The AI Lab Effect: How OpenAI and Anthropic Are Reshaping Data Center Demand Through 2030
Two companies with no data centers of their own five years ago are now driving some of the largest infrastructure commitments in history.
By Debra Brewster | Partner / CEO, Axiom AI Group USA, LLC
It is worth sitting with the scale of what has happened here. OpenAI and Anthropic are, first and foremost, AI research and product companies. Neither built its own data centers a few years ago. Today, the infrastructure commitments tied to their growth are among the largest capital deployments in the history of the industry, and the ripple effects are reshaping site selection, power procurement, and construction timelines well beyond the two companies themselves.
OpenAI’s disclosed hardware and cloud infrastructure spending commitments now total more than
$1.09 trillion running from 2025 through 2035, spread across partnerships with Oracle, Microsoft, Broadcom, Nvidia, AMD, Amazon Web Services, and CoreWeave. As of last September, the company had contracts in place for roughly 8 gigawatts of capacity by 2028, and its Stargate joint venture with Oracle and SoftBank has reportedly reached about 7 gigawatts of planned capacity toward a stated 10-gigawatt target, with a cumulative investment estimate exceeding $400 billion. That buildout has not been without friction: a major Stargate-linked project was set back this year when New Mexico blocked a planned gas pipeline meant to fuel its on-site power plant.
Anthropic’s approach has been more deliberately diversified across cloud partners, but no less ambitious in scale. The company published research this year arguing that the US needs at least 50 gigawatts of additional AI electric capacity by 2028 to maintain its competitive position, and that training a single frontier model will require data centers with 5 gigawatts of capacity by that same year. To meet its own growth, Anthropic expanded its partnership with Google and Broadcom this year to secure multiple gigawatts of next-generation TPU capacity beginning in 2027, building on its existing Project Rainier training cluster with Amazon Web Services and a previously announced $50 billion commitment to US computing infrastructure. The company’s run-rate revenue reportedly surpassed
$30 billion this year, up from roughly $9 billion at the end of last year, growth that is directly driving its infrastructure appetite.
The combined effect of these two companies alone helps explain why the broader industry figures look the way they do. Electric Power Research Institute modeling suggests data centers could consume between 9 and 17 percent of total US electricity generation by 2030, a range revised sixty percent higher than the same institute’s 2024 estimate. Anthropic itself has been notably candid about the tradeoffs involved, publicly committing to help offset electricity price increases in communities affected
by its data center buildout, and cautioning against simply exporting the largest AI training runs to facilities overseas as a way of avoiding domestic infrastructure constraints.
Looking toward 2030, the practical implications for the broader data center and infrastructure development industry are significant regardless of exactly how the competition between the major AI labs plays out. Both companies are hiring seasoned data center executives away from established players like Google and major developers such as Stack Infrastructure, signaling they intend to exert direct influence over site selection, power sourcing, and construction standards rather than simply purchasing capacity as a passive tenant. That shift toward direct involvement by the AI labs themselves, rather than only their cloud and hyperscale partners, is likely to be one of the more consequential structural changes in how large-scale AI infrastructure gets planned, financed, and built between now and the end of the decade.
For developers, landowners, and municipalities evaluating opportunities tied to this wave of demand, the lesson is straightforward: the AI labs driving this growth are no longer distant customers several layers removed from a project. Increasingly, they are direct counterparties with their own infrastructure teams, their own site criteria, and their own multi-billion-dollar balance sheets behind the commitments. Understanding their specific power, timeline, and scale requirements has become as important to a successful project as understanding any traditional hyperscale tenant, and arguably more so, given how quickly their capacity needs continue to grow.