The AI data center buildout has split into two tiers. The giant projects that define the narrative are bogged down in local fights and construction lag, while the capacity that actually gets bought right now comes from far smaller facilities. For US chipmakers, server vendors, and power-equipment suppliers, the near-term opportunity is moving down in scale and outward in geography, even as the industry's public story stays fixed on gigawatts.
The Gigawatt Plan Is Not Delivering On Schedule
OpenAI and Anthropic are spending billions on massive data center deals, yet they are simultaneously seeking immediate capacity through smaller data center deals, as Tom's Hardware reported. Those smaller facilities run 20 to 30 MW and are meant to provide capacity while mega structures undergo construction, according to the same report. That is the tell. When the two most prominent AI labs are shopping below the size of their flagship commitments, the flagship commitments are not yet producing usable compute at the pace demand requires. The gigawatt project remains the destination. It is no longer the mechanism.
The reason is partly physical and partly political. Construction timelines for very large sites have stretched, and the deals that get signed do not convert to racks and accelerators on the schedule the models need. The labs have responded by buying what is available rather than what is impressive.
Pennsylvania Shows The Political Ceiling
TechCrunch's account of two years of fraught AI data center debates in Pennsylvania is a case study in why the big tier is slow. The piece is built around a simple finding: everyone can find a reason to dislike data center construction. That is not a minor public-relations problem. It is a siting problem. Opposition is not concentrated in one constituency that can be negotiated with, because the objections arrive from different directions and do not share a remedy.
For US technology companies, this means the constraint on AI capacity is not only silicon or capital. It is permission. A project that needs local approval, power interconnection, water arrangements and community tolerance faces multiple veto points, and each one can add months. Pennsylvania is not exceptional in the TechCrunch reporting. It is illustrative. The pattern suggests that the marginal large project in the United States now carries a political discount on top of its construction cost.
Small Sites Become The Pressure Valve
The 20 to 30 MW facilities that OpenAI and Anthropic are pursuing, per Tom's Hardware, occupy a different political and physical footprint. They need less land, less transmission buildout and fewer headline-generating concessions. They can be placed in more jurisdictions, which spreads siting risk rather than concentrating it. That makes them a sensible hedge even though they are expensive per unit of compute and modest in absolute terms.
This matters for the US AI hardware supply chain in a specific way. Large campuses reward different products than small ones. The mega-site economics favor very high-density racks, liquid cooling at scale, and dedicated substation infrastructure. The smaller tier still needs accelerators, but the balance shifts toward equipment that can be deployed quickly in existing or brownfield facilities, where power and cooling are bounded. Vendors whose roadmaps assume only the gigawatt customer may find the near-term order book composed of something more modest.
