Will AI Data Centres Make Shale Gas the Winner? What the Data Says

Most investors file AI under chips, cloud software and clean power. The less glamorous story is that a large share of the electricity behind it may come from gas fracked out of shale in Texas and Pennsylvania. Goldman Sachs projects US data-centre power demand rising roughly 160% by 2030 from 2023 levels, and the fuel that can be switched on fastest is gas.

That demand forecast collides with a sober price signal. In early October 2026, Henry Hub spot gas, the main US benchmark, traded near $3.12-$3.18 per million British thermal units (MMBtu). That is well below the bullish bank scenarios.

The gap matters to anyone allocating capital to gas. If AI creates a durable demand floor, today’s price looks cheap. If it does not, the thesis is ahead of the evidence.

Here is where the AI-gas case stands on solid data, where it rests on investor inference, and which signposts to watch before you position.

Why the AI power surge lands on gas first

Start with the size of the number. Goldman’s 2024 work has US data centres rising from about 3-4% of national power demand to roughly 8% by 2030, with later work citing 8-11%. Data centres alone add about 1 percentage point to US power-demand growth, lifting total demand growth to roughly 3-3.5% a year, measured as a compound annual growth rate (CAGR). After a decade of near-flat consumption, that is a sharp break.

The scale of AI data centre energy demand explains why banks now treat load growth as structural, since a decade of flat consumption has given way to sustained annual increases that utilities must plan around.

Goldman Sachs projection US data-centre power demand is forecast to grow about 160% by 2030 versus 2023. Goldman’s separate global estimate, in later GS SUSTAIN work, is roughly 170-175%.

The distinction matters. The 160% figure is US-specific; the 170-175% figure is global.

Morgan Stanley puts the AI share of incremental US load at about 75%, while Goldman separately cites about two-thirds. The near-term path keeps steepening: Goldman’s October 2026 updates lifted US data-centre capacity forecasts to 64 GW by end-2026 and 90 GW by end-2027, above earlier outlooks.

Then comes the timing mismatch. AI campuses need power within months or a few years. Renewables paired with storage, new transmission lines and nuclear plants take longer to permit and connect.

Goldman therefore allocates about 60% of new data-centre-serving capacity to gas, split roughly evenly between combined-cycle plants and peakers (smaller plants that run during demand spikes). Morgan Stanley’s estimate is lower, at about 30% of AI-driven demand met by gas.

What you are looking at is two separate claims. The demand shift is what banks call structural; the gas share depends on what can be built quickly. That speed advantage is what turns an AI story into a gas story.

How does fracking connect to the grid? The shale basin transmission chain

Speed to power explains the fuel choice. The next question is how a data-centre load request actually reaches a wellhead.

The chain runs like this. Higher load on the Texas grid, run by the Electric Reliability Council of Texas (ERCOT), and on PJM, the operator covering much of the US Mid-Atlantic, prompts new gas-fired capacity. Those plants draw volumes from the Permian Basin (for ERCOT) and the Marcellus and Utica shales (for PJM), tightening basis differentials, the price gap between a regional hub and Henry Hub, and potentially lifting Henry Hub itself if pipeline capacity lags.

One popular claim, that the two grid operators approved more gas in 2024-2026 than in the prior decade combined, could not be confirmed in available research. The queue data tells a more mixed story.

ERCOT and the Permian

ERCOT moves faster. Gas interconnection requests reached about 80 GW in 2026 updates, 17.6% of the active queue, up 23% since mid-2026, although solar and batteries still dominate overall. Average waits run 20+ months.

There is a catch. Permian gas is largely associated gas, produced as a by-product of oil drilling, so supply arrives regardless of gas prices. AI demand here may show up as tighter regional basis rather than a Henry Hub spike.

PJM and the Marcellus

PJM’s new cycle looks enormous on paper: 617 projects totalling 168.6 GW in Phase One, including 81.1 GW of gas, after reforms cut speculative entries. Yet some PJM analyses found no new gas plant above 100 MW submitted since 2018 had reached commercial operation by early 2026.

The PJM capacity shortfall shows why queue size alone misleads: even with 81.1 GW of gas proposed, new supply is struggling to arrive before data-centre load does.

AI is also one pull among several. The US Energy Information Administration (EIA) forecasts dry gas production of 112.20 Bcf/d (billion cubic feet per day) in 2026 and 116.13 Bcf/d in 2027, alongside liquefied natural gas (LNG) exports of 17.6 Bcf/d and 18.6 Bcf/d.

Queue size is not capacity built. Read the grid data as a pipeline of intent; its conversion rate, not its headline gigawatts, decides how much shale volume AI adds to your thesis.

What price floor does AI demand actually justify?

If conversion is uncertain, price forecasts should be too. They are, and the spread is wide.

Same driver, different answers Deloitte, in secondary analysis, sees Henry Hub near $5.40 in 2030 on LNG and data-centre demand. The EIA projects $3.80 for the same year.

Now lay that beside the market. The EIA’s Short-Term Energy Outlook (STEO) has 2027 at $3.16, and early-October spot sits at $3.12-$3.18. Both are below a $4 floor.

The two Morgan Stanley figures were never reconciled, and neither should be read as the bank’s definitive view. Morgan Stanley itself notes Henry Hub averaged $3.23 over the past decade and warns rising production can dampen price effects for consumers.

For a $4 floor to hold, conversion would need to accelerate, LNG demand would need to stay firm and production growth would need to slow. None of those is assured. The evidence supports AI as demand support rather than a proven floor, so size your exposure across a range of outcomes, not a single bank’s scenario.

Past performance does not guarantee future results. Financial projections are subject to market conditions and various risk factors.

Who benefits, and what could break the thesis?

That price range feeds directly into the equities investors usually name.

The equity thesis

Targa Resources (Permian midstream), EQT (Marcellus) and Coterra are commonly framed as volume beneficiaries. That framing is an investor thesis. Explicit AI or data-centre commentary was not prominent in their reviewed guidance, and no named hyperscaler gas-supply deals tied to them were found.

Supporting evidence sits elsewhere in the chain. Oilfield services firms report growing orders for data-centre power solutions such as turbines and modular systems, and utility Entergy references large-load data-centre agreements.

Risks and signposts

• Efficiency gains: better chips and cooling could cut power per unit of compute, lowering volumes producers sell.
• Forecast overstatement: weak AI monetisation or hyperscaler capex cuts could slow the build-out.
• Turbine and labour bottlenecks: delayed gas plants push demand growth further out.
• Permian takeaway and associated gas: oversupply could depress regional prices for Permian-exposed names.
• Power-price politics: siting limits or capacity-market reform could favour renewables over gas.
• Emissions policy: stricter standards or carbon pricing could shorten gas’s bridge role.

The net effect is a wide cone of uncertainty. AI may deliver moderate price and volume support rather than a multi-year supercycle.

Capacity-market reform is already live in PJM, where an auction clearing at its price cap still left a reserve shortfall, and any redesign could tilt the economics between gas and other resources.

Track these five signposts:

1. Gas queue-to-commissioning conversion in ERCOT and PJM.
2. Turbine order backlogs.
3. Hyperscaler capex guidance.
4. Permian takeaway additions.
5. Producer commentary on data-centre contracts.

Treat these three equities as leveraged to the thesis only to the extent demand converts into contracted volumes. The signposts matter more than the narrative.

Sizing the AI-gas thesis without overpaying for it

The evidence splits cleanly. Data-centre power demand growth is well supported by bank research. Gas as the fast-build fuel is plausible. The price floor and company-level volume uplift remain inferences.

For a global energy portfolio, treat AI demand as one support alongside LNG exports rather than the whole case. Weigh basin-specific takeaway, since Permian and Marcellus exposure carry different risks, and size positions against the full span from the EIA base case to the bullish bank calls.

Two variables are most likely to settle the debate: how many queued gas plants actually reach operation, and whether contracted data-centre gas volumes begin appearing in producer disclosures.

For readers weighing utilities against producers, our deep-dive into AI energy infrastructure investing explains how the regulated rate-base model converts AI load into multi-decade earnings.

This article is for informational purposes only and should not be considered financial advice. Investors should conduct their own research and consult with financial professionals before making investment decisions.