What are the highest quality energy stocks in the US that will benefit from AI data center buildout?

Four core candidates earn their place in different ways

The shortlist combines direct exposure with diversified and regulated businesses; Vistra requires greater tolerance for merchant-power and execution risk.

Sources: company disclosures; assessments are our judgments. [4] [5] [6] [7] [8] [2] [9]

The distinction begins with the electricity business itself. An operating generator can improve its economics by selling existing output under a more attractive contract. A regulated utility typically earns through investment in approved infrastructure. A gas infrastructure company can collect transportation fees or, as Williams is doing, invest further along the chain in power generation. Each route has a different capital requirement and a different claim on the customer’s electricity bill.

Our preference is for evidence along that entire chain: a customer commitment, an asset capable of serving it, a funding plan and a plausible return. Constellation provides the clearest starting point.

Constellation owns the scarce asset customers already want

Constellation’s approximately 22 GW operating nuclear fleet gives it a substantial base of existing supply. Its 2025 fleet produced 183 TWh of electricity. That operating footprint matters because the company can negotiate over electricity already being generated while new projects work through construction and approvals. [4]

Meta’s Clinton agreement shows the mechanism. Beginning in June 2027, the 20-year contract supports 1,121 MW of nuclear output, continued operation and relicensing, alongside a 30 MW uprate. Meta obtains long-duration access to reliable, emissions-free electricity; Constellation obtains a customer commitment supporting investment in the plant. [5]

“Securing clean, reliable energy is necessary to continue advancing our AI ambitions,” Meta’s head of global energy, Urvi Parekh, said when the agreement was announced. The customer is buying certainty as well as electricity. Our view is that an owner of scarce, operating nuclear capacity is well placed to supply both. [5]

The incremental earnings mechanism is the improvement over the sale the plant would otherwise have made. On Clinton’s contracted capacity, assumed utilization of 90% and a hypothetical realized-price improvement of $10–$30/MWh produce $88 million–$265 million of annual gross revenue uplift, before additional operating, investment and financing costs. The arithmetic is capacity multiplied by annual operating hours and the pricing improvement. This sensitivity shows why even a modest premium on a large operating asset can matter. [5]

There is also contracted expansion potential. The Microsoft-backed, 835 MW Crane restart gives Constellation another route to supplying a large technology customer, subject to restart execution. Its additional 920 MW of long-term contracts announced with second-quarter results begin between 2029 and 2032 and include customers such as Walmart. That broader contracting activity strengthens the clean-power franchise beyond AI alone. [4] [1]

Our preference for CEG comes with a substantial qualification: its corporate risk has changed. Calpine closed in January 2026, bringing a broader generation portfolio and integration obligations. Investors must assess the combined business’s financing and cash generation alongside the nuclear opportunity. [1]

We prefer that diversified franchise to Talen for a quality-focused shortlist. Talen has an authentic Amazon agreement reaching 1,920 MW at full ramp by 2032 and running through 2042. The concentrated plant and customer exposure makes it a less suitable first answer to this question, even though the commercial opportunity is substantial. [10]

Vistra’s contract is compelling, but most of the output already exists

Vistra provides a second example of the value in operating nuclear assets. Its Meta package covers 2,609 MW, comprising 2,176 MW from operating Perry and Davis-Besse generation and 433 MW of planned uprates. Purchases begin in late 2026, with the full package coming online through 2034. [9]

Existing generation accounts for 83% of the agreement, making repricing and contract durability central to the economics.

Source: Vistra–Meta announcement. Shares are calculated from 2,176 MW operating output and 433 MW planned uprates. [9]

The agreement gives Vistra greater certainty to invest in equipment upgrades and continued plant operation. At an assumed 90% utilization, a hypothetical $10–$30/MWh pricing improvement on the operating portion would translate into $172 million–$515 million of annual gross uplift before incremental costs. The investment question is the net benefit after those costs and the required uprate spending. [9]

Vistra’s existing business is also growing. Company-wide second-quarter ongoing adjusted EBITDA increased from $1.35 billion to $1.77 billion, approximately 31%. That growth preceded the start of purchases under the Meta agreement. For 2026, management guided to $6.8 billion–$7.6 billion of ongoing adjusted EBITDA and $3.93 billion–$4.73 billion of ongoing free cash flow before growth investment. [11]

Our view is that VST belongs beside the core shortlist as a higher-risk addition. Its contracts and generation assets are attractive, but merchant exposure, acquisition obligations and growth spending warrant greater tolerance for uncertainty. We retain CEG as our preferred direct-generation franchise.

There is another way to meet a data center’s need for reliable electricity: bring the fuel and the generating plant to the customer. That is where Williams becomes interesting.

Williams has progressed from gas-demand exposure to operating power

Williams’ strongest evidence is an operating project. By its August earnings call, the company had brought the first 200 MW of Socrates into service. Management said the phase was delivered in under 18 months from commercialization, on time and within budget, with the next phase expected before year-end. [7]

That is a more tangible link to data-center demand than a national gas-consumption forecast. Williams also has a signed customer agreement for Neo, a 682 MW behind-the-meter project with an indicated $2.3 billion cost, and a separate Atlas gas infrastructure agreement providing up to 164 MMcf/day of capacity to a Northeast data center. Neo’s scheduled service date is in the second half of 2028. [12] [13]

The investment case combines a gas network with the ability to deliver fuel and power infrastructure around a customer commitment. But Williams must fund and execute the new assets, which makes the financing structure as important as the project list.

Blackstone and partners committed $5.34 billion for a 49% noncontrolling interest in five Power Innovation projects. Approximately $4.4 billion represents their share of expected construction capital, while approximately $0.9 billion is additional consideration to Williams. Williams obtains funding and monetizes part of the opportunity while retaining 51% of the economics. [6]

The construction contribution implies approximately $9.0 billion of total project capital: $4.4 billion divided by 49%. Williams’ share is approximately $4.6 billion, or roughly $3.7 billion after the additional consideration, before timing, taxes and other adjustments. That is a meaningful funding commitment even with outside investors participating. [6]

Our estimate is $0.2 billion–$0.8 billion of incremental annual Williams economic-share EBITDA around 2029–2031, with a central scenario near $0.46 billion. The range reflects assumed project returns and commissioning progress. It measures Williams’ economic participation before interest, taxes and depreciation; the consolidated reporting treatment may differ. [6] [7]

Williams’ future earnings contribution depends on project returns and commissioning, even after outside investors absorb part of the funding burden.

Note: analyst assumptions. Each case applies the return and rollout assumptions to $8.98 billion of implied capital and Williams’ 51% interest, then subtracts the modeled initial-phase baseline. The low case discounts commissioning; the high case assumes full rollout and the strongest project return. Source inputs: Williams financing announcement and operating update. [6] [7]

We prefer WMB to Kinder Morgan for direct exposure to this theme. KMI remains a credible conventional gas infrastructure alternative: its Amarillo expansion is fully subscribed under a long-term contract, supports demand including data-center development and targets third-quarter 2028 service. KMI’s share of the approximately $200 million project is about $75 million. [14]

Scale and attribution matter. KMI’s total project backlog was $9.6 billion at the second quarter, with more than 60% supporting power generation and local distribution demand. Those customers span a much broader market than AI. Applying the company’s 5.6-times first-full-year project EBITDA multiple for a selected backlog to its Amarillo investment share yields roughly $13 million of annual project EBITDA; the actual project return and data-center share remain open. [14]

The strongest objection to our Williams preference is the change in its risk profile. Onsite generation adds construction and operating obligations to a gas infrastructure franchise. The successful first phase and outside capital are favorable evidence; subsequent projects must earn that confidence again.

NextEra and Duke trade some direct exposure for resilience

NextEra is our diversified growth choice. Florida Power & Light supplies a regulated operating foundation, while Energy Resources provides a renewables and storage development platform. Second-quarter adjusted EPS rose from $1.05 to $1.15, a 9.5% increase. Energy Resources’ backlog reached approximately 35.1 GW after 1.1 GW of projects entered service since the preceding earnings call. [8]

That combination gives the investment case more than one source of growth. The backlog encompasses the broader renewables and storage business, so its conversion into profitable operating projects matters more than attaching an AI label to the whole quantity. FPL, meanwhile, expected $12 billion–$13 billion of capital investment during 2026, illustrating the financing demands behind utility growth. [8]

NextEra also has nuclear optionality. Its 615 MW Duane Arnold restart targets operation no later than the first quarter of 2029, subject to approvals, supported by up to $1.9 billion of federal loan financing. The opportunity adds another technology to the development platform while carrying its own execution requirements. [15]

We would not automatically put NEE ahead of CEG and WMB for this particular question. Its diversified franchise is appealing, but the other two have more direct, identifiable routes from data-center demand to contracted assets. NEE earns its core place through breadth and operating growth.

Duke is our defensive choice on the valuation evidence available. Its August presentation identified 7.8 GW of signed data-center electric-service agreements within a 15.4 GW high-confidence late-stage pipeline. The signed agreements represent approximately half that pipeline and provide a firmer starting point for assessing future load. [2]

The regulated mechanism is straightforward. A utility builds approved infrastructure, places it in service and earns the authorized return on the equity supporting that investment. In a simple sensitivity, $1 billion of approved operating rate base, financed with 40%–60% equity at a 9%–11% allowed return, produces $36 million–$66 million of annual equity earnings before timing effects. Those financing and return ranges are assumptions; the investor ultimately receives the result per share.

That last step is crucial for Duke. Management expects $10 billion of common equity issuance during 2027–2030, equivalent to $2.5 billion a year if evenly distributed. New shares help finance the assets, but spread their earnings across a larger ownership base. Our conviction depends on growth in per-share earnings, supported by actual load and appropriate cost recovery. [2]

Southern is a close regulated alternative. Georgia Power’s OpenAI agreement covers 3.2 GW of prospective demand and requires the customer to cover the full cost of dedicated infrastructure. The company projects approximately $950 million of annual portfolio-wide customer savings beginning in 2029. Those savings show how the economics can be shared with existing customers through regulation. [16]

AEP also deserves attention, but its large development pipeline needs a conversion discount. Its 69 GW outlook through 2030 combines electric-service agreements and letters of agreement, alongside a $78 billion capital plan. Its Ohio data-center tariff provides minimum-bill, collateral and contract-term protections. We would move it up the list as those commitments become energized load and attractive per-share earnings. [17] [18]

The 17× versus 22× gap changes where we would start

The clearest current valuation comparison favors Duke. Its September 28 close of $113.41 divided by the $6.675 midpoint of 2026 adjusted EPS guidance gives 17.0 times. Constellation’s $260.43 close divided by its $12.00 guided adjusted operating EPS midpoint gives 21.7 times. The earnings definitions differ, but the comparison makes the premium visible. [2] [1] [3]

Duke trades within our entry discipline; Constellation’s stronger direct exposure comes at a multiple above the range we prefer.

Note: points divide September 28 closing prices by guided 2026 adjusted EPS midpoints; bands show our discretionary entry multiples. Sources: company guidance and Fiscal.ai price records. [1] [2] [3]

Our entry discipline is 18–20 times Constellation’s current midpoint, equivalent to $216–$240 per share, and 16–18 times Duke’s, equivalent to $106.80–$120.15. These ranges express our valuation preferences: we want more margin of safety around CEG’s merchant and acquisition exposure. Changes in earnings guidance would change the corresponding prices. [1] [2]

The strongest case against our shortlist is that attractive assets may already command prices that absorb much of their future benefit. Higher construction costs, interest expense, equity issuance or less favorable regulatory allocation can then leave shareholders with disappointing returns despite strong electricity demand. That objection deserves substantial weight, particularly for CEG’s premium valuation and Williams’ modeled project returns.

Our firmest price judgment is therefore Duke versus Constellation. A low-looking share price or a large EBITDA figure would be insufficient grounds to call the other shortlisted stocks cheap.

Demand supports the opportunity; delivery determines the winner

The demand backdrop is large enough to matter. LBNL’s updated scenarios put total US data-center electricity consumption at 521–843 TWh in 2030, with a 649 TWh reference case, against a modeled 192 TWh in 2024. Subtracting the baseline gives 329–651 TWh of additional annual consumption, or approximately 38–74 GW of average delivered power; the reference case is about 52 GW. The estimates cover all data centers. [19]

Regional supply and delivery constraints decide where that demand can be served. PJM’s procurement for the 2028–29 delivery year left a 6.831 GW shortfall against its reliability requirement. That is a concrete reason to value dependable capacity in the relevant market, while recognizing that new customers still require workable transmission and service arrangements. [20]

Gas infrastructure also has room to benefit. Using the electricity scenarios above, assumed gas shares of incremental generation of 15%–40%, a 7,548 Btu/kWh heat rate and 1,037 Btu per cubic foot yields approximately 1.0–5.2 Bcf/day of additional gas consumption by 2030. The central case uses a 25% gas share and produces about 2.3 Bcf/day. The range supports a substantial infrastructure opportunity while leaving geography, contracts and investment returns to determine its distribution. [19] [21]

This is why we remain skeptical of using a generic gas-price thesis as the first answer to a quality-focused AI question. EQT’s ten-year agreement to supply 325,000 Dth/day to CPV Shay is a measurable connection to power demand, but its pricing is linked to PJM power prices. Investors still accept producer and commodity economics. [22]

ONEOK’s disclosed agreement to supply gas to a 1 GW generation system provides another link, with management indicating more than $100 million of required capital. The project evidence is less developed than the Williams and KMI cases. Nor do we see a sufficiently direct domestic data-center case for Cheniere to displace the shortlist. [23]

We do not need every proposed data center to arrive on schedule for the selected franchises to benefit. We do need their particular customers to pay, their particular assets to operate and the investment economics to survive the funding bill.