Anthropic and OpenAI are actively exploring partnerships for smaller data center projects, according to sources familiar with the matter.

As market demand surges, both companies are racing to deploy AI compute capacity, having signed numerous AI infrastructure agreements over the past year. One analyst noted that smaller compute projects are attractive primarily because they offer faster access to usable processing power.

With the competition for AI compute infrastructure intensifying, multiple sources reveal that Anthropic and OpenAI are scouting for smaller AI data center collaboration opportunities. Over the last 12 months, these two AI labs have inked large-scale data center deals ranging from hundreds of megawatts to gigawatts. However, insiders now indicate they are simultaneously pursuing much smaller compute resources, with individual projects scaling only 20-30 megawatts.

Four individuals with direct knowledge of the negotiations, who requested anonymity, stated that Anthropic has already approached potential partners in the UK and Nordic region for compute solutions at this scale. Two additional sources said OpenAI has also explored similarly sized small compute deployments in the Nordics.

Another insider mentioned that both companies have discussed this tier of compute deployment in the US market as well.

To support model training and deliver services to end users, both firms have closed multiple AI infrastructure transactions over the past year. Amid the AI industry boom, securing smaller batches of compute capacity can help companies get their operations up and running faster.

An OpenAI spokesperson stated, "We are building a diverse portfolio of compute resources to meet the growing global demand for AI." The spokesperson added, "Different workloads require different infrastructure. We engage with numerous partners and evaluate opportunities based on our needs, performance, reliability, delivery timelines, and cost. We do not comment on specific commercial negotiations."

Faster Access to Usable Compute

Both AI labs typically lease compute from data center operators and emerging AI cloud providers, known as neoclouds, previously focusing on large-scale, long-term contracts. In August, two sources disclosed that Anthropic and Nscale signed a substantial agreement to lease approximately 460 megawatts of compute at a data center under construction in West Virginia.

OpenAI, meanwhile, stated that its "Stargate" AI infrastructure project, launched in April, has already surpassed its initial commitment of 10 gigawatts. The company has since pledged an additional combined 3 gigawatts across Georgia and Ohio.

Megaprojects in the US and abroad are facing pushback from local communities, while much of Europe also grapples with industry pressure due to limited land and tight power supply.

Jabez Tan, Head of Research at Structure Research, told that the appeal of smaller compute deals lies in "faster access to usable compute." "Securing megawatts at existing sites that are already energized is far more practical than waiting for a massive compute block to come online at one location. For workloads that operate across multiple sites, stacking several smaller projects can still yield substantial total compute capacity."

Shifting Toward Inference Workloads

Training AI models requires vast arrays of chips to process enormous data volumes. However, the daily serving of models to end users, known as inference, can rely on smaller clusters of chips.

Tan explained, "Training large models often demands many chips working in tight coordination. In contrast, many inference tasks can distribute requests across multiple small clusters, unlocking more location possibilities."

This shift, as more AI compute transitions from model training to production environment inference, is significant. The scale of compute required for inference is expected to continue rising.

A report from real estate services firm JLL forecasts that by 2027, the share of global data center compute used for inference workloads will surpass that used for training. The report shows that in 2025, inference accounts for 9% of global data center workloads, with training at 14%. By 2030, inference is projected to consume 37% of compute, while training will account for only 13%.

In February, it was announced that Nvidia, alongside several data center-related companies, is jointly researching small data center solutions for distributed inference.

Crusoe, a US company that previously built a large data center for OpenAI in Texas, is now investing in constructing smaller facilities. Reports indicate that large data centers across the US are experiencing widespread construction delays, while these smaller installations are faster to build and more cost-effective.

Crusoe did not respond to requests for comment.

Crusoe is among the neoclouds experiencing rapid business growth amid the AI buildout. The company announced Thursday that it had closed a $3.9 billion funding round, achieving a post-investment valuation of $30.9 billion.