Even with that correction, the outlook is consequential. A BloombergNEF projection points to a much larger US data-center buildout than it anticipated previously, while grid-connection constraints could push operators toward substantial on-site gas generation. For Windows users and IT decision-makers, this is not merely a debate about distant power plants: electricity availability, utility costs, cloud-region capacity, construction timelines, and the sustainability claims surrounding AI services can all be shaped by how this buildout is supplied.
What the 15 Bcf/d forecast actually says
Reported coverage of BloombergNEF's outlook says that gas consumption used to generate electricity for US data centers could grow by 15 Bcf/d in the ten years through 2035. That is an incremental-demand projection, not a disclosed estimate of the sector's total gas consumption at the end of the period.
The difference matters because total 2035 demand would require a starting baseline, assumptions about existing data-center load, the share served by gas, generator utilization, and the evolution of grid supply. Those underlying methodological details and an uncertainty range were not available in the retained material.
It also means a claim that US data centers will become the world's fifth-largest natural-gas consumer is not established by this forecast alone. Such a ranking would need a like-for-like comparison: total 2035 data-center consumption against total consumption in other countries in the same year, using comparable definitions. Comparing a projected increase in one category with a country's current consumption can illustrate the scale of the increment, but it cannot prove a future global ranking.
The forecast is nevertheless a sign that assumptions about AI infrastructure have moved sharply. BloombergNEF expects 194 GW of US data-center capacity to be online by 2035, an 83% upward revision from its December 2025 estimate. Separate reporting characterizes that outcome as around 20% of total US electricity consumption, versus 5.9% at present.
Those two figures should not be casually merged. Gigawatts describe power or capacity, while electricity consumption is energy used over time. A facility's actual annual electricity use depends on how continuously it runs and how much of its installed capacity it uses. But in practical terms, both measures point in the same direction: the forecast envisions data centers becoming a system-level load rather than a niche addition to the grid.
A large forecast is not a settled outcome
There is substantial uncertainty around data-center electricity demand and, by extension, gas use. PwC's conservative-to-bull scenarios put incremental AI-linked US gas demand in 2035 at 7.6 Bcf/d to 11.5 Bcf/d. That is materially below the reported 15 Bcf/d BloombergNEF increment.
Rhodium Group likewise describes the answer as highly uncertain, with estimates varying substantially according to methodology and assumptions. In one of its higher-demand cases, data centers reach 18% of US generation in 2035. That does not disprove the BloombergNEF case; it demonstrates that plausible outcomes depend heavily on variables that remain unsettled.
Among them are how quickly AI demand grows, whether inference workloads become more efficient, how rapidly newer chips improve performance per watt, where campuses are built, how much demand can be shifted in time, and how fast transmission, generation, and fuel infrastructure arrive. A projection should therefore be treated as a planning scenario, not as a measured future fact.
There is an important counterargument to the idea that more AI automatically means proportionately more fuel consumption. Improvements in hardware and software can reduce energy required for a given workload. Yet efficiency can coexist with rapidly rising total demand if the number, size, and utilization of AI workloads grows faster than efficiency improves. The evidence in this outlook does not resolve that tension; it shows why planning cannot rely on efficiency gains alone.
The binding problem may be grid delivery, not generation alone
The data-center question is not simply whether enough electricity can be produced nationally. It is whether electricity can be delivered to a particular site when the operator needs it.
BloombergNEF says that, if the grid cannot connect data centers faster than an estimated historical maximum of 10 GW in a year, 48 GW of on-site gas generation would be needed by 2035 in its base case. This is a striking conditional result. It does not mean 48 GW of on-site gas will definitely be constructed. It means that slow interconnection could turn private generation from a resilience option into a major part of the buildout plan.
The market is already responding. The same outlook identifies 124 GW of announced on-site gas capacity, though commissioning dates have been mapped for only 56% of that total. Announcements, then, should not be confused with operating equipment. The gap is a practical warning for investors, local communities, and enterprise customers assessing whether a planned AI campus can meet a promised service date.
Gas-turbine demand provides another indicator of pressure. S&P Global reported 51 GW of gas-turbine orders in 2025, the highest annual level since 2000, followed by 18 GW in the first quarter of 2026. This does not establish that every order is for a data center, nor does it establish a specific turbine shortage for any one customer. It does show that rising data-center loads are occurring during a period of exceptional demand for this equipment.
For public policy, the implication is that interconnection reform and transmission construction affect more than queue-management paperwork. Faster, predictable grid connections could reduce the case for long-lived on-site fossil generation. Conversely, delayed wires and substations can lock in choices made to keep construction schedules moving.
Gas supply and price effects are likely to be regional
The reported BloombergNEF outlook forecasts US producers increasing output by 35 Bcf/d between 2025 and 2035, but says another 11 Bcf/d would be required to meet projected demand. It also projects power-sector gas use reaching 54 Bcf/d in 2035.
Those figures imply a difficult infrastructure challenge, not a guaranteed national shortfall. Production must be accompanied by gathering systems, pipelines, storage, power plants or generators, and the electric transmission equipment needed to use the output where demand emerges. A national production figure does not ensure that gas or power will be readily available in a constrained local market.
That is why a blanket prediction that consumer gas prices will “shoot up” goes beyond the available evidence. S&P Global sees likely upward pressure on natural-gas and electricity prices in places where load growth exceeds supply additions. Rhodium finds that greater data-center demand can raise grid operating costs. Neither finding establishes that every household in the country will face the same result, or that a dramatic nationwide retail-price rise is inevitable.
The more defensible conclusion is regional and conditional. Communities hosting rapid data-center growth may face greater exposure to higher wholesale fuel and power costs, expensive network upgrades, and disputes over who pays. The distribution of those costs will depend on state utility rules, long-term contracts, new generation, and whether large customers pay adequately for the infrastructure built on their behalf.
The on-site generation trade-off is visible in Memphis
The Colossus 2 example illustrates the collision of speed, local air quality, and permanent infrastructure. An FTC early-termination notice confirms a transaction involving Elon Musk as the acquiring party and CF APR Super Holdings as the acquired party, with New APR Energy among the acquired entities. The available record does not establish a reported purchase price, so claims of a roughly $1 billion deal should not be presented as confirmed.
SpaceXAI says it has an agreed order with Mississippi regulators to remove 69 temporary mobile turbines in Southaven by July 2027 while bringing a 1.2 GW permanent plant online. It says that permanent facility will use 41 turbines and was authorized by a Clean Air Act permit granted in March 2026.
That account describes a transition from temporary equipment to permitted permanent generation, not an end to local generation. It also does not settle the environmental dispute. Reporting on claims from the Southern Environmental Law Center and the NAACP said that an alleged increase from 27 to 57 turbines corresponded to calculated increases of 111% for nitrogen oxides, 83% for fine particulate matter, and 88% for formaldehyde emissions. Those are advocacy-side calculations based on the turbine-count change, not independently verified stack-monitoring results.
This distinction is essential. Communities deserve transparent information about actual emissions, monitoring, permit conditions, and health implications. At the same time, percentage estimates in litigation or advocacy should not be recast as measured pollution outcomes before independent verification supports that conclusion.
Nuclear commitments help explain the long-term search for firm power
Major technology companies are also looking beyond gas. Amazon says it made a capital investment in X-energy through its Climate Pledge Fund and is partnering on an initial four-reactor, 320 MW small-modular-reactor phase that could expand to 960 MW. Google says it signed an agreement to purchase electricity from multiple Kairos Power small modular reactors, with the first unit intended to be online by 2030.
These are meaningful commitments to prospective firm, lower-carbon power, but they are not interchangeable. Amazon's documented arrangement includes a capital investment and a specified phased project. Google's documented arrangement is an agreement to purchase future reactor output. Neither fact establishes that every large AI or cloud company has invested in advanced-reactor startups, nor does either make nuclear capacity an immediate substitute for the gas generation contemplated in the nearer-term data-center outlook.
What this means for Windows users and enterprise IT
An individual Windows user is unlikely to see a direct charge tied to a single AI query. The consequences are more likely to appear indirectly: in the price and availability of cloud-backed AI services, the geography of new digital infrastructure, the credibility of vendor sustainability commitments, and potentially local electricity costs where large campuses cluster.
For enterprise IT teams, power has become a procurement and resilience issue. When evaluating AI platforms, hosted infrastructure, or data-center partners, it is reasonable to ask where capacity will be located, whether power supply is grid-connected or on-site, how long interconnection is expected to take, and what generation mix supports reliability claims. These questions matter especially for workloads that cannot tolerate curtailment or regional capacity delays.
For regulators, the central challenge is not choosing between AI and reliable electricity. It is making the expansion more legible and accountable: accelerate viable grid upgrades, require transparent cost allocation, distinguish announced capacity from operational capacity, enforce air-quality rules, and avoid treating a forecast as a certainty. The 15 Bcf/d outlook is a warning about the possible scale of the buildout—not proof that its most fuel-intensive path is unavoidable.