A creative director asks when the new AI feature ships. The honest answer, more often now, is: when the substation clears. Not when the model is trained, not when the interface is approved, but when a utility signs off on enough megawatts to run the racks that serve it. That is not a metaphor. It is becoming the actual critical path on a growing number of launches, and most marketing calendars have no line for it.

The spend is real, the power isn't waiting

U.S. ISPs spent $92.6 billion on capex in 2025. Google alone is planning to spend $205 billion in 2026. Those are not marketing budgets or content budgets, they are infrastructure budgets, and the infrastructure in question is increasingly AI specific: racks built for inference and training loads that draw far more power per square foot than the data centers most of us grew up assuming were interchangeable and fungible.

That spending gap between what companies want to build and what the grid can currently deliver is the thing agencies and brand teams need to internalize. A budget number that big does not mean the capacity shows up on schedule. It means demand for power is now large enough to be a line item that utilities, regulators and local governments all have opinions about, and none of those parties answer to a launch date set by a marketing team six months ago.

The instinct in our business is to treat capex announcements as a proxy for readiness. If a company says it's spending $205 billion, surely the thing gets built. But capex is a commitment to spend, not a guarantee of delivery, and the gap between the two is exactly where power politics lives.

Four systems, one bottleneck

Every data center runs on four core systems, and new chips are changing all four of them at once, not just the compute layer marketers tend to think about. Power delivery, cooling, networking and the physical rack itself all have to be redesigned together when the chips inside get denser and hungrier. That is not a software update. It is a construction project, and construction projects have permitting timelines, utility timelines and neighborhood timelines that don't move at the speed of a product roadmap.

Denser racks and grid delays are forcing what the industry is now openly calling an infrastructure rethink, not an optimization. When the rack changes, the cooling changes. When the cooling changes, the power draw changes. When the power draw changes, the interconnection request changes, and that request goes back into a queue that a marketing calendar has no visibility into and no leverage over. Every one of those dependencies is sequential, not parallel, which is exactly why "we're building it now" and "it will be powered by launch" have become two separate claims that need to be checked separately.

Politics now sits between you and your launch date

Builders confronting the politics of power is the part of this story that doesn't show up in a product brief. A data center's ability to get built, and to get powered, increasingly depends on local approval, utility capacity planning and the willingness of a community to host the substation upgrades a new facility requires. None of that is under the control of the company promising you a Q3 feature, and none of it moves faster because a campaign has already been booked around it.

The chip is ready. The model is ready. The substation is not.

That is the sentence agencies should start expecting to hear from clients, and it is worth saying out loud in planning meetings before it becomes an excuse delivered after a missed date. A local zoning board that meets once a month and a utility capacity study that takes a quarter are now, functionally, part of the marketing supply chain.

The governance layer is catching up, unevenly

Regulators are aware this is happening, even if their timelines don't match the industry's. The FCC has relaunched its Tech Advisory Council specifically to guide policy on spectrum, AI and 6G, which is an admission that the old guidance doesn't cover what's being built now. The House Commerce Committee has cleared five telecom bills, another sign that legislative attention is catching up to infrastructure that outran the rules written for it. A bipartisan coalition has separately called on House leaders to advance AI guardrails, which tells you the appetite for oversight is not confined to one party or one chamber, even if the bills themselves move slowly.

Meanwhile a StateScoop report on state CIOs found that AI and modernization pressure are reshaping how state governments think about IT investment and governance, which matters because state and local approval is often exactly where a data center's power request lives or dies. A state that is rethinking its own IT governance is also a state that is rethinking how quickly it says yes to a private company's substation upgrade.

None of this produces a single predictable timeline. It produces the opposite: a patchwork of state and local processes, each with its own pace, each capable of adding delay to a launch that a national campaign calendar assumed would be uniform. A feature that ships on time in one region and slips two quarters in another is no longer a coordination failure on the agency's part. It's the actual shape of the infrastructure it depends on.

What this means for the people who plan campaigns

If your launch, feature, or campaign depends on new AI infrastructure coming online, the planning conversation has to include a question that used to belong entirely to engineering and real estate: has the power been secured, and by whom. Not "is the data center built" but "has the interconnection agreement cleared," because those are now different milestones with different, less predictable timelines, and only one of them determines when the thing actually turns on.

  • Build buffer into launch dates that depend on new AI infrastructure, not just new code, and treat power availability as a dependency you track, not a detail you assume someone else is handling.
  • Ask vendors and partners directly whether their capacity is contracted and interconnected, or merely planned and budgeted, since a $205 billion spending plan and a signed grid agreement are not the same thing and shouldn't be presented as if they were.
  • Watch local and state approval processes the way you'd watch a competitor's launch, because a permitting delay in one city can quietly become your delay too, and you won't hear about it until it's already late.
  • When a client tells you a feature is "on track," ask which of the four systems, power, cooling, networking, or rack, is the one still waiting on sign off. There is almost always one.

The roadmap hasn't stopped being a technology document. It has just added a utility bill and a zoning hearing to the list of things that can move the date, and the sooner that shows up in the planning deck, the fewer surprises land on launch week.