BloombergNEF Data Center Gas Forecast Puts AI Growth on a Fossil-Fuel Track

BloombergNEF has raised its US data center gas forecast to about 18 billion cubic feet per day by 2035. That projection is nearly double its estimate from nine months earlier. It turns the AI infrastructure race into a much larger test of pipelines, power plants, utility planning, and climate commitments.

The forecast does not mean every proposed AI campus will be built. It reflects a base case that already discounts projects likely to stall. Even after that adjustment, data centers become one of the largest new sources of US natural gas demand.

The central conflict is no longer simply AI growth against grid capacity. It is speed against long-term energy strategy. Meta, Microsoft, Google, Amazon, and other operators need dependable electricity before slower grid and generation projects can arrive.

Natural gas offers that dependable output, either through the grid or through power plants built beside data centers. However, gas also introduces fuel-price exposure, local air pollution, carbon emissions, and infrastructure commitments lasting decades.

The BloombergNEF Data Center Gas Forecast Has Doubled

BloombergNEF’s revision suggests that AI infrastructure plans are moving faster than previous energy models could absorb.

US data centers could consume about 18 billion cubic feet of natural gas daily by 2035, according to the new outlook. That would represent roughly 15 Bcf/d of growth over ten years.

The projected increase alone would exceed current gas consumption in almost every country. Only China, Russia, Iran, and the United States consume more, according to the comparison reported with the forecast.

The scale also exceeds the combined consumption of Germany and Japan. That comparison is imperfect because countries use gas across several economic sectors. However, it illustrates how a specialized digital industry could become a major energy consumer.

BloombergNEF’s estimate is not based only on public project announcements. The model reportedly accounts for the likelihood that many proposed data centers never reach operation.

That qualification matters because announced capacity greatly exceeds what existing grids, equipment suppliers, and construction markets can deliver. Projects can fail because they lack transmission access, permits, turbines, water, financing, or firm customers.

Even after those projects are filtered, BloombergNEF expects data centers to become the second-largest source of US gas demand growth. Only liquefied natural gas exports would add more demand over the period.

The gas projection follows another sharp revision to BloombergNEF’s electricity outlook. Its capacity forecast expects 194 gigawatts of US data centers online by 2035.

That estimate was 83 percent higher than the firm’s previous projection. It also implied that data centers could consume around one-fifth of US electricity by 2035.

BloombergNEF calculated that the grid would still face a 19 GW supply gap, even under an unusually fast connection schedule. Filling that gap would require as much as 48 GW of onsite gas generation in its base case.

Onsite generation sits behind the customer’s meter, rather than depending entirely on the regional grid. It can help a campus begin operating before transmission upgrades are finished.

The new gas forecast therefore represents more than a routine adjustment to fuel demand. It connects three escalating estimates: data center capacity, electricity consumption, and dedicated generation.

Each estimate depends on assumptions that remain uncertain. AI hardware efficiency could improve, announced campuses could be canceled, and alternative generation could arrive faster.

Yet the direction is becoming harder to dismiss. Every upward revision tells utilities that waiting for perfect information carries its own risks. It also tells technology companies that compute procurement now depends on energy procurement.

That changes the business of building AI. Companies cannot secure advanced chips, construct a campus, and assume sufficient electricity will follow. Power must be contracted, generated, stored, and delivered as part of the original design.

The result is a new bottleneck for AI deployment. The industry has spent years competing for GPUs. It is now competing for grid connections, gas turbines, pipelines, and permission to operate them.

AI Power Demand Is Moving Faster Than the Grid

The pressure comes from the mismatch between data center construction schedules and the slower timetable for expanding regional power systems.

A large AI campus can move from planning to construction faster than utilities can approve and complete major transmission projects. New power lines often cross several jurisdictions and require lengthy reviews.

Power plants also face equipment constraints. Gas turbine manufacturers must manage long order books, while transformers and switchgear remain difficult to source quickly.

This timing problem gives natural gas a practical advantage. Gas-fired generators can provide dispatchable power, meaning operators can raise or lower production when electricity demand changes.

AI facilities value that control because they run expensive computing hardware around the clock. An unreliable power supply leaves chips idle while financing, staffing, and cooling costs continue.

The appeal becomes stronger when a company can locate generation beside the data center. A dedicated plant can reduce dependence on an overloaded connection and avoid some transmission delays.

However, onsite gas does not remove every constraint. The plant still needs turbines, fuel delivery, environmental approvals, and interconnection equipment. It may also need backup systems for maintenance or supply interruptions.

Pipelines present another limitation. A region can possess large gas resources without having enough local delivery capacity for several new power plants.

That is why existing rights-of-way, storage sites, and pipeline connections are gaining strategic value. PwC’s gas infrastructure analysis describes permits, turbines, pipelines, water, and time as the scarce resources.

PwC estimates that incremental AI-linked US gas demand could reach 7.6 to 11.5 Bcf/d by 2035. Its range is lower than BloombergNEF’s latest total, partly because models define AI-related demand differently.

The difference should discourage false precision. There is no single meter recording national AI gas consumption. Analysts must combine proposed capacity, completion rates, utilization, efficiency, and expected power sources.

The models still agree on the mechanism. More data center capacity creates a large block of continuous electricity demand. Grid delays then increase the attraction of dedicated generation.

Lawrence Berkeley National Laboratory offers another view of the same pressure. Its 2025 update estimates data centers could account for 11.8 percent of US electricity by 2030.

The national energy report gives a scenario range from 9.5 to 15.3 percent. That range reflects uncertainty about equipment shipments, facility types, cooling, and operating practices.

Those percentages matter because data centers concentrate demand geographically. A national grid might appear capable of serving additional load while a specific utility territory faces a severe shortage.

Northern Virginia illustrates this concentration. Its established fiber networks and cloud infrastructure attracted clusters of facilities, placing exceptional pressure on local transmission planning.

New projects are spreading into Texas, Ohio, Pennsylvania, Louisiana, and other markets with available land or energy resources. The movement does not eliminate infrastructure pressure. It redistributes it.

Utilities must decide how much generation and transmission to build before every proposed campus becomes certain. If they build too little, projects wait or install private generation.

If they build too much, ordinary customers could inherit unnecessary infrastructure costs. Regulators must determine which expenses belong to data center developers and which belong in general electricity rates.

That decision gives the AI energy debate a direct consumer dimension. A delayed model release affects technology companies. A poorly allocated utility investment can affect every household and business in a service territory.

Natural Gas Solves the Timing Problem but Creates a Lock-In Problem

Gas can bring firm power to an AI campus sooner, but the resulting infrastructure can outlive the demand assumptions that justified it.

A gas plant built for a data center is not a temporary software deployment. Its pipelines, turbines, permits, and financing structures can remain in place for decades.

That long life creates a mismatch with AI forecasting. Compute demand changes quickly, while physical energy assets recover their investment over much longer periods.

A developer might expect a campus to operate near full capacity for years. Yet newer chips could deliver more computation per unit of electricity, or workloads could move to a different region.

Demand could also grow faster than expected. In that case, dedicated gas plants might expand from bridge infrastructure into a permanent foundation for the AI economy.

The word “bridge” therefore carries more uncertainty than it appears to carry. A bridge should lead toward another energy system. It becomes lock-in when the replacement never arrives.

The International Energy Agency expects renewables to provide nearly half of additional global data center electricity through 2030. Natural gas and coal together would supply more than 40 percent.

In the IEA base case, electricity generation serving data centers rises from 460 terawatt-hours in 2024 to more than 1,000 TWh in 2030. It reaches 1,300 TWh by 2035.

The IEA’s energy supply outlook also models a faster AI expansion. Under that case, data center electricity approaches 2,000 TWh by 2035.

Grid connection queues limit how quickly clean generation can serve that higher demand. Consequently, fossil fuels provide a larger share of the unexpected increase.

This mechanism explains why corporate renewable contracts do not settle the physical question. A company can purchase clean-energy certificates while its facility draws from a regional system that still burns gas.

Location and timing determine the actual generators responding to a new load. A solar project in another market does not necessarily supply an AI cluster operating overnight.

Batteries can move electricity between hours and reduce peak demand. They cannot independently produce the continuous energy required by a high-utilization campus.

Nuclear power offers firm, low-carbon generation, and several technology companies are pursuing it. However, new reactors usually face longer development schedules than data centers.

Restarting existing nuclear units or buying output from operating plants can help individual projects. Those options remain limited by geography and the available fleet.

Geothermal energy also attracts interest because it can deliver steady output with low operational emissions. Its near-term contribution depends on drilling performance, project execution, and suitable locations.

Natural gas wins many current comparisons because it is available, familiar, and dispatchable. It can support both grid generation and isolated campus systems.

Yet that advantage has a price beyond emissions. An AI operator using dedicated gas becomes exposed to fuel delivery, commodity markets, and environmental requirements.

The arrangement can also shift corporate expertise. Technology companies must negotiate with pipeline operators, turbine manufacturers, utilities, and air-quality regulators.

For cloud customers, these decisions may eventually affect computing costs and service locations. Energy-intensive workloads could migrate toward regions offering faster and more predictable power delivery.

Developers and enterprise buyers should therefore treat geography as part of AI architecture. Model performance matters, but the energy system behind a computing region shapes capacity, reliability, and future cost.

The Forecast Does Not Guarantee an 18 Bcf/d Outcome

The strongest challenge to the BloombergNEF data center gas forecast is that both AI demand and project completion remain unusually difficult to predict.

The headline number compresses many variables into one figure. Small changes in those variables can produce a different result by 2035.

One variable is the number of announced facilities that reach operation. The US development pipeline includes speculative projects seeking land, connections, or customers before financing is complete.

Some proposals compete for the same limited grid capacity. Counting all of them as independent future demand would exaggerate the likely outcome.

BloombergNEF says its estimate discounts projects that will not be completed. The public reporting does not expose every project-level assumption in the underlying proprietary model.

That limits independent evaluation. Readers can assess the direction and reported totals, but they cannot reproduce the entire forecast from public information.

A second variable is utilization. A data center with a stated maximum capacity does not necessarily consume that amount every hour.

Training clusters can create intense demand, while inference workloads may vary with user activity. Operators can also schedule some computing tasks around electricity availability.

A third variable is hardware efficiency. Each new generation of accelerators can perform more work per unit of energy, although total consumption can still rise.

Efficiency lowers the energy required for a fixed workload. It does not guarantee lower demand when cheaper computation causes customers to use much more of it.

This rebound effect sits near the center of AI energy uncertainty. Better chips could restrain electricity growth, or they could make additional models and automated services economical.

A fourth variable is the power mix. BloombergNEF’s forecast assumes gas serves a substantial portion of the new load, especially where grid connections remain constrained.

Faster transmission construction would weaken that assumption. So would quicker deployment of renewables, storage, nuclear restarts, enhanced geothermal systems, or flexible computing.

The US gas market introduces another uncertainty. Total consumption already averaged a record 92.0 Bcf/d in 2025, according to the Energy Information Administration.

Electric power accounted for 35.8 Bcf/d that year. The sector remained the country’s largest gas consumer despite a decline from 2024.

The EIA’s consumption data shows how weather, solar additions, batteries, and power demand can move annual gas use in different directions.

Adding as much as 15 Bcf/d for data center electricity would interact with LNG exports, industrial demand, household heating, production growth, and pipeline constraints.

Prices could rise in regions where supply or delivery fails to keep pace. Higher prices would then change the economics of gas generation and alternative power sources.

Environmental policy could also alter the outcome. Gas plants emit carbon dioxide, while production and transportation can release methane, a potent greenhouse gas.

Local opposition may focus on nitrogen oxides, water consumption, noise, and the distribution of electricity costs. Permitting resistance could delay both generation and the data centers themselves.

The projected 18 Bcf/d should therefore be read as a scenario, not a scheduled result. It describes the system required by one plausible path for AI expansion.

That distinction does not make the forecast unimportant. Planning decisions happen before final demand becomes certain. Utilities and developers must act while the range remains wide.

The risk runs in both directions. Underbuilding can delay economically valuable infrastructure. Overbuilding can leave customers paying for assets designed around demand that never appeared.

Who Bears the Cost of the AI Gas Buildout

The decisive policy question is not simply whether new gas capacity gets built, but who pays for it and who absorbs its risks.

Data center developers can finance dedicated generation themselves. That approach places more project risk on the company benefiting from the electricity.

The boundaries become less clear when a regulated utility builds generation, pipelines, substations, or transmission for expected data center growth. Utilities recover approved expenses through customer rates.

Regulators can require large customers to sign long-term contracts or provide financial guarantees. Those protections reduce the chance that other customers pay for abandoned facilities.

However, contract design becomes difficult when demand arrives in stages. A campus might begin with a fraction of its planned capacity and request repeated expansions.

Utilities also value large customers because they can spread fixed system costs across more electricity sales. That benefit disappears if the new infrastructure costs exceed the revenue it produces.

Rate design must therefore address several questions. How much capacity is genuinely committed? How long must the customer remain? Who pays if construction stops?

The same questions apply to gas infrastructure. A pipeline extension serving several power plants could remain useful after one data center closes.

Conversely, a dedicated lateral line can become stranded if a single campus changes direction. Its value depends on alternative customers and regional network design.

Communities also bear nonfinancial costs. New generation can increase local air pollution, water demand, traffic, and industrial land use.

The benefits may include construction work, tax revenue, and investment. The balance varies by location and by the terms negotiated with the developer.

Climate commitments create another allocation problem. Large technology companies have announced ambitious emissions goals, but their AI expansion increases electricity demand.

Purchasing renewable energy can support new clean projects. It does not erase the emissions from dedicated gas plants supplying facilities before those projects become available.

Carbon capture may reduce emissions at some gas plants. Its effectiveness depends on capture rates, operating performance, transportation networks, and permanent storage.

Using offsets introduces another layer of accounting rather than changing the plant itself. Stakeholders increasingly distinguish contractual claims from the physical emissions of electricity production.

Enterprise customers should watch this distinction because cloud emissions become part of their own reporting. A workload’s footprint depends partly on where and when it runs.

Companies using AI at scale may need more detailed information about regional generation and hourly emissions. Annual renewable matching can hide important differences between locations.

This is also a knowledge-management challenge inside enterprises. Infrastructure teams, sustainability groups, procurement staff, and executives must work from the same assumptions.

A searchable AI knowledge base can help teams preserve energy contracts, capacity forecasts, and regulatory decisions. It cannot resolve uncertainty, but it can make assumptions visible.

That visibility matters when forecasts change this quickly. BloombergNEF’s sharp upward revision shows how an infrastructure strategy can become outdated within months.

The cost debate will intensify as utilities convert speculative requests into formal investments. The strongest protection is transparent allocation of project costs and cancellation risks.

Without those protections, communities may conclude that AI developers receive preferential access while other customers carry the financial and environmental consequences.

Three Signals Will Show Whether the Gas Forecast Holds

Grid connections, actual campus construction, and gas-equipment delivery will determine whether the 2035 projection strengthens or weakens.

The first signal is the conversion of announced data center capacity into operating load. Press releases and land purchases are not enough.

Investors should watch completed buildings, installed computing equipment, contracted customers, and measured utility demand. A growing gap between announcements and operation would weaken the forecast.

The second signal is the rate of new grid connections. BloombergNEF’s gas outlook depends partly on transmission and generation failing to keep pace with data center construction.

Faster interconnection approvals would allow more projects to use regional electricity. That would not automatically make their power clean, but it could reduce dedicated gas construction.

A persistent connection backlog would strengthen the case for onsite generation. It would also increase the strategic value of campuses near existing power plants and pipeline networks.

Lawrence Berkeley National Laboratory’s forecasts provide a nearer checkpoint. Its 2030 range will become easier to test as equipment shipments and utility loads accumulate.

If national consumption approaches the upper end of that range, the 2035 BloombergNEF scenario will look more credible. Slower load growth would support a lower gas trajectory.

The third signal is firm equipment and fuel infrastructure. Turbine orders, manufacturing capacity, pipeline approvals, and completed plants matter more than preliminary proposals.

Developers can announce a gas-powered campus before securing a turbine delivery slot. Long equipment lead times can push operation several years beyond the intended date.

Pipeline projects face their own approval and construction risks. A plant without reliable fuel delivery cannot serve as dependable onsite power.

The International Energy Agency estimates that 15 to 27 GW of onsite natural gas could serve data centers by 2030, mostly in the United States. Actual deployments within that range will offer an early test.

Technology companies’ clean-energy projects belong in the same measurement. New nuclear, geothermal, renewable, and storage capacity can weaken the gas forecast if it arrives where AI loads are growing.

Contract announcements should not count as completed substitution. The relevant signal is electricity reaching an operating campus during the hours when its computing equipment needs power.

Readers should also watch natural gas prices and utility rate proceedings. Rising fuel costs can improve the relative economics of other generation, efficiency, and workload flexibility.

Regulators may impose stronger financial guarantees on large-load customers. That would filter speculative projects and reveal which developers will accept the cost of their forecasts.

The next several quarters will not settle a projection running through 2035. They can show whether its underlying mechanism is accelerating.

If operating data center load rises, connection queues persist, and gas equipment reaches construction, BloombergNEF’s estimate gains support. If any link breaks, the outcome moves lower.

The BloombergNEF data center gas forecast ultimately measures a choice, not an unavoidable law. The United States can meet AI demand through several combinations of grids, generation, storage, and flexibility.

Natural gas currently offers the fastest familiar route to firm power. Its advantage persists only while alternative infrastructure arrives too slowly.

For developers, policymakers, and enterprise AI buyers, the useful question is concrete: which projects have secured power, who finances that power, and what replaces gas later?

Tracking those answers will reveal whether today’s bridge becomes a temporary response or the enduring energy foundation of American AI.