The market spent 2025 arguing about whether there were enough chips. That argument is over, and it was the wrong one. The binding constraint on artificial intelligence in mid-2026 is not silicon, not capital, and not data. It is a physical connection to an electricity grid that was not built for this.
The problem is a wire, not a wafer
Nvidia will sell you compute. Capital is so abundant that the IEA notes the largest tech firms’ capex exceeded $400bn in 2025 and is expected to jump by another ~75% in 2026. None of that money buys a connection to the grid if the substation is full.
The bottleneck has also spread along the electrical chain. It is no longer a single missing component: the IEA flags both transmission lines and the transformers and cables that connect to them. The shortage spans the path from the substation to the server, not one chokepoint that a clever procurement team can route around.
The numbers are concrete. Data centres consumed around 415 TWh in 2024, about 1.5% of world electricity, and the IEA expects that to more than double to roughly 945 TWh by 2030. Consumption has grown about 12% a year since 2017. The load is also concentrated: the United States accounted for ~45% of data-centre electricity in 2024, China ~25%, Europe ~15%. AI demand piles onto grids that are already the tightest.
Here is the part the GPU headlines skip. The IEA warns that around 20% of planned data-centre projects could be delayed by grid constraints. A new transmission line takes four to eight years to build in advanced economies. Wait times for transformers and cables have doubled in the past three years. A GPU rack ships in weeks. The 765-kV transformer that feeds it does not.
Analysis: the queue is the new moat
Compute scarcity was always going to be temporary. Fabs expand, yields improve, competitors ship accelerators. Grid scarcity is structural, because the limiting inputs are permitting, copper, skilled line crews and large power transformers — none of which respond to a funding round.
So the hyperscalers stopped waiting in the interconnection queue and started buying their own power. The clearest signal came in January, when Vistra and Meta signed 20-year agreements for more than 2,600 MW of nuclear from the Perry, Davis-Besse and Beaver Valley plants — 2,176 MW of operating output plus 433 MW of uprates, deliveries beginning late 2026. That deal is one slice of Meta locking up to 6.6 GW of nuclear across Vistra, Oklo and TerraPower. When a software company signs a two-decade contract on a reactor’s entire output, it is not buying clean energy for the brand. It is buying certainty of interconnection that the public grid cannot promise.
This reframes the whole supply chain. The scarce asset is no longer the accelerator at the top of the stack. It is firm, dispatchable power with a guaranteed point of connection. That is why the procurement scramble has moved upstream — to existing nuclear plants with live grid ties, to gas turbines whose air permits are the actual gating item, and to small modular reactors that mostly do not exist yet. The IEA’s own supply math is telling: to 2035 it sees renewables adding over 450 TWh to meet data-centre demand, with natural gas (~175 TWh) and nuclear (a similar amount) filling the firm-power gap. Gas and nuclear are doing the heavy lifting precisely because intermittent supply does not satisfy a 24/7 training cluster.
The catch is that the firm-power options each carry their own queue. Restarting or uprating an existing reactor depends on regulatory sign-off and is rare by design — the Meta–Vistra uprates are billed as the largest nuclear capacity additions ever backed by a corporate buyer. New gas turbines are gated less by the machines than by air permits and pipeline capacity. Small modular reactors are real on paper and largely unbuilt in practice, with first commercial output years out. So the captive-power strategy does not remove the constraint. It swaps a public interconnection queue for a private one, available only to buyers who can write a twenty-year cheque against a power plant.
Implications: who actually wins
Three consequences follow, and none of them are priced into a pure GPU thesis.
First, the centre of gravity in AI infrastructure shifts toward whoever controls firm generation and transmission rights. Independent power producers, nuclear operators and turbine makers become structural beneficiaries of the AI build, not bystanders. A GE Vernova or a Vistra is now an AI infrastructure name whether it likes the label or not.
Second, the captive-power workaround splits the market. The handful of companies with the balance sheet to sign 20-year PPAs and underwrite new reactors get to keep building. Everyone else sits in an interconnection queue that, in the tightest US markets, runs many years. Capital abundance does not fix this; it concentrates the advantage with the firms that can self-supply. The grid becomes a filter on who is allowed to scale.
Third, the constraint is geographic and slow to relocate. The US carries the largest share of the load and the longest transmission lead times at once. Capacity will migrate toward jurisdictions that can actually energise it — places with spare firm generation, faster permitting, or a willingness to let private builders bypass the public grid entirely. That is a quieter sovereignty story than chip export controls, but it may matter more for where the next training run physically happens.
The tell for the next twelve months is not the next benchmark score. It is the interconnection agreement, the transformer order book, and the air permit. AI’s growth curve now runs through a switchyard. The labs that understood that early bought reactors. The ones that did not are still holding their place in line.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.