ONEOK Turns AI Power Demand Into Gas Capex • The Weighted Average

AI infrastructure is turning gas pipelines into part of the compute stack. ONEOK says it secured a supply agreement for 1 gigawatt of power-plant demand and will spend more than $100 million on the associated project; divided by the disclosed demand, that is a minimum $100 million per gigawatt of gas-linked power infrastructure. The contract is a useful warning for data-center buyers: power procurement is not an afterthought to the GPU order.

The details come from ONEOK’s Q2 earnings call, where Chief Commercial Officer Sheridan Swords described the agreement as supplying natural gas into a power-generation system serving the AI-data-center opportunity. The company’s earnings release reported $2.12 billion of second-quarter adjusted EBITDA and kept 2026 capital-expenditure guidance at $2.7 billion to $3.2 billion. The AI-linked project is not the whole budget, but it shows how a software demand forecast becomes a physical contract.

The two sources yield a more useful budget ratio. ONEOK’s call puts the project above $100 million, while the release gives a 2026 capex range whose midpoint is $2.95 billion. 100 ÷ 2,950 = 3.39%, so the project consumes more than 3.4% of the midpoint budget before any overrun or additional scope. That is a lower bound, not a disclosed return.

The company’s earnings summary adds the operating frame: record NGL throughput, 7% year-over-year NGL growth, and 15% growth in the Gulf Coast and Permian region. Those figures do not prove completion, but they show why ONEOK can treat the contract as an extension of an existing network.

This brief extends the archive’s AI data-center power analysis and the NVIDIA photonic-network story: the bottleneck is becoming a chain of dependencies, not a single accelerator specification. The Business Arena lead applies the same rule to software: capacity is useful only when the system can carry it safely.

The power bill arrives before the model bill

ONEOK’s management called the project “not a high capital,” but the number is still material: more than $100 million tied to one gigawatt of power-generation demand. The company says it has firm demand and expects a good return, but it does not disclose the project’s full schedule, customer, fuel volume, rate structure, or absolute return. The correct reading is not “AI power is cheap.” It is that a dedicated supply relationship can make an infrastructure decision legible to an energy company.

The 3.4% budget ratio is a floor, not a quote. It does not establish a final cost per megawatt, a levelized cost of electricity, or a data-center tariff. The arithmetic is useful because it puts a lower bound beside the demand claim. A buyer can now ask which costs sit upstream of the meter and which ones the campus must finance itself.

ONEOK also says commercialization of some data-center projects has taken longer than expected and that several additional opportunities remain in discussion. That caveat is as important as the contract. A large announced load is not the same as a final investment decision, an energized substation, or a customer taking power at the promised ramp. The company’s own call distinguishes the secured supply agreement from other multi-large-scale data-center developments that had not reached FID.

The company’s wider numbers show why the project fits its strategy. ONEOK reported 7% year-over-year NGL throughput growth and raised its 2026 net-income guidance midpoint to $3.6 billion. Management describes future AI and power-generation projects as “singles and doubles” in the $100 million to $400 million or $500 million range. AI demand is being absorbed into an existing energy network.

That integration creates an operator implication for both sides. Data-center developers should involve gas suppliers, utilities, and transmission planners before the site is treated as “under development.” Energy companies should demand evidence that the load can achieve FID and construction milestones before reserving capital. A power contract can reduce one risk while exposing another: counterparty concentration if the compute campus slips.

Switch the planning unit from GPUs to deliverable megawatts

The right first switcher is a campus developer with a firm load profile, a plausible energization date, and enough balance-sheet capacity to coordinate fuel, generation, interconnection, and cooling. The cost is project management and contractual complexity before a single production cluster goes live. Teams that cannot yet specify when and how the site will consume power should not turn a headline gigawatt into a procurement commitment.

The strongest counterpoint is that gas is not the only route. A project may use grid supply, renewables, storage, nuclear, or a hybrid portfolio; the ONEOK contract does not demonstrate that natural gas will dominate AI power. It also does not establish that this project serves one named data center, only that ONEOK described the supply as supporting power-generation demand in the AI-data-center context. Operators should not generalize one commercial agreement into an industry-wide fuel mix.

There is a timing risk too. ONEOK’s earnings materials show strong existing capital activity, including $2.7 billion to $3.2 billion of 2026 capex guidance, while the call says AI commercialization is taking longer than expected in some cases. If permits, transmission, equipment, or customer financing slip, a gas lateral or power plant can be ready before the compute load. The conclusion breaks if the contract is not followed by FID, construction, fuel delivery, and a disclosed offtake schedule.

Evidence that would change the verdict is concrete: an executed FID, an identified customer or project phase, a construction and energization timetable, disclosed fuel volumes, and performance against the supply agreement. For a buyer, the equivalent evidence is a signed utility interconnection, a guaranteed delivery date, a take-or-pay exposure that matches the ramp, and a fallback plan if the cluster arrives late.

The OpenAI Stargate expansion makes the same point at larger scale. OpenAI and partners described nearly 7 gigawatts of planned capacity and more than $400 billion of investment over three years, with Lordstown, Ohio, among the sites. Those are plans, not proof every site is energized. ONEOK’s contract reveals where the promise becomes a fuel and financing decision.

Data-center developers should switch their site checklist from GPU availability to deliverable megawatts, including fuel, interconnection, permits, cooling, and the customer ramp.
Energy suppliers should price the project against firm demand, FID timing, and counterparty credit; the disclosed floor is more than $100 million per gigawatt, not a guaranteed return.
Procurement teams should require a construction and energization schedule before treating a power agreement as usable capacity.
Operators should watch whether ONEOK’s additional AI-data-center discussions reach FID; that evidence would separate a durable demand channel from a promising pipeline.

The practical lesson is not that gas wins the AI power race. It is that compute commitments now pull capital into the infrastructure beneath the rack. A campus is not online when the GPUs arrive. It is online when the power system, fuel contract, and counterparty can keep the electrons coming.