Could the departure of one partner unsettle the footing of a $10 billion AI mega-project? The latest market downturn at Oracle implies that it can and has. The company’s shares slumped by nearly 5% over news that Blue Owl Capital, its lead data center partner in the United States, refused to invest in its proposed 1 gigawatt campus in Saline Township, Michigan. However, the company has rejected this claim, saying that Blue Owl Capital was never chosen, and that progress is being made with a lead equity partner.
The Michigan project is one of the most ambitious ones that Oracle is pursuing, aimed at providing compute power with high-performance computing for workloads related to OpenAI under the Stargate AI infrastructure program. With 1 GW of capacity, the campus will make it one of the largest AI-centric facilities in the U.S., competing with hyperscalers’ infrastructure. As per the plans for the project, the facility, called “The Barn,” will span 250 acres of land and feature three single-story buildings with 550,000 square feet of floor space, with the use of a closed-loop cooling system to cut down on water usage.
Contemporary AI data centers of this scale are designed for GPUs-clusters to train cutting models. Rack power densities are already in excess of 100 kW, with a projected increase to 1,000 kW in 2029. As a consequence, there is an increasing need for AI-ready electrical infrastructure, including medium-voltage distribution, UPS with a high-power handling capacity, and DC feeds to server rooms. Cooling systems are transitioning from liquid-to-air to liquid-to-liquid designs, with liquid-to-liquid requiring half of the space of a liquid-to-air system, resulting in more than 50% reduction in non-IT energy costs.
Blue Owl’s announced retreat also points out that there are challenges in financial structuring. In previous investments involving Oracle, Blue Owl has constructed and maintained ownership and built the facility through long-term leases with Oracle, thus keeping it off the balance sheet of the former. Also, according to sources, Michigan faced unfavorable terms of debts and repayment, along with possible delays in politics. There are also disputes over rezoning and lawsuits, although an agreement has been reached.
The capital commitments made by Oracle are simply mind-boggling. The lease liability for data center and cloud infrastructure is now roughly $248 billion over a period of 15-19 years, up by a staggering 148% over the levels in August last year. In September, Oracle raised $18 billion in new debt, and by late November, total debt, including operating leases, topped $124 billion. It is estimated that the company may have to borrow as much as $100 billion to underpin its AI infrastructure developments. The free cash flow has become negative, with estimates pegging the figure to be roughly $10 billion in the latest quarter.
The broader backdrop is a hyperscaler race for AI-optimized compute. World spending on data center infrastructure reached $290 billion in 2024, with servers accounting for a remarkable 61% of overall information and communications technology spending. Hyperscalers are at the forefront of innovation in rack-scale GPU architectures, high-density power management, and emerging storage solutions such as Heat-Assisted Magnetic Recording (HAMR) that can reduce power usage by as much as 39% per terabyte. Oracle’s “Supercluster” fabric solution with a potential maximum deployment of 131,072 NVIDIA GPUs puts it on track for petaflop-to-exaflop capabilities for OpenAI and other clients.
Yet the risk of execution is very real. Supply chain issues may extend the delivery schedules for turbines, for transformer components, and for the specialty cooling equipment into multiple years. Jonathan Koomey, a researcher focusing on the topic of data centers, added, “The world of bits moves fast. The world of atoms doesn’t. And data centers are where those two worlds collide.”
And for an equity investor, the collision is further complicated by the fact that Oracle’s entry into the market for hyperscale AI infrastructure has come relatively late and is dependent on a single ‘anchor’ customer. The bond markets are already warning. The credit-default spreads for Oracle are at their highest levels since 2009, and some of their newer debt is trading at junk-bond levels. While the company is committed to preserving investment grade status, debt investors, prioritizing security over high rewards, are scratching their heads. The experience in Michigan, whether a misunderstanding or a glitch, has underlined that in this era of AI infrastructure investment, confidence in capital is just as important as competence.
