AI CEOs call for slowing AI boom as billions flow to data centers

Amodei's proposal

"A race to the bottom, spurred by commercial incentives, can make these risks more acute," Amodei wrote of his concerns in his proposal this weekend. AI companies "must slow the pace at which we improve the capabilities of AI models" to give safety measures time to catch up, Amodei added. He didn't call for an end to model training or technical progress.

As part of the proposal, Amodei said Anthropic would give embedded third-party evaluators ongoing, employee-like access to assess its safety practices and the ability to report incidents and assess the alignment of models and training processes. He proposed that other frontier AI companies give independent evaluators similar access. He also called for coordination among frontier companies and democratic governments around safety standards and limits on unchecked capability gains.

OpenAI CEO Sam Altman said pacing the frontier has been a primary topic of recent discussions at OpenAI. He also said OpenAI would adopt independent evaluators with employee-like access. In an X post, SpaceX CEO and xAI founder Elon Musk responded simply: "Dario is right."

AI commitments are already baked into the buildout

The CBRE "North America Data Center Trends H1 2026" report, released Aug. 27, found that primary-market supply reached a record 10,903 megawatts (MW) in the first half of 2026, up 33.7% from a year earlier, while vacancy fell to a record-low 1.4%. Hyperscale and AI occupiers drove demand as they competed for large amounts of power.

Developers had 7,481 MW under construction at the end of the first half, a record high, and 80.4% of that capacity was already preleased. Less than 1,500 MW remained available for preleasing across the primary markets, according to CBRE.

Much of that capacity already has a customer attached to it. The workloads, utilization rates and expansion plans behind those leases can still change. A project designed around a large frontier training run might serve inference or enterprise workloads, operate at a different utilization rate or come online later than expected.

The commitments aren't equivalent. Some are cloud-spending targets, others are capacity agreements, development plans or operating facilities. Those differences determine who carries the risk if demand shifts in timing or workload mix.

Anthropic said in April that its annualized revenue run rate had surpassed $30 billion and that demand for Claude had accelerated so quickly that its infrastructure was under strain. The company has committed more than $100 billion over 10 years to AWS technologies and secured access to up to 5 gigawatts (GW) of capacity to train and serve Claude.

Anthropic also announced a $50 billion investment in U.S. computing infrastructure with Fluidstack and committed to purchase $30 billion of Azure compute capacity from Microsoft, with additional capacity of up to 1 GW.

OpenAI has disclosed similarly large infrastructure ambitions. Its Stargate project intends to invest $500 billion over four years in new U.S. AI infrastructure. In April, OpenAI said it had already surpassed its original goal of securing 10 GW of U.S. AI infrastructure by 2029, with more than 3 GW added during the preceding 90 days.

Musk's xAI has also expanded rapidly, saying its Colossus supercomputer in Memphis doubled to 200,000 GPUs in 92 days and could eventually reach 1 million GPUs.

The announcements represent different kinds of commitments. Some involve cloud spending, some capacity agreements and some physical infrastructure. They don't guarantee that every planned megawatt will be needed at the same time or for the same purpose.

By the time an AI customer revises a training plan, a developer might already have tied up land, ordered transformers and committed capital to a substation plan. The resulting exposure is therefore less about the disappearance of AI demand than about whether demand arrives soon enough, at sufficient scale and in a form that matches the asset.

Julie Bolthouse, director of land use at the Piedmont Environmental Council, said Virginia illustrates another distinction that can get lost in discussions of AI demand: Some data centers have been built but can't obtain enough power to operate.

"Most of the colocation companies … are primarily being leased up by hyperscalers like AWS, Microsoft, Meta, Oracle and Google," Bolthouse said. The next question, she said, is "leased up by who?"

Bolthouse said her research indicates that hyperscaler leasing is largely supported by compute agreements with frontier AI companies, including Anthropic and OpenAI. Demand could therefore be visible at the hyperscaler level while actual use of capacity still depends on an AI company's product, training and deployment schedules.

Var Shankar, senior director analyst at Gartner, said a coordinated slowdown at the frontier wouldn't necessarily reduce overall AI infrastructure demand. It could instead shift compute demand from training cutting-edge models toward deployment at scale.

"This episode illustrates why companies should already be building optionality into major AI capacity commitments," Shankar said. "They should emphasize staged investments, flexible capacity and regular reassessments."

Large frontier-training runs can concentrate enormous demand in a smaller number of campuses. Inference and enterprise workloads can distribute demand across cloud regions and expand in smaller increments. A shift between those workloads could preserve aggregate demand while changing the economics, location and operating profile of the infrastructure required.

Shankar said tech entrepreneur and venture capitalist David Sacks has pushed back against government-supported coordination to pace the frontier.

Even if frontier development slows materially, Shankar expects substantial demand from inference, enterprise deployment and agentic workloads. Companies will scrutinize that demand more closely, he said, with greater emphasis on measurable business value and sustainable operating economics rather than expectations about what the next generation of frontier models might enable.

The exposure isn't evenly distributed. A developer with a long-term, creditworthy hyperscaler lease might be protected even if the hyperscaler's AI customer changes plans; the utilization risk could instead sit with the cloud provider. More vulnerable are speculative or partially contracted projects, phased campuses dependent on future expansion decisions, and investments tied to one concentrated customer or a specific training-driven load forecast.

Physical constraints can magnify timing risk

Power availability, transmission, interconnection, permitting and zoning can delay projects even when customers are ready to use capacity. Those delays make it harder to shift infrastructure if AI demand changes direction.

An Annenberg Public Policy Center survey released in August found that 61% of Americans somewhat or strongly opposed building new data centers in their communities, up from 49% in a February-March survey. The nationally representative survey covered 1,320 U.S. adult citizens and opposition crossed party lines, with majorities of Democrats, Republicans and independents opposed.

In July, New York paused consideration of certain discretionary state environmental permit applications for new hyperscale data centers of 50 MW or more for up to one year, with exceptions for applications already deemed complete, while the state develops a new review framework.

Power remains the biggest constraint on Virginia's data center pipeline, according to Bolthouse.

The Piedmont Environmental Council (PEC) tracks roughly 280 million square feet of mostly approved but unbuilt data center space in Virginia, Bolthouse said. PEC says Dominion Energy has committed to roughly 27 GW of data center load with in-service dates before 2032, while its broader pipeline includes more than 70 GW of data center delivery-point requests. Many of those requests don't yet have connection dates.

CBRE likewise said U.S. capacity demands are significantly outpacing utility transmission buildout, with developers facing energization delays as new last-mile projects encounter permitting, planning and zoning hurdles.

CBRE said the 80.4% of the capacity under construction that's already preleased reduces developers' exposure on those projects compared with campuses still waiting for customers, power or both. It doesn't eliminate risk if lease commencements, expansion options or utilization assumptions depend on a particular pace of AI development.

Shankar expects inference, enterprise deployment and agentic workloads to keep generating demand even if frontier development slows. Those workloads can require different locations, power profiles and delivery schedules from large training runs.

Anthropic is already expanding Claude inference capacity in Asia and Europe and exploring data centers in other democratic countries. Its infrastructure commitments extend beyond the U.S. frontier-training buildout, giving the company more ways to shift where and how it deploys compute.

The physical infrastructure is harder to move. An AI company can delay or resize a training cycle quickly. A developer, utility or power producer that has already committed land, transformers, generation and financing can't adjust nearly as fast.

Shane Snider is a senior news writer at TechTarget, covering AI infrastructure, hyperscale data centers, cloud platforms, and the power and energy systems driving modern compute expansion. You can reach Shane at [email protected] or on LinkedIn.