Anthropic and OpenAI are exploring smaller AI data center deals as competition for access to the infrastructure needed to deploy workloads intensifies, people told CNBC.
Both AI labs have secured large contracts for AI data centers in the past year for facilities in the hundreds of megawatts and gigawatts, but the companies are now also seeking contracts for computing capacity for much smaller deployments of 20 to 30 megawatts, the people said.
Anthropic is talking about agreements within that range in the U.K. and across the Nordic countries, four people familiar with the conversations told CNBC, requesting anonymity to discuss the private company deal. Two of the people said OpenAI had been exploring opportunities for such small-scale capacity deployments in Northern Europe.
One source said he was also familiar with discussions involving Anthropic and OpenAI about bringing in manufacturing capacity at that scale in the United States.
The two companies have announced a flurry of AI infrastructure deals over the past year, with the goal of training and delivering models to end users. Deals to secure smaller compute allocations will allow businesses to deploy workloads faster amid the AI boom.
“We are building a diverse computing portfolio to meet the growing demand for AI around the world,” an OpenAI spokesperson told CNBC.
“Different workloads require different infrastructure, so we are talking with a variety of partners and evaluating opportunities based on requirements, performance, reliability, timing and cost,” they added. “We do not comment on specific commercial discussions.”
Anthropic did not comment when contacted by CNBC.
“Speed to Usable Capacity”
Both AI labs typically rent computing power from data center operators or neoclouds and seek large, long-term contracts.
Anthropic has signed a roughly $45 billion cloud deal with Nscale that will see AI Labs lease about 460 MW of computing power in a data center development in West Virginia, two people familiar with the matter told CNBC in August.
OpenAI announced in April that it had exceeded its initial commitment of 10 GW for the Stargate AI infrastructure project, and has since committed to developing an additional 3 GW in Georgia and 8 GW in Ohio.
Large-scale data center projects in the United States and other regions increasingly face pushback from local communities. The sector is also under pressure in many parts of Europe where there is a lack of available land and electricity.
Jabez Tan, head of research at Structure Research, told CNBC that smaller capacity contracts are often more attractive in terms of “speed to available capacity.”
“Securing a few megawatts at an existing power distribution site may be more practical than waiting for a much larger block in one location,” he said. “For workloads that can span separate sites, adding a collection of smaller deployments can provide significant capacity gains.”
Moving to inference
Training AI models requires large amounts of computing power to process vast amounts of data, but the day-to-day deployment of these systems, a process called inference, can be done on clusters of smaller chips.
“Training large models typically requires many chips to work closely together,” Tan says. “Many inference workloads can instead serve individual requests across multiple small clusters, allowing for more locations.”
This shift will be important as more AI computing moves from training models to production delivery. Therefore, the amount of capacity used to perform inference is expected to increase.
According to a report from real estate firm JLL, the proportion of total data center capacity used for inference workloads is expected to exceed training workloads by 2027. According to the report, by 2025, inference will account for 9% of the world’s workloads in data centers, while training will account for 14%. By 2030, it is predicted that 37% of that capacity will be used for inference, but only 13% for training.
Announced in February, Nvidia We plan to work with multiple data center stakeholders to explore small data centers designed for distributed inference.
Crusoe, the US company that built a huge data center complex in Texas used by OpenAI, is now investing in smaller data centers, the Wall Street Journal reported on Thursday. These facilities would be faster and cheaper than large-scale construction, which has been delayed across the United States, the magazine said. Mr. Crusoe did not respond to requests for comment.
Crusoe is one of several neoclouds experiencing a business boom in building AI. The company announced Thursday that it has raised a $3.9 billion funding round at a post-money valuation of $30.9 billion.
