Natural gas for data centers is one of the biggest stories in the gas business, and most of the work is commercial. Before a molecule becomes a megawatt, someone buys it, moves it through a pipeline, and delivers it to a turbine on a schedule a hyperscaler set. On Wednesday, November 11, I join the technology panel at the NatGas To Power Forum in Austin, "From Molecules to Megawatts: Transforming the Natural Gas Commercial Value Chain Through Innovative Technology." This post is my view of where AI fits in that chain, and a look inside the agent I built for the panel. It is written for the producers, pipelines, utilities, power generators, engineering and construction firms, developers and operators in the room.
Key takeaways
- Texas shows the scale. ERCOT is tracking about 474 GW of large loads seeking to connect, about 90% of them data centers (as of June 2026), against an all-time demand record of 91,134 MW set July 22, 2026.
- Gas capacity in ERCOT's generation queue is up 85.5% in a year, to 81,033.87 MW in September 2026, and EIA expects U.S. power plants to burn 37.0 Bcf/d of natural gas in 2026, up from 35.8 Bcf/d in 2025.
- The hard part is commercial: firm supply, pipeline capacity, turbine lead times and contracts.
- The platforms on this panel already run the inside of the chain, from nominations to pipeline notices, and some already ship AI agents with a person approving.
- The open lane is outside public data: state air permits for new gas turbines, ERCOT's interconnection queue and state large-load rules, checked and handed to a person.
- Accuracy is the design problem. In the agent I built, Python does the math, AI writes the words, a checker verifies every number, and a person decides.
Why Natural Gas for Data Centers Became the Default
Texas shows the scale. ERCOT told the Texas Senate in July that it is tracking approximately 474 GW of large loads seeking to connect to its grid, about 90% of them data centers (ERCOT, figures as of June 2026). On July 22, 2026, ERCOT set an all-time demand record of 91,134 MW, unofficial until final settlement. In September, ERCOT's board was told that 204 large-load projects totaling 66.4 GW had qualified to be conditionally included as base load in its planning.
The state is tightening the gate at the same time. After Gov. Greg Abbott's August 3 directive ordered an audit of proposed data centers, ERCOT paused approvals to energize new data centers and crypto mining loads of 75 MW or more, pending audit reports due December 10. And the Public Utility Commission of Texas's new large-load interconnection rule, 16 TAC §25.194, effective October 8, requires financial security of $50,000 per MW before a large load is included in an ERCOT interconnection study.
Nationally, the International Energy Agency's Key Questions on Energy and AI (April 2026) sees data center electricity use roughly doubling worldwide, from 485 TWh in 2025 to 950 TWh in 2030, and the U.S. Energy Information Administration (EIA) estimates that servers alone used about 7% of U.S. commercial-sector electricity in 2025.
Natural gas carries more of that load than any other fuel. It was the top source of U.S. utility-scale electricity in 2025, about 41%, according to the EIA, and the IEA puts it at over 40% of the electricity U.S. data centers use. EIA's October 2026 Short-Term Energy Outlook has U.S. power plants burning 37.0 billion cubic feet per day of natural gas in 2026, up from 35.8 Bcf/d in 2025, and the IEA calls natural gas the largest source of additional supply for U.S. data centers through 2030.
Data Centers That Build Their Own Power, Behind the Meter
A growing share of natural gas for data centers never touches the grid. In a behind-the-meter setup, gas turbines or reciprocating engines sit on or next to the campus and serve it directly. The IEA estimates that around 15 to 27 GW of on-site natural gas generation may power data centers by 2030, mostly in the United States, though the IEA notes that most data centers still prefer a grid connection.
- Williams, Socrates (New Albany, Ohio). Two power facilities with a combined 400 megawatts, built for an affiliate of Meta and designed to run without a physical connection to the grid. Williams expanded the scope to about $2 billion and reported phase one complete in its second-quarter 2026 results, with phase two on track for the fourth quarter.
- Chevron and Microsoft, Project Kilby (West Texas). A 20-year power purchase agreement for about 2.67 GW, mostly from GE Vernova gas turbines, with first power targeted for 2028, as reported by World Oil.
Each one is a supply contract, a transport arrangement, a turbine order, and a construction schedule before it is a single megawatt.
The Commercial Value Chain, From Molecule to Megawatt
Every link in the panel's chain runs on information.
| Link in the chain | What the work involves | Where data center demand adds pressure |
|---|---|---|
| Supply and procurement | Buying gas, long-term supply deals, price risk | Firm, around-the-clock supply for a load that cannot go down |
| Transport and capacity | Pipeline nominations, firm transport, laterals | New laterals and expansions on a developer's timeline |
| Storage | Injections, withdrawals, balancing | Backing up a load with very little tolerance for a miss |
| Power generation | Turbines, engines, plant operations | Turbine lead times and on-site plant construction |
| Delivery to the load | Interconnection, metering, contracts | Behind-the-meter designs that sit outside the usual utility model |
Pipelines are already building for it. Kinder Morgan's third-quarter 2026 results describe the largest natural gas project backlog in the company's history, at $13.2 billion, within a $14.4 billion total backlog that is more than 75% tied to power generation and local distribution company demand, including a lateral backed by take-or-pay agreements with data center and industrial customers. In the same report, Kinder Morgan's Natural Gas Pipeline Company of America took a final investment decision on its Permian Link project, about 2.8 Bcf/d of new transportation capacity that "will serve new power generation for an A-rated hyperscaler," with service targeted for the third quarter of 2030.
Where AI Fits: Nominations, Forecasting, Capacity, and Gas Flow
The panel's first topic is AI and predictive analytics for pipeline nominations, forecasting, capacity management and gas flow optimization. Much of this work is already well served, by companies on the same stage.
What the platforms on this panel already do
- Nominations and scheduling run every day on the platforms of NatGasHub, Quorum, Trellis and nGenue, along with confirmations and imbalances.
- Pipeline notices. NatGasHub's gNOTICE already uses AI to scan incoming pipeline critical notices, alongside gNOM for nominations and gAGENT, its agentic AI service.
- Forecasting, capacity and gas flow. Load forecasting, capacity release, imbalance tracking and measurement sit inside these platforms, next to the deals they depend on.
- Agents with a person in charge. Quorum's agents work inside existing workflows and make recommendations a person can approve, modify or override.
AI agents with a person approving are not new in this industry. They are shipping.
What those platforms don't watch
Those platforms are built around a company's own deals, nominations, contracts and measurements, and around what the pipelines post. Future demand for gas shows up first in public filings that sit outside the commercial systems:
- State air permits for new gas turbines, each naming a company, a county and a megawatt figure before the plant needs its first nomination.
- ERCOT's interconnection queue, a monthly report on gas generation seeking to connect to the Texas grid, by study stage, capacity, county and projected online date.
- State large-load rules, written in public projects at the Public Utility Commission of Texas.
These files are free, and messy: confidentiality rules, layout changes, project names that shift from month to month. That is where I think AI earns its next place: read the outside signal, check it, and hand it to a person who takes it into the platform the team already runs, as a supply conversation, a capacity question, a forecast input or a regulatory deadline.
What I Built for Austin: the Texas Data Center Gas Watch
To test that idea, I built an agent for this panel. Each month it reads ERCOT's interconnection queue, the Texas Commission on Environmental Quality (TCEQ) gas turbine air permit list and five Public Utility Commission of Texas (PUCT) large-load projects, works out what changed, writes a brief, checks every number, and stops for a person.
The brief reports what the state's files show and nothing more: the company, the county, the megawatts and the dates.
Make.com runs it on a schedule:
| Step | Runs on | What it does |
|---|---|---|
| 1. Watch | Python | Checks ERCOT's public list of monthly Generation Interconnection Status (GIS) reports for a new one |
| 2. Skip repeats | Make.com data store | Stops if that report was already briefed |
| 3. Read and do the math | Python, no AI | Downloads this month's and last month's ERCOT GIS spreadsheets, TCEQ's pending and issued gas turbine air permits, and filings from PUCT Projects 58317 (Senate Bill 6), 58479, 58480, 58481 and 58482; does all the math |
| 4. Draft | Claude (Sonnet) | Writes the wording only, from those facts, copying every number exactly |
| 5. Check and format | Python | Confirms every number in the draft appears in the source facts, sets "Attention" by fixed rules, builds the email |
| 6. Stop | Email to a reviewer | A person replies RELEASE or HOLD; RELEASE sends the commercial team a hand-off checklist: supply, pipeline capacity, regulatory deadline, owner confirms |
The Python uses only the standard library and reads each spreadsheet straight from its XML, so the same code runs on a laptop and inside Make. The AI never does arithmetic. "Attention" is not an AI judgment either: Python sets it by fixed rules: a gas turbine permit application of 500 MW or more, a project of 500 MW or more appearing in or leaving ERCOT's public list (splits and re-filings are recognized and excluded), a 1,000 MW move in the gas in ERCOT's queue, and any PUCT order. If a number in the draft is not in the source facts, the subject line says the check failed. In the latest test on ERCOT's September report, all 115 figures matched.
The checker is strict about numbers, not meaning. It proves each figure came from the source; the reviewer confirms each figure sits beside the right project. Code for the arithmetic, a person for the judgment. There is no live connection to anyone's software: the person takes what matters into the platform the team already runs.
What the September run found
- Gas in ERCOT's interconnection queue is up 85.5% in a year. ERCOT's own trend table shows 43,695.24 MW of gas in the queue in September 2025 and 81,033.87 MW in September 2026. September was also the first monthly decline after 12 straight increases: down 670.09 MW from 81,703.96 MW in August.
- ERCOT's public project list does not add up to ERCOT's own total. ERCOT's note on the list: "Due to Protocol confidentiality provisions, only those projects for which a Full Interconnection Study has been requested are included." September's public list held 79,075.8 MW of gas. The rest is 1,899 MW in the confidential early study stages and 59.07 MW of small generators: 79,075.8 plus 1,899 plus 59.07 is 81,033.87 MW.
- A 3,000 MW project that was most likely not new capacity. Brazos Energy Security (Fisher County, 3,000 MW, combined cycle) appeared in the public list for the first time. The same month, confidential early-stage combined-cycle gas fell by exactly 3,000 MW, from 4,708 MW to 1,708 MW. ERCOT never names confidential projects, so the match is an inference, but it is the one that balances the month.
- A split, not an addition. Blue Energy Center Phase 1a (Victoria County, 780 MW, combustion turbine) appeared in the same month that 29INR0248, same county and same developer, went from 1,110 MW to 330 MW. 780 plus 330 is 1,110: no net change.
- A re-filing. Wolf Bone North Energy Center (Reeves County, 1,174.4 MW) appeared after its earlier request, 29INR0162 (1,204 MW), was canceled September 4. The net change is down 29.6 MW.
- Every online date that moved, moved later. All 20 gas projects that changed their projected commercial operation date in September pushed it later.
- TCEQ's pending list (version 2026.10.7) shows 21 gas turbine air permit applications totaling 17,249.78 MW, though 2 of those permit numbers also appear on TCEQ's issued list. New in the window: Gulf Nexus I LLC, 500 MW, Wharton County, received September 23. The list does not say what any plant will serve.
- A new rule. The PUCT adopted 16 TAC §25.194, its large-load interconnection standards, by order on September 18, 2026, in Project 58481, effective October 8.
Read month over month without ERCOT's stage data, the report says 3,000 MW of new gas appeared in Fisher County. Read with it, a project most likely changed stages. The fixed rule flagged Brazos, as it should, and the same brief showed the matching 3,000 MW drop in the confidential stages so the reviewer could see both. That is why Python does the math, a checker verifies every number before a person sees the brief, and the person decides what it means. It is the difference between a sales call built on a real project and one built on a spreadsheet artifact.
Connecting AI to ERP, ETRM, CRM, and SCADA
The panel's second topic is system integration, where AI projects in this industry tend to succeed or stall. The room already runs serious platforms: enterprise resource planning (ERP) for finance and procurement, energy trading and risk management (ETRM) for deals and positions, customer relationship management (CRM) for counterparties, and supervisory control and data acquisition (SCADA) for operations.
- The first connection can be a person. For outside signals that arrive monthly, a reviewer who enters what matters as a deal, a prospect, a forecast input or a deadline is often the right design, not a shortcut. It also keeps outside content out of your systems of record.
- When AI touches a business system, start read-only and narrow. Give the AI its own account that reads only the records the task needs, logs what it reads and cannot send data out. AI drafts the summary. The record's owner makes the change, and anything that writes, sends or deletes waits for a person's approval.
- Treat SCADA differently. AI can analyze a copy of historian data sent out of the control network through the DMZ, the buffer network between operations and the business side, one-way where possible, and recommend. AI assistants and agents get no path to send commands or change setpoints. AI built into a vendor's control product runs inside engineered limits, and operators can switch it off and fall back to normal automation or manual control. AI advises. Operators and engineers decide.
- Write the data rules down. Use only company-approved AI tools whose contracts spell out storage, retention and no training on your data. Contracts and counterparty terms stay inside those tools. Network diagrams and control-system details need security team approval first, and where NERC CIP or TSA pipeline security rules apply, AI access gets the same approval, review and logging as any other account.
That is the line drawn in the joint CISA, NSA and FBI guidance on AI in operational technology, which asks operators to limit active control of OT infrastructure by AI without a human in the loop, and to move data out of OT without granting persistent inbound access.
The Bottlenecks in Delivering Natural Gas to Data Centers
The panel's third topic is the commercial and operational bottlenecks in serving booming data center markets. Four stand out:
- Turbines. Gas turbines for data centers are the tightest link in the chain. The IEA notes a 70% surge in gas turbine orders in 2025, which has highlighted chokepoints in the supply chain. GE Vernova reported its gas power equipment backlog and slot reservations at 122 GW in its third-quarter 2026 results, up from 116 GW a quarter earlier, and expects at least 130 GW by year-end. Siemens Energy's chief financial officer has said lead times across the company now run three years or more.
- Pipeline capacity. New supply needs firm transport and, often, new laterals, built on schedules that rarely match a developer's.
- Permits and policy. Chevron's vice president for power solutions told Semafor in October that the final investment decision on Project Kilby could slip into 2027, after Gov. Greg Abbott's September directive halting state permits for data centers until ERCOT completes its audit.
- Contracts. Firm supply, take-or-pay transport, and power purchase agreements have to line up with a construction schedule that keeps moving.
AI does not build a turbine or lay a pipe. It can make sure the people managing these constraints see the permit, the queue change or the new rule the month it appears.
Six AI Workflows Built on Outside Public Signals
Each reads public data the commercial platforms are not built around and hands a checked result to a person.
| Team | Outside signal the AI reads | What a person decides |
|---|---|---|
| Gas supply | New gas turbine air permit applications in the counties you supply | Whether to open a supply conversation |
| Pipeline commercial | Interconnection queue entries, splits and cancellations near your system | Whether a lateral or expansion is worth scoping |
| Scheduling | Already runs on the platforms on this panel; the outside add is which new plants are coming, and where | Which future delivery points to plan for |
| Forecasting | New load in permits and the queue, before it appears in any history | Whether to change the long-range demand view |
| Regulatory | Orders and filings in state large-load rulemakings | What the company files or says, and by when |
| Business development | The same permits and queue entries, read as leads by engineering, construction and power firms | Who to call first |
Pick one, give it an owner, and measure it against how the work is done today.
The Austin Panel: Details
"From Molecules to Megawatts: Transforming the Natural Gas Commercial Value Chain Through Innovative Technology" runs Wednesday, November 11, 2026, from 10:15 to 11:45 AM Central at the JW Marriott Austin, as part of the NatGas To Power Forum, November 11 to 13. Cleve Hogarth, Principal of Cleveland Advisory, moderates. The panel: Brandon Penksa, Solutions Architect Director at Quorum Software; Jay Bhatty, CEO of NatGasHub.com; Jason Langella, Founder and Chairman of SEO Agency USA; Scott Vogan, SVP of Sales and Marketing at nGenue; and Dhruv Venkatraman of Trellis Energy Software.
Thursday's panels cover the pipeline infrastructure behind the data center boom, with speakers from Kinder Morgan, Sequent Energy Management (Williams), Entergy, NRG Energy, and Enbridge, and the commercial race to deliver gas to data centers, moderated by Jay Bhatty. Friday opens with Robert Bryce's keynote, "AI, Gas, And The Grid."
What SEO Agency USA Builds for the Data Center Ecosystem
SEO Agency USA is a sponsor of the NatGas To Power Forum and of all six 2026 LDC Gas Forums. We build three things for the companies that power and build the data center economy: gas producers, pipelines, utilities, power generators, engineering and construction firms, developers and operators.
- AI agents like the Gas Watch: public data in, math in code, every number checked, a person in charge. See AI agent development.
- Automated workflows, the scheduled steps around an agent, from the first download to the hand-off checklist, built on tools your team can inspect. See AI workflow automation. An AI strategy and readiness assessment picks the first workflow and its owner.
- Websites. When a developer's site-selection team, or its AI assistant, looks for gas, power or construction partners, your website decides whether you make the list. We build websites for that search, and the AI visibility work that gets them cited.
Visit the SEO Agency USA kiosk at the JW Marriott throughout the forum, see how we work across energy and the data center ecosystem, or contact the team.
The Bottom Line
- Natural gas for data centers is a commercial race, and it runs on information.
- The platforms on this panel already handle nominations, scheduling and notices, and some already ship AI agents.
- The open lane is outside public data: permits, queues and rules, checked and handed to a person.
- Let code do the math, let AI write, check every number, and let people decide.
About the Author
Jason Langella is the Founder and Chairman of SEO Agency USA. He was named the 2025 Associate Member of the Year by the Florida Municipal Electric Association (FMEA) and has taken part in more than 100 energy and utility industry events. He has spoken on technology and AI at every 2026 LDC Gas Forum, from Ponte Vedra Beach to Austin. His earlier posts cover what AI is changing in energy and what jobs AI will replace in natural gas and LNG. Connect with Jason on LinkedIn.
Frequently Asked Questions
Why are data centers using natural gas?
Data centers need large amounts of power around the clock, often faster than the grid can deliver it. Natural gas plants can be sited near the load, built in modular blocks, and run continuously. The IEA calls natural gas the largest source of additional supply for U.S. data centers through 2030.
What is behind-the-meter power for a data center?
Behind-the-meter power is generation built on or next to the data center campus that serves it directly, rather than through the utility grid. The IEA estimates that around 15 to 27 GW of on-site natural gas generation may power data centers by 2030, mostly in the United States.
How much power are data centers asking for in Texas?
ERCOT told the Texas Senate in July 2026 that it is tracking approximately 474 GW of large loads seeking to connect, about 90% of them data centers, as of June 2026. That is more than five times ERCOT's all-time demand record of 91,134 MW, set July 22, 2026.
How is AI used in the natural gas industry?
As with AI in oil and gas generally, the commercial platforms already use AI inside nomination, scheduling and forecasting work, for example to scan pipeline notices or recommend actions a person approves. The newer opportunity is outside public data, such as turbine air permits, interconnection queues and large-load rules, checked and handed to a person.
How does the Texas Data Center Gas Watch avoid wrong numbers?
Python reads the public files and does every calculation. Claude only writes the wording, under a rule to copy every number exactly. A second Python step checks that every number in the draft appears in the source facts and flags any that does not. A person reviews before anything is released.
Can AI connect to ETRM and SCADA systems?
Yes, in stages. Give AI its own read-only account for ETRM, ERP and CRM data, so it drafts and a person updates the system of record. For SCADA, AI analyzes a copy of historian data sent out of the control network and recommends. AI assistants get no path to send commands. Operators and engineers decide.
What are the biggest bottlenecks in delivering natural gas to data centers?
Four stand out: gas turbine supply, with multiyear lead times after a 70% surge in orders in 2025, per the IEA; pipeline capacity and new laterals; state permitting and large-load rules; and contracts for firm supply, transport and power that have to line up with a construction schedule that keeps moving.
When is the NatGas To Power Forum technology panel?
"From Molecules to Megawatts: Transforming the Natural Gas Commercial Value Chain Through Innovative Technology" runs Wednesday, November 11, 2026, 10:15 to 11:45 AM Central at the JW Marriott Austin. Cleve Hogarth moderates, with panelists from Quorum Software, NatGasHub.com, nGenue, Trellis Energy Software and SEO Agency USA.
Stop Focusing on Yesterday's Search
See how your brand appears in the new AI Search Funnel. Our free GEO AI Grader reveals if ChatGPT, Perplexity, and Google AI Overviews can find and recommend your brand. Top-of-funnel visibility starts here.
