AI has already escaped the software stack.
It is now bidding for transformers, substations, cooling water, land, transmission, generators, gas pipelines, turbines, construction crews, and years on permitting calendars.
Global Energy Monitor's August update makes the physical scale visible. The organization now tracks 378 gigawatts of U.S. gas-fired power capacity in development, up from 252 GW in January. It says 189 GW of the pipeline is intended to meet data-center demand.
That is a spectacular number. It is also not the same thing as 378 GW of future operating plants.
Development is a funnel, not a promise.
GEM's total spans projects at different stages: announced, pre-construction, and construction. Those categories have radically different probabilities of reaching commercial operation.
An announced project can disappear because a customer changes plans, a permit fails, financing changes, a turbine delivery slips, a pipeline cannot be expanded, a transmission study changes the economics, or the underlying data-center demand never arrives.
The useful metric is therefore not only capacity announced. It is capacity surviving each gate.
announcement → site control → permits → fuel → equipment → interconnection → construction → commissioning → sustained load
Every arrow is an attrition point.
AI demand changes the schedule.
Traditional power planning assumes long-lived load growth and utility timelines measured in years. Large data-center developers often want power much faster.
That mismatch creates pressure for generation that can be built near a campus, contracted directly, or developed alongside the load. Gas plants can look attractive because they are dispatchable and familiar to utilities and project financiers.
But “faster than transmission” does not mean “fast.” A gas plant still needs equipment, environmental review, fuel delivery, construction, controls, grid protection, and often some form of interconnection.
The AI buildout therefore shifts the bottleneck rather than removing it.
The turbine is becoming part of the compute supply chain.
Model developers talk about accelerators. Infrastructure developers increasingly have to talk about prime movers.
Large gas turbines are specialized industrial machines with finite manufacturing capacity and long procurement cycles. If many utilities and data-center developers order at once, turbine slots can become a schedule constraint in exactly the same way advanced transformers or switchgear already can.
This creates an uncomfortable systems truth: a software company can have capital, chips, land, and customers and still be delayed by an industrial component manufactured in a completely different sector.
Fuel is an architecture decision.
A plant does not consume “natural gas” in the abstract. It consumes gas at a specific pressure, volume, location, and time.
That means pipeline capacity and local delivery infrastructure become part of the data-center architecture. A proposed generator may be technically feasible while the regional gas network is not ready to serve it at full output during peak conditions.
Behind-the-meter generation can reduce dependence on a congested transmission queue while increasing dependence on a fuel network that has its own weather, compressor, storage, and pipeline constraints.
There is no infrastructure-free option. There are only different dependency trees.
The grid may inherit stranded assumptions.
Data-center forecasts are unusually uncertain because individual projects can be enormous and because AI hardware efficiency, model architecture, utilization, and business demand are all moving at once.
If utilities or independent developers build generation around demand that arrives late or never reaches the forecast level, ratepayers and investors can inherit oversized assets. If they underbuild and demand arrives faster than expected, interconnection delays and reliability concerns intensify.
The planning problem is therefore not simply “AI needs more power.” It is matching lumpy, uncertain digital load to slow, capital-intensive physical assets.
Count completion, not announcements.
The 378 GW pipeline is most useful as a leading indicator of pressure. It shows what developers believe might be needed and where capital is trying to go.
To understand what actually happens, Cyberdelia will track the funnel: projects canceled, projects entering construction, turbine orders, pipeline expansions, interconnection agreements, commercial-operation dates, and the share of announced data-center load that becomes real metered demand.
That turns a giant headline number into a falsifiable industrial forecast.
The AI electricity story has entered a second physical phase. The first was demand: data centers require enormous amounts of power. The second is buildability: the proposed response now depends on turbines, fuel networks, permitting, interconnection, financing, and construction capacity. The strategic number is not 378 GW announced. It is the fraction of that pipeline that reaches operation on time, at the locations where compute actually materializes.
What we want next.
The next useful dataset is project-level attrition. For each data-center-linked gas proposal: stage, developer, intended customer, turbine supplier, fuel source, interconnection status, construction start, expected commercial-operation date, and whether the associated data-center load is contracted or speculative.
That is where the AI map stops being a press-release map and becomes an industrial one.