Commentary argues AI infrastructure buildout (specifically OpenAI's Stargate initiative) requires continued aggressive data center construction.

Suddenly, much of the AI industry wants to slow down. Earlier this month, Anthropic CEO Dario Amodei called for "pacing the frontier" of AI development over concerns that increasingly powerful models could outpace developers' ability to control them. Other industry leaders have expressed support for greater oversight and stronger safeguards. Lawmakers, too, are scrambling to respond, with some even proposing a pause on advanced AI development and a ban on artificial superintelligence.

At the same time, others are pushing for a different kind of pause: a halt to construction of the data centers that power AI. Facing public backlash and pressure to get tough on the industry, state and local officials are increasingly pursuing bans, moratoria, and other restrictions on new data-center construction.

The two efforts would seem consistent on their face. If AI is advancing too quickly, why keep expanding the infrastructure that makes those advances possible? Yet the push for more AI safety points toward a counterintuitive conclusion: pacing the AI frontier may require more computing power, not less. Policies that restrict that capacity will make the task of ensuring AI safety even more difficult.

Consider what the frontier labs are—and are not—advocating. In his essay calling for an industry-wide slowdown, Amodei argues for more rigorous safety testing and alignment research, as well as third-party evaluators embedded inside AI companies to scrutinize their work. He acknowledges that this may involve limiting the amount of computing power used to train new models, but he notably does not call for a slowdown in data-center construction. In fact, Anthropic has reportedly lined up compute deals worth as much as $517 billion. Just last week, the Swiss bank UBS maintained its forecast that AI-industry capital spending would reach $1.2 trillion in 2027, up from $900 billion this year. If "pacing the frontier" means slowing down the compute buildout, the industry's actions suggest otherwise.

AI safety itself consumes a lot of computing power. Understanding how models behave, testing their safeguards, and monitoring their actions all require immense amounts of compute. Devoting more resources to alignment, safety testing, and evaluation will therefore increase demand for compute even if companies slow the development of ever-more-powerful models.

The recent Hugging Face security incident helps explain why. OpenAI test agents broke out of their restricted environment and hacked into another company's systems. After discovering the breach, OpenAI paused training on its latest models and placed its largest planned reinforcement-learning experiment on hold, instead devoting more resources to safety and alignment work, including expanded monitoring that sends suspicious activity to increasingly sophisticated automated investigators.

Those safeguards require significant compute resources. OpenAI estimates that its expanded monitoring system adds roughly 20 percent to the computing power needed to run the model activity it monitors. In a subsequent report on the incident, the company said it was "investing significantly more compute resources" to examine its models' step-by-step reasoning and to more quickly intervene on "misaligned behavior." The company later extended similar monitoring to its newest model, Astra, "with significant compute cost."

Anthropic researchers have reached similar conclusions. In tests of AI systems designed to monitor other AI systems for malicious behavior, researchers found that "more compute means better detection." Allowing monitors to execute code helped them verify changes they otherwise struggled to detect—for instance, running a program to compare two versions of a file to identify harmful alterations. This more active approach required roughly ten times as many requests to the AI model as alternative methods.

Monitoring a capable AI system is itself computationally demanding. "The way we will achieve safety and alignment for these very capable models is by actually spending more compute on building the safety models," OpenAI's compute strategy chief Sachin Katti said. "If anything, we believe that we will need more compute to make sure that future models are more safe and more aligned."

Another reason demand for computing power could keep growing during a slowdown is that companies would still have powerful models to deploy more widely. Every chatbot response or coding agent requires computing power. Even if frontier capabilities froze tomorrow, there would be enormous work in putting existing models to use. Addressing safety and alignment issues could even strengthen the case for continued AI infrastructure investment. "The buildout doesn't stop because the CEOs asked for guardrails," Erik Kratz of Arena Private Wealth told Reuters. A credible safety framework, he argued, could make long-term infrastructure investments easier to justify.

The growing backlash against data centers undermines that safety work. The same facilities needed to train new models also support the monitoring and alignment work that can make existing models safer.

Policymakers concerned about AI safety should therefore have an interest in making AI infrastructure easier to build. This means establishing workable permitting rules for data centers and requiring developers to pay the costs their projects impose. If policymakers want AI companies to spend more time testing and evaluating their models, they should allow them to build the infrastructure needed to do it.