Anthropic and OpenAI are reportedly pivoting their data center strategies, shifting focus from colossal, multi-gigawatt deployments to smaller, more agile 20-30 megawatt (MW) deals. This move signals a nuanced evolution in the relentless pursuit of AI infrastructure, driven by the need for faster deployment and the increasing demands of inference workloads.
Sources familiar with the matter revealed to CNBC that both AI powerhouses are actively exploring smaller compute capacity agreements. Anthropic has been in discussions for such deals across the United Kingdom and the Nordic region, while OpenAI has also been evaluating opportunities in the Nordics for these scaled-down deployments. Furthermore, there are indications of discussions involving both companies regarding similar-scale capacity deployments within the United States.
This strategic recalibration comes after a year of significant investment in hyperscale data centers. OpenAI, for instance, has been instrumental in the development of vast AI infrastructure projects, exceeding its initial 10 GW commitment and securing additional capacity in Georgia and Ohio. Similarly, Anthropic announced a substantial $45 billion cloud deal with Nscale for approximately 460 MW of compute capacity in West Virginia.
However, the sheer scale of these massive projects is increasingly encountering headwinds. Numerous large-scale data center developments in the U.S. and Europe are facing significant pushback from local communities concerned about environmental impact and resource strain. Additionally, the scarcity of available land and power in many European regions is adding to the logistical challenges.
The appeal of smaller capacity deals, according to Jabez Tan, head of research at Structure Research, lies in the “speed to usable capacity.” He explained that securing a few megawatts at an existing, powered site can often be a more practical and time-efficient solution compared to waiting for the construction of much larger facilities. “For workloads that can operate across separate sites, a collection of smaller deployments can add up to substantial capacity,” Tan noted.
This shift also aligns with a growing trend in the AI industry: the increasing prominence of inference workloads. While training large AI models demands immense computational power, the day-to-day operation of these models, known as inference, can be effectively handled by smaller clusters of chips.
“Training a large model typically requires many chips working closely together,” Tan elaborated. “Many inference workloads can instead serve separate requests across multiple smaller clusters, opening up more locations.”
The implications of this evolving demand are significant. As more AI compute shifts from model training to production deployment, the capacity dedicated to inference is expected to surge. Real estate advisory firm JLL projects that inference workloads will constitute a larger proportion of total data center capacity than training workloads by 2027. Their report indicates that in 2025, inference accounted for 9% of global data center workloads, compared to 14% for training. By 2030, this figure is projected to rise to 37% for inference, while training is expected to utilize only 13%.
This trend is already being reflected in industry developments. Nvidia, a key player in AI hardware, announced in February its collaboration with data center stakeholders to explore smaller-scale data centers optimized for distributed inference. Furthermore, U.S.-based Crusoe, known for building large data centers for OpenAI, is reportedly investing in smaller, more agile facilities that promise faster deployment and reduced costs, a strategic move to circumvent the delays plaguing larger construction projects. Crusoe recently secured a substantial $3.9 billion funding round, underscoring the burgeoning demand for innovative data center solutions.
OpenAI acknowledged the evolving landscape in its statement, noting, “We’re building a diversified compute portfolio to meet growing demand for AI around the world.” The spokesperson added, “Different workloads need different infrastructure, so we have conversations with a range of partners and assess opportunities based on our requirements, performance, reliability, timing and cost.” While they declined to comment on specific commercial discussions, the company’s emphasis on a diversified strategy clearly indicates an adaptation to the dynamic needs of the AI sector.
Original article, Author: Tobias. If you wish to reprint this article, please indicate the source:https://aicnbc.com/25874.html
