Artificial intelligence chips are getting upgraded more quickly than data centers can be built, a market reality that exposes a key risk to the AI trade and Oracle’s debt-fueled expansion.
OpenAI is no longer planning to expand its partnership with Oracle in Abilene, Texas, home to the Stargate data center, because it wants clusters with newer generations of Nvidia graphics processing units, according to a person familiar with the matter.
The current Abilene site is expected to use Nvidia’s Blackwell processors, and the power isn’t projected to come online for a year. By then, OpenAI is hoping to have expanded access to Nvidia’s next-generation chips in bigger clusters elsewhere, said the person, who asked not to be named due to confidentiality.
Bloomberg was first to report on the companies ending their plans for expansion in Abilene. In a post on X on Sunday, Oracle called news reports about the activity, “false and incorrect,” but the post only said existing projects are on track and didn’t address expansion plans.
Oracle secured the site, ordered the hardware, and spent billions of dollars on construction and staff, with the expectation of going bigger.
An Oracle spokesperson declined to comment.
It’s a logical decision for OpenAI, which doesn’t want older chips. Nvidia used to release a new generation of data center processors every two years. Now, CEO Jensen Huang has the company shipping one every year, and each generation offers a leap in capability. Vera Rubin, unveiled at CES in January and already in production, delivers five times the inference performance of Blackwell.
For the companies building frontier models, the smallest improvement in performance could equate to huge gaps in model benchmarks and rankings, which are closely followed by developers and translate directly to usage, revenue, and valuation.
That all points to a bigger problem at play. For infrastructure companies, securing a site, connecting power and standing up a facility takes 12 to 24 months at minimum. But customers want the latest and greatest, and they’re tracking the yearly chip upgrades.
Oracle’s added challenge is that it’s the only hyperscaler funding its buildout primarily with debt, to the tune of $100 billion and counting. Google, Amazon and Microsoft, by contrast, are leaning on their enormous cash-generating businesses.
Meanwhile, Oracle partner Blue Owl is declining to fund an additional facility, and plans to cut up to 30,000 jobs.
Oracle reports fiscal third-quarter results on Tuesday, and investors will be paying close to how the company addresses a $50 billion capital expenditure plan with negative free cash flow, and whether the financing pipeline can hold up.
The stock is down 23% so far this year and has lost over half its value since peaking in September.
Beyond Oracle, GPU depreciation is a risk for the broader market and could have ramifications across the AI landscape. Every infrastructure deal signed today may result in a commitment to outdated hardware before the power is even connected.
WATCH: Jefferies’ Brent Thill talks to CNBC ahead of Oracle earnings
