Microsoft leases Abilene data center abandoned by Oracle and OpenAI

Microsoft has agreed to lease a roughly 700-megawatt data center site in Abilene, Texas that had been pursued by Oracle and OpenAI before they walked away, according to Bloomberg and cited by The Decoder. The facility sits next to the planned Stargate campus, one of the flagship projects in the US race to build AI infrastructure. Reuters reported that Microsoft, Oracle, OpenAI and the developer Crusoe declined to comment.

The handover is a reminder that the “AI race” is being decided as much by real estate and grid connections as by model benchmarks. A 700MW site is not a software feature; it is land, transformers, cooling loops, diesel backup, water rights, fiber routes and a queue position with the local utility. When a single tenant steps back, the asset does not vanish—it becomes a bargaining chip for whoever can absorb it, reconfigure it, and sign a long-enough contract.

That dynamic favours the hyperscalers with diversified cashflow and the ability to shift workloads across regions. Microsoft can treat capacity as an option: lease now, fill later, and arbitrage price differences as demand and power availability change. In a recent podcast, CEO Satya Nadella said he expects an oversupply of compute and falling prices by 2027 or 2028, effectively describing a cycle where today’s scarcity premium becomes tomorrow’s discounted rental market.

For developers and financiers, the same cycle looks different. Data center projects are increasingly being underwritten on the assumption that AI demand is both large and urgent—an assumption that becomes fragile when model roadmaps change, training runs move, or a major customer decides it needs a different configuration. Bloomberg previously reported that Oracle and OpenAI had stepped back from Abilene expansion talks after financing discussions stalled and OpenAI’s needs shifted; Oracle disputed the characterization at the time. Either way, the result is an expensive, partially committed asset searching for a tenant that can move faster than the debt clock.

The second-order effect is consolidation. If smaller developers and single-purpose projects get stranded, the buyers are typically the firms that already control cloud distribution, enterprise relationships, and long-term procurement. The AI buildout still expands, but ownership concentrates: the same companies that sell the software subscriptions end up controlling the physical bottlenecks that determine who can scale.

Abilene’s 700MW did not become unnecessary when Oracle and OpenAI stepped away. It simply changed hands to the one company that can afford to wait for the next pricing cycle.