The AI boom's grid problem: why gas turbines are suddenly back-ordered

01 How much power an AI datacenter actually draws

A data center, in the standard definition, is a physical facility housing the systems that process, store, and transmit digital information. The AI buildout has changed what such a facility draws. Traditional cloud halls were designed around racks consuming tens of kilowatts; AI training clusters pack accelerators that each pull on the order of a kilowatt, plus cooling, into the same floor space. A single large AI campus is now specified in hundreds of megawatts, and several announced projects sit at the gigawatt scale — the output of a large nuclear reactor unit.

Aggregate numbers are catching up with the anecdotes. The IEA estimates global data centres consumed about 415 TWh in 2024, up from roughly 290 TWh in 2020, and projects about 945 TWh by 2030 — close to a tripling within the decade. For scale, 415 TWh is on the order of the annual electricity consumption of a mid-sized European country. The 2030 figure is a projection, not a measurement, and it assumes the announced buildout largely proceeds.

What makes AI load different from the cloud era is density and constancy. It runs around the clock, does not tolerate interruptions, and concentrates where cooling, land, and fiber allow. For grid planners, a hyperscale AI campus is less like a normal commercial customer and more like a mid-sized factory that appears in eighteen months.

02 Why turbines, not just solar, are the bottleneck

Gas turbines sit at the center of this story because they are the fastest dispatchable power that can be manufactured at scale. In engineering terms, a gas turbine is a continuous-flow internal combustion engine: air is compressed, fuel is burned in the compressed air, and the expanding hot gas drives a turbine that spins a generator. That simplicity is the point — high reliability, quick ramping, and a mature global supply base that took decades to build.

Solar and wind remain the cheapest new energy in most markets, but AI load is round-the-clock and dense, and firming it today means batteries or long-duration storage at costs and build times many developers will not accept for a 2027 go-live. Grid operators therefore treat gas not as a rival to renewables but as the bridge that keeps new load fed while clean firm capacity and transmission catch up.

The practical pattern in 2025-26 is behind-the-meter generation: developers install on-site turbines to power campuses before grid connections are ready, then synchronize with the grid later. This is interpretation rather than measurement, but the turbine surge is fundamentally a timing problem — the mismatch between how fast AI load arrives and how fast anything else can be built.

03 Order backlogs stretch years

The result is an order book the power industry has not seen in decades. GE Vernova, Siemens Energy, and Mitsubishi Power — the heavy-duty turbine makers that dominate the global market — have publicly reported backlogs stretching multiple years, and industry reporting through 2025-26 describes popular heavy-duty frames as effectively sold out toward the end of the decade, with prices up sharply from pre-boom norms.

Supply cannot sprint, because the limiting parts are heavy metallurgy: castings, rotor forgings, blades, and combustor components that require specialized foundries and certified processes, plus a finite pool of engineers who can install and commission the machines. A new foundry or a new workforce takes years, not quarters.

The consequences are visible upstream and downstream. Energy developers reserve turbine slots speculatively; hyperscalers sign supply agreements directly with equipment makers and utilities; and the power-plant slot itself becomes the gating asset for AI capacity — ahead of chips, land, or fiber.

04 Grid interconnection queues: the hidden delay

Even with hardware in hand, connecting to the grid is its own multi-year wait. New large loads and new generators enter interconnection queues, where grid operators study impacts and order network upgrades. In congested US markets, industry reporting commonly puts the wait at several years, and the studies themselves can take longer than the construction of the data center that triggered them.

Equipment compounds the delay. Large power transformers — the units that step voltage up and down across the network — carry industry-reported lead times around four years, and high-voltage circuit breakers around two and a half, in the current market. These are not exotic components; they are commodity-by-design parts whose supply chains were sized for a world without gigawatt campuses.

This is why behind-the-meter gas became the default workaround: it substitutes a procurement the buyer controls, a turbine order, for one the buyer does not, a queue position. The strategy defers rather than solves the interconnection problem — campuses still need the grid for resilience and for growth beyond the first phase.

05 Clean-energy pledges meet reality

The major technology companies built part of their public identity on clean-energy commitments, including round-the-clock carbon-free targets. AI demand is testing them in public. To power new campuses on schedule, companies have renewed or extended gas plant contracts, kept fossil units running longer than planned, and signed power deals that fall short of their stated timelines for carbon-free supply.

The clean-firm response is real but slow. Nuclear restart agreements and small-reactor purchases announced since 2024 signal genuine demand for always-on carbon-free power, yet almost none of that capacity arrives before 2030. Meanwhile the accounting is shifting: buyers increasingly report market-based emissions alongside location-based ones, and the gap between the two is widening.

Our reading is that the pledges are being reconciled with physics rather than abandoned. The likely steady state is gas as an explicit, temporary bridge, with its emissions on someone's ledger — a politically uncomfortable but arithmetically honest position that most large buyers have quietly adopted.

06 Efficiency offsets: better chips vs more chips

Efficiency is the counterargument, and it is not nothing. Each accelerator generation delivers more compute per watt, inference techniques such as quantization and distillation cut the cost of serving a query, and model families keep shipping smaller variants that handle routine work. Energy per unit of useful computation has fallen substantially every year of the boom.

The catch is the rebound effect, and the IEA projection already embeds it: cheaper intelligence increases usage faster than efficiency reduces its cost. Total data-centre demand keeps climbing because the number of tokens, queries, and agent steps grows much faster than the energy per unit falls. Efficiency widens what is possible; it has not yet narrowed the buildout.

Through 2030, algorithmic gains are best treated as a second-order offset against fleet growth — meaningful for unit economics, small against the construction wave. The demand curve is set by how much agentic work the economy absorbs, not by how efficiently each step runs.

07 The buildout through 2030

The buildout math follows the IEA trajectory: roughly 945 TWh of data-centre demand projected for 2030 against 415 TWh in 2024, with the United States and China absorbing the largest shares and AI the dominant driver of growth. Meeting it requires turbines that are already back-ordered, transformers on multi-year leads, and interconnection queues moving faster than they do today.

The credible path is a sequence, not a choice: gas bridges new load through the decade, grid equipment capacity and interconnection reform slowly catch up, and clean firm power — nuclear above all — begins to scale into the early 2030s. Every step in that sequence has a visible bottleneck, which is why completion dates keep slipping rather than projects collapsing.

The strategic conclusion is uncomfortable for the industry but hard to avoid: through 2030, the binding constraint on AI looks less like chips and more like electrons. The signals to watch are turbine backlog clearance, interconnection rule reform, transformer manufacturing capacity, and final investment decisions in nuclear — each a leading indicator of whether the projected curve bends.