare capacity [1]. Developers can protect returns through a higher price per megawatt, but each project needs more capital, and AI customers pay more for the same amount of power. Equipment spending may therefore shift toward switchgear and battery storage suppliers, and power technologies that can follow load on their own may be able to charge for that capability.

This view has a clear limit. Williams's CEO said that as operators learn load profiles over time, they may be able to carry less redundancy and use spare capacity more efficiently [4], so part of the extra spending may reflect conservative early designs. Two things to track next are whether Atlas's capex per megawatt keeps rising on new projects and whether Williams's new projects carry less redundancy.


Companies exposed to this change

  • Powell Industries (POWL): Makes switchgear and power control systems. Its CEO said behind-the-meter power islands raise Powell's equipment content per project, and on August 4 the company disclosed a behind-the-meter onsite generation order for a data center worth more than $400 million [5]. That maps to the power-distribution share of non-generator plant spending, but Powell did not tie the order to load swings.
  • Fluence Energy (FLNC): Supplies battery storage systems. On August 6 it disclosed a $300 million behind-the-meter order with a data center developer and said its data center pipeline had grown to 16 GWh [6]. Storage is the equipment onsite plants add to manage load, but Fluence did not say this storage was sized to follow AI load.
  • Bloom Energy (BE): Provides onsite fuel-cell power. On July 28 its CFO said customers pay for capabilities such as following the load profile of an AI campus, and that its pricing reflects this [7]. It may be able to price in a capability that gas plants must buy extra equipment to achieve.

Sources

[1] Drillr · Atlas Energy Solutions (AESI) · 2026-11-03 · FY2026 Q3 earnings call

previously we've talked about it being anywhere from 1.5 to 2.5 million dollars per capex per megawatt. That's probably trended at a low level above $2 million. Just the low end of that range now starts at the $2 million, and it continues to climb, but obviously that passes through into incremental price per megawatt.

[2] Drillr · Atlas Energy Solutions (AESI) · 2026-08-04 · FY2026 Q2 earnings call

[3] Drillr · The Williams Companies (WMB) · 2026-08-04 · FY2026 Q2 earnings call

[4] Drillr · The Williams Companies (WMB) · 2026-11-03 · FY2026 Q3 earnings call

[5] Drillr · Powell Industries (POWL) · 2026-08-04 · FY2026 Q3 earnings call

[6] Drillr · Fluence Energy (FLNC) · 2026-08-06 · FY2026 Q3 earnings call

[7] Drillr · Bloom Energy (BE) · 2026-07-28 · FY2026 Q2 earnings call


This article highlights industry changes and companies that may be overlooked. It is not a stock recommendation.

Want deeper analysis?

Ask drillr anything about AESI, BE, FLNC, POWL, WMB — powered by SEC filings, earnings calls, and real-time data.

Try drillr.ai for free

drillr can make mistakes. Information only — not investment advice. Learn more