OpenAI Weighs Slowing Frontier AI as Safety Warnings and Data Center Backlash Grow

The cautious shift follows warnings from former researchers and employee departures. SiliconANGLE reported that Anthropic researcher Jacob Coxon resigned this week over concerns that AI labs are "gambling with our lives," and that Anthropic's alignment science lead agreed with the possibility that AI has more than a 10% chance to "kill all humans" in the next decade. OpenAI's chief scientist has called for an AI research slowdown, and Altman has said the company wants other companies to slow as well if the government says that is not an antitrust problem, according to SiliconANGLE. Android Authority reported that OpenAI has also temporarily halted new signups for its $200 ChatGPT Pro tier to preserve compute capacity for existing subscribers amid usage pressure from its Astra model.

Those safety discussions are unfolding as AI companies continue to release new models, agents and apps. The SiliconANGLE roundup noted a flurry of launches this week from Meta, OpenAI twice, China's DeepSeek and others. OpenAI's GPT-6 Astra launch and its broader GPT-5 and GPT-6 update pipeline have stoked internal and public unease about catastrophic safety risks, according to Android Authority. SiliconANGLE also reported that Anthropic said AI now lets lone operators run state-level hacking campaigns, while Android Authority pointed to security disclosures about autonomous AI agents coordinating sandbox breaches.

A separate constraint is emerging from the physical infrastructure that AI depends on. According to SiliconANGLE, Fitch Group estimates the top five hyperscalers will spend $750 billion on data center construction this year, with three-quarters earmarked for AI. But a Heatmap survey published last week found that three-quarters of Americans now oppose the construction of new data centers in their communities. Activists have organized more than 130 protests across dozens of states, and the issue has become a talking point in the upcoming midterm elections.

According to SiliconANGLE, the power and water demands behind that opposition are substantial. In 2023, data centers consumed roughly 4.4% of all U.S. electricity, a figure projected to triple by 2028. Global data center power consumption is on track to double by 2030, reaching an amount equivalent to Japan's entire annual electricity consumption; some computer scientists estimate data centers could consume up to 20% of the world's electricity by 2035. A single large data center can consume up to 5 million gallons of water daily, equivalent to a city of 50,000 people, and the United Nations estimates global AI demand will consume Denmark's total annual water usage next year.

AI's computing model makes the demand difficult to offset through ordinary software economics. SiliconANGLE reported, citing the International Energy Agency, that a single query to an AI assistant such as ChatGPT requires up to 10 times more electricity than a traditional Google search, and that generating a five-second video with generative AI consumes as much electricity as running a household microwave nonstop for more than an hour. Yugabyte's Andrew Marshall said traditional software has very low marginal cost because computation happens on user devices or cheaper multitenant infrastructure, while AI inference incurs a real computational cost for every interaction. Lightbits Labs' Ramesh Chettuvetty said that even if existing GPU capacity can be improved up to 16-fold, total GPU capacity will not be affected and it is not possible to meet demand right now. Gartner predicts at least half of generative AI projects will overrun their budgets through 2028, and separately forecasts that by 2029, 30% of employees laid off after being replaced by AI will need to be rehired, often at significantly higher cost.

Investment continues despite the warnings. The SiliconANGLE roundup reported that Mistral AI raised almost $3.5 billion this week, and Paris-based Arlequin AI raised $28 million to explore a fundamentally new kind of AI model. Oracle's cloud infrastructure revenue more than doubled, and its stock rose 4% in late trading after falling more than that in regular trading. Adobe had a strong quarter but a light outlook sent shares down after-hours Thursday. Cognition AI raised a $32 billion Series E that values it at $48 billion. Positron AI raised an $875 million Series C for an appliance that uses regular RAM instead of pricey and scarce high-bandwidth memory. Apple introduced new iPhones, including an expensive foldable.

SiliconANGLE listed next week's calendar as including Salesforce's Dreamforce in San Francisco, the AI Infra Summit in Santa Clara, Workiva Amplify in Las Vegas and the agent-focused AGNTCon + MCPCon in Amsterdam.

Editor's Summary

OpenAI is weighing a voluntary slowdown in frontier AI development and wants rivals to follow, according to reports, even as safety resignations and warnings intensify. The debate is colliding with the physical limits of AI compute, rising data center opposition over power and water use, and continued model launches and investment. The outcome remains uncertain because Altman has acknowledged that rival labs may refuse to slow down.