The Grid Decides What AI Gets Built

Everyone covering the AI data centre boom keeps writing about money, because the money is enormous and easy to report. U.S. ISPs spent $92.6 billion on capex in 2025. Google alone says it will spend $205 billion in 2026. Those numbers get quoted as proof that the industry is unstoppable, that capital is the only constraint left. It isn't. The constraint that actually decides what gets built, where, and on what timeline is the electricity grid, and grid decisions are made by utilities, regulators and local officials who do not answer to anyone's roadmap.

The spending figures hide the real bottleneck

$205 billion from a single company in a single year is the kind of figure that makes capex sound like the whole story. It isn't. Money can buy chips, land, cooling systems and construction crews. It cannot buy a faster interconnection queue from a regional grid operator, and it cannot conjure new transmission lines into existence on a timeline that matches a chip refresh cycle. Reporting on denser racks and grid delays has already documented the industry rethinking its own infrastructure assumptions, because the assumptions that worked for a data centre built five years ago do not hold for the power draw of the racks going in now.

A separate piece on data centre core systems makes the same point from the engineering side: new chips are changing all four core systems inside a facility, not just the compute layer. That is not a story about semiconductors. It is a story about power distribution, cooling and physical plant design all having to be rebuilt around a chip generation that draws more electricity per rack than the building was designed to carry. When the hardware outpaces the building, the building becomes the bottleneck, and the building's power supply is not something a chip vendor controls.

Power is political in a way capex never was

Coverage of builders confronting the politics of power gets at something the spending totals cannot: a grid connection is not a market transaction, it is a negotiation with public institutions. Utilities have obligations to existing customers. Regulators have rate cases and reliability standards. Local officials have constituents who will notice a new substation, a new transmission corridor, or a rate increase tied to a facility most of them will never see the inside of. None of that moves at the speed of a funding round.

Put plainly: a company can have $205 billion sitting in its capex budget and still be stuck at the back of a queue set by a grid operator with dozens of other demands to balance. The money buys intent. It doesn't buy a connection date. The money determines what a company wants to build. The grid determines what it is allowed to build, and when. Those are different questions, and right now only one of them is getting serious daily coverage.

Federal policy is starting to notice, slowly

Two pieces of federal activity this cycle show the policy apparatus beginning to catch up, even if neither directly touches the interconnection queues. The FCC has relaunched its Tech Advisory Council specifically to guide spectrum, AI and 6G policy, which puts AI infrastructure formally inside a body that already has standing relationships with utilities and grid-adjacent telecom infrastructure. Separately, a bipartisan coalition has called on House leaders to advance AI guardrails, a sign that Congress recognises the AI buildout has moved past the stage where it can be left entirely to state utility commissions and corporate siting decisions.

Neither of these is a grid policy in itself. But both point at the same recognition: the AI industry has grown large enough, fast enough, that its infrastructure choices are now a matter for federal bodies that were not built with data centre power demand in mind. The advisory council and the guardrails push are early moves in a fight that has not yet been framed as a grid fight, even though that is what it will become.

States are already rewriting the governance question

A report on state CIOs finds them actively rethinking IT investment as AI and modernisation reshape governance. That is a quieter story than the capex figures, but it is arguably closer to where the real decisions land. State governments run the procurement processes, the permitting offices and, in many cases, the public utility commissions that will decide whether a proposed data centre gets its power allocation approved and at what rate. If state CIOs are rethinking governance around AI now, that rethink will eventually intersect with state-level energy policy, because the two cannot be separated once a data centre proposal reaches a siting board.

This is the level at which the grid politics angle becomes concrete, not "AI needs more power" as an abstract statement, but a specific state utility commission deciding whether to approve a rate structure that lets one company's data centre draw enough electricity to affect the bills of every other customer on that grid. Those hearings do not make national headlines the way a $205 billion spending commitment does. They are, nonetheless, where the actual limits on AI infrastructure get set.

Congress is legislating telecom while the grid question waits

The House Commerce Committee has cleared five telecom bills this cycle, which is a reminder of how much legislative attention is still organised around telecom's old categories: spectrum, carriers, consumer protection. That work matters, but none of it is framed around the specific problem of a data centre operator trying to get power approved fast enough to match a construction schedule. The FCC's relaunched advisory council touches AI and spectrum together, which is progress, but spectrum policy and grid interconnection policy are governed by entirely different bodies, different statutes and different timelines. A company can win every regulatory fight over spectrum and connectivity and still be sitting on an empty building waiting for a substation upgrade.

That mismatch is the actual policy gap here. The institutions moving fastest on AI right now, whether that's a bipartisan guardrails push or a telecom committee clearing bills, are not the institutions that control power allocation. The institutions that do control it, state utility commissions and regional grid operators, are not built for the pace this industry is trying to set, and nothing in the current legislative activity changes that.

What the numbers actually add up to

$92.6 billion in ISP capex and $205 billion from one company are real figures, and they describe a real willingness to spend. But spending totals answer the question of ambition, not the question of delivery. The stories about denser racks forcing an infrastructure rethink, about all four core systems inside a data centre needing to change, about builders confronting the politics of power directly, are the stories that answer the delivery question. And the answer they give is that delivery is gated by grid capacity and by the political processes that allocate it, not by how much any single company is prepared to spend.

The policy conversation happening in Washington right now, the AI guardrails push, the FCC's advisory council, the telecom bills moving through committee, is necessary but incomplete. It is being conducted largely without reference to the fact that the physical limit on how much AI infrastructure gets built in the next few years is set by utilities and state regulators working through interconnection queues one application at a time. Until that becomes the centre of the policy conversation rather than a footnote to the capex figures, the industry's actual pace will keep being decided somewhere other than where the public debate is happening.