OpenAI halts "Stargate" expansion, trillion-yuan debt-driven growth model suffers setback, Oracle faces AI infrastructure "outdated" dilemma

Zhitong Finance learned that the update speed of artificial intelligence chips has already far exceeded the construction cycle of data centers, a market reality that exposes the core risks of AI trading and the vulnerability of Oracle's (ORCL.US) debt-driven expansion model.

According to sources, OpenAI is no longer planning to expand its collaboration with Oracle in Abilene, Texas. This location is home to the "Stargate" data center project, and OpenAI made this decision because it aims to acquire clusters of newer generation Nvidia GPUs.

It is understood that the current phase of the Abilene facility is expected to use Nvidia’s Blackwell processors, while power supply is estimated to be available about a year later. The source said that at that time, OpenAI hopes to obtain even larger-scale clusters of Nvidia’s next-generation chips elsewhere.

On Sunday, Oracle posted on its X platform, stating that the related reports are "false and inaccurate," but the post only mentioned that the current project is progressing smoothly and did not comment on expansion plans.

It’s reported that Oracle had previously acquired the land, ordered hardware, and invested billions of dollars in construction and staffing, with expectations to further scale up.

For OpenAI, this decision is perfectly reasonable—it is naturally unwilling to use outdated chips. Nvidia used to release a new generation of data center processors every two years, but now CEO Jensen Huang has accelerated the pace to one generation per year, with each generation seeing a qualitative leap in performance. The Vera Rubin chip, introduced at CES this January and already in production, features inference performance five times that of Blackwell.

For companies building cutting-edge models, even the slightest performance improvement can lead to a huge difference in benchmark tests and rankings. Developers closely monitor these metrics, which directly impact user base, revenues, and valuations.

All of this points to a deeper structural issue. For infrastructure companies, selecting a site, connecting to power supplies, and completing construction all take at least 12 to 24 months. But customers are chasing the latest and best technology, keeping a close eye on Nvidia’s annual chip upgrades.

Oracle faces the added challenge of being the only major hyperscale cloud service provider mainly relying on debt to finance expansion, and its debt load has soared to $100 billion with no signs of stopping. In contrast, Google, Amazon, and Microsoft all have massive cash-generating businesses backing them.

Meanwhile, Oracle's partner Blue Owl has declined to finance another facility and plans to lay off up to 30,000 employees.

Oracle will release its third-quarter earnings on Tuesday. Investors will be watching closely to see how management addresses the contradiction between its $50 billion capital expenditure plan and negative free cash flow, as well as whether its financing channels remain sustainable.

The stock has fallen 23% so far this year, with its market cap halved since peaking last September.

Looking across the market, the risks of GPU obsolescence extend far beyond Oracle. It could create a ripple effect throughout the AI sector, as every current infrastructure deal signed may already be doomed to outdated hardware before power is even supplied.