OpenAI’s audacious push for AI supremacy through the Stargate project—once billed as a $500 billion infrastructure blitz—now faces mounting hurdles. Key leaders bolt for Meta. Deals crumble in Europe. Yet Amazon looms large, injecting $50 billion to fuel OpenAI’s compute hunger. The shift underscores a brutal truth in AI’s infrastructure arms race: ambition collides with reality.
Back in January 2025, Sam Altman stood beside SoftBank’s Masayoshi Son and Oracle’s Larry Ellison at the White House. President Trump hailed Stargate as a cornerstone of American AI dominance, promising $500 billion over four years for data centers to power next-generation models. OpenAI would handle operations; SoftBank, financing. Partners like Nvidia, Microsoft, and Arm signed on. The goal? 10 gigawatts of capacity—enough to light 7.5 million homes. (OpenAI announcement)
Fast forward to April 2026. Cracks appear everywhere. Three senior Stargate executives—Peter Hoeschele, who launched the project; Shamez Hemani, on compute strategy; and Anuj Saharan, building AI systems—defect to Meta’s new compute unit. Their exit signals OpenAI pivoting from owning data centers to renting vast compute from hyperscalers. (Bloomberg)
Europe bears the brunt. OpenAI shelves Stargate UK, a £31 billion ($39 billion) pledge under a UK-US deal. High energy costs—Europe’s worst—and regulatory thickets cited. “We see huge potential for the UK’s AI future,” an OpenAI spokesperson insisted, but they’d wait for “the right conditions.” The government spun it as part of £100 billion in private AI investment, yet critics called out “phantom investments.” (The Guardian)
Norway follows suit. Talks with Nscale for a 230-megawatt Narvik facility collapse. Microsoft snaps it up, agreeing to supply OpenAI compute. This marks the second European Stargate pullout in a week, after UK. Analysts point to OpenAI’s IPO prep: trim costs, clean books amid $1.15 trillion in long-term commitments. Abilene, Texas—Stargate’s flagship—sees expansion scrapped for better sites with advanced chips. (Network World)
But OpenAI presses on. Stargate evolves into a brand for all compute strategy, spanning Oracle, SoftBank, CoreWeave, AWS. U.S. sites in Texas, Ohio, New Mexico, Michigan advance; Abilene houses 450,000 Nvidia GB200 GPUs by mid-2026. Revenue hits $25 billion yearly, yet $600 billion infrastructure spend looms over four years. (Forbes)
Enter Amazon. Reports swirl of outreach for deeper ties. A $50 billion investment materializes—$15 billion upfront, $35 billion milestone-tied, perhaps to AGI or IPO. It caps a $38 billion AWS deal from 2025, now ballooning by $100 billion over eight years. OpenAI taps 2 gigawatts of Trainium chips. AWS becomes exclusive third-party distributor for OpenAI Frontier agents on Bedrock. “AWS will be the exclusive third-party cloud distribution provider,” per announcements. Microsoft blesses it; Azure stays primary for stateless APIs. (Yahoo Finance; OpenAI; About Amazon)
Why Amazon now? Diversification. OpenAI sheds Microsoft exclusivity post-restructure. Oracle’s $300 billion cloud pact powers Abilene. AMD, Broadcom chips enter mix to cut Nvidia reliance. But demand surges; no single provider suffices. Amazon’s Project Rainier rivals Stargate scale, touting the “world’s largest AI compute cluster.” (Time)
Challenges persist. Energy grids strain—Stargate sites pledge to fund upgrades, avoiding bill hikes. Talent wars rage; Meta poaches. Geopolitics bites: UK prices soar from wars. OpenAI’s $122 billion funding round—largest private ever—includes Amazon’s stake, SoftBank’s $30 billion, Nvidia’s $30 billion. Losses? $14 billion projected this year. (WSJ)
Stargate’s original vision fades. No longer bespoke megacenters. Instead, a web of rentals from AWS, Azure, Oracle. OpenAI compute lead Sachin Katti calls it mobilizing the “full ecosystem.” Sites progress toward 9 GW by 2029. But pullouts hand capacity to rivals—Microsoft, Google snag Norway, UK scraps.
Industry watchers see pragmatism. Hyperscalers like Amazon, with $100 billion 2025 capex, cram GPUs into existing facilities faster. OpenAI’s custom builds lag. Renting buys time for Titan chips on TSMC 3nm, due late 2026.
Amazon gains frontier access, Bedrock boost. OpenAI secures scale sans capex traps. Microsoft? Watches warily, but agreements hold. The race intensifies—Meta builds, Google expands, all chasing OpenAI’s lead.
Stargate stumbles. Partnerships endure. AI’s infrastructure battle rages on, with Amazon now all-in.
WebProNews is an iEntry Publication