OpenAI procurement doubts grow
- DigiTimes says OpenAI is rethinking parts of its AI-server buying plan, after months of Stargate expansion talk pulled Quanta, Wistron, and Foxconn deeper in. - The tension is between giant public ambitions and shifting near-term needs: Stargate promised $500 billion, but one Texas expansion was reportedly scrapped in March. - That matters because suppliers staffed and built for a wave that may arrive later, smaller, or through different partners.
AI infrastructure is the physical side of the generative-AI boom — the racks, power gear, cooling loops, buildings, and contract manufacturers behind the models. That buildout has been running on one huge assumption: OpenAI would keep pulling more and more hardware into Stargate at breakneck speed. Now that assumption looks shakier. A DigiTimes report on May 1 says OpenAI is reassessing parts of its server procurement plan, and that lands awkwardly because suppliers already expanded around the idea of relentless demand. (digitimes.com) ### What exactly is being questioned? The immediate issue is not whether OpenAI still wants more compute. It clearly does. The issue is whether the pace, mix, and routing of purchases are changing enough to disrupt the supply chain that geared up for them. DigiTimes points to uncertainty around orders tied to OpenAI’s data-center push and names Quanta, Wistron, and Foxconn as the hardware groups most exposed if volumes slip or timing moves. (digitimes.com) ### Why were suppliers leaning in so hard? Because OpenAI and its partners set expectations absurdly high. On January 21, 2025, OpenAI launched Stargate with SoftBank and Oracle and said the venture intended to invest $500 billion in U.S. AI infrastructure over four years, with $100 billion to begin immediately. Later, OpenAI said Stargate and related Oracl(digitimes.com)s for a living, that sounds like a once-in-a-generation order book. (openai.com) ### So what changed in practice? The cleanest sign is Texas. In March 2026, Bloomberg reported that Oracle and OpenAI had scrapped plans to expand a flagship Abilene data center after financing talks dragged and OpenAI’s needs changed. That does not mean Abilene is dead — far from it. The first campus is live, Oracle started delivering Nvidia GB200 racks in June 2025, and early training and infe(openai.com)m ambition” and “every planned expansion happens on schedule” are not the same thing. (bloomberg.com) ### Why does Abilene matter so much? Because Abilene became the proof point for the whole story. Bloomberg said the site was expected to house 64,000 Nvidia GB200 chips by the end of 2026, with an initial 16,000 in the first phase. Crusoe and partners also pushed the campus to 1.2 gigawatts, backed by a $15 billion joint venture. That is the kind of anchor project suppliers use to justify factory lines, hiring, and inventory commitments. (bloomberg.com) ### Does this mean the AI buildout was fake? No — that is the wrong read. Buildings are going up, chips are shipping, and OpenAI is still publicly expanding Stargate. The more realistic read is that hyperscale AI infrastructure is lumpy. Financing changes. Site plans move. A customer may still spend enormous sums overall while changing which partner supplies what, and when. That is enough to hurt manufacturers who planned for a smooth ramp. (openai.com) ### Why are Quanta, Wistron, and Foxconn the ones to watch? Because they sit in the awkward middle. They are not the end customer, and they are not the chip designer. They assemble the expensive boxes and broader systems that turn Nvidia silicon into usable AI clusters. If OpenAI delays, reroutes, or resizes orders, these companies can get caught with underused capacity, weaker ut(openai.com)Times’ point is basically that the demand story may still be real, but the timing risk just got much more visible. (digitimes.com) ### What is the bottom line? The story is not “OpenAI stops building.” It is “the market priced in a straight-line surge, but the real buildout looks messier.” For AI-server suppliers, that distinction is everything. (digitimes.com)
