Oracle is seeking a significant data center presence in the US to support OpenAI training workloads.

The company has a specific focus on individual ~1GW sites with power being delivered in late 2026 (early 2027 at the latest), analysts at financial services firm TD Cowen said.

The partnership comes partially from a preference for Oracle's RDMA Superclusters, with its high throughput and low latency, and builds on the existing collaboration at the Abilene Stargate campus - developed by Crusoe and Oracle.

However, TD Cowen believes that this five-year deal is separate from Stargate, and was negotiated directly between Oracle and OpenAI.

As part of the deal, Oracle will procure data center capacity built to OpenAI's specifications, in which Oracle will charge $2.60-3.00/GPU/hour, with pricing varying by GPU type across around 2.3m GPUs.

In our profile of Stargate, in the next issue of the DCD Magazine, OpenAI said that it was looking to work with suppliers like Oracle and Crusoe for projects like Abilene, but that it was also planning on designing its own data centers.

TD Cowen analysts do not yet know if Oracle has exclusivity on servicing this 5GW capacity, or if others can compete for it. OpenAI has also signed a $12bn cloud deal with CoreWeave, which could potentially compete for this.

"Regardless of the entity servicing the demand, we believe 5GW of US demand from OpenAI training positions the industry for potential record data center leasing in 2025," TD Cowen said in its latest report.

"However, we are admittedly skeptical that there is 5GW of available US data center supply across individual sites with ~1GW scale that have power delivering in late 2026, aside from capacity held by certain historical crypto miners converting facilities to support HPC."

The company also put the cost at about ~$92 billion in GPU capex (assuming $40K/GPU) and $67 billion+ in data center capex (assuming $13.5m/MW). Given this $159bn cost, it's likely that other financial backers may be involved.

Despite the deal not being negotiated with Stargate, it is possible that its funding could be involved.

Stargate, which expects to spend $500bn in the US, is also looking to build around the world. The UAE announced the first such project, at 1GW, although the terms of the deal are still in the earliest stages.

TD Cowen believes Stargate is targeting a total of 5GW internationally, backed by investments from the states involved. Analysts mentioned European countries, Israel, and Malaysia as being under consideration.

OpenAI’s chief strategy officer is currently touring Asia Pacific, including India, Japan, South Korea, Australia, and Singapore - with the trip in part to hear potential data center investment incentives.

In Europe, OpenAI is reportedly in discussions with the UK, France, and Germany.