Major tech companies are facing a severe infrastructure crisis in their race to build the massive computing facilities required for artificial intelligence. Projects are being canceled, tens of thousands of workers are being laid off, and some firms are proposing radical solutions—including moving data centers into space—as terrestrial expansion hits social, regulatory, and logistical walls.
OpenAI's ambitious plan to establish a major UK facility has ground to a halt. Announced in September with British firm Nscale, the project at Cobalt Park in Tyneside was slated to install roughly 8,000 Nvidia computing chips by early 2026. The facility remains unopened, with OpenAI declining to provide a new timeline. This British project is part of the broader $500 billion "Stargate" program revealed by CEO Sam Altman in January 2025. In the US, OpenAI has also scrapped plans to expand a Texas site developed with Oracle, and discussions with investors like SoftBank are moving slowly.
Oracle is facing its own financial squeeze from the AI infrastructure push. On March 31, the company shocked employees by sending termination emails at 6 a.m., announcing the elimination of their roles as part of a "broader organizational change." Estimates of the job cuts range from 10,000 to 30,000 people, with the higher figure representing nearly 19% of Oracle's 162,000-strong workforce. This comes despite the company reporting a 95% jump in profits last quarter to just over $6 billion. Oracle's stock price has fallen about 55% from its peak last September, closing at $147.11 on the layoff day.
The cuts are seen as a cost-saving measure, with analysis suggesting eliminating 20,000 to 30,000 positions could save Oracle up to $10 billion. The company is grappling with doubled borrowing costs as banks pull back from financing its expansion, despite signing a $300 billion deal with OpenAI last year to provide computing infrastructure.
The problem is industry-wide. Research from Sightline Climate indicates that between 30% and 50% of large AI facilities planned for the US this year will face delays or be scrapped entirely. Of 140 building projects representing at least 16 gigawatts of capacity scheduled to open by end-2026, only around 5 gigawatts are currently under construction. Power supply is the primary bottleneck. Global electricity use by AI data centers hit roughly 415 terawatt-hours in 2024, with the International Energy Agency projecting it could top 1,000 terawatt-hours by 2026. In some regions, like Virginia, these facilities already consume 26% of all electricity.
In response, a new frontier is being proposed: orbital data centers. SpaceX filed paperwork on January 30 to launch up to one million satellites to create a constellation with "unprecedented computing capacity." Seven weeks later, Blue Origin filed for 51,600 satellites under its Project Sunrise. Startups are also entering the fray; Starcloud raised $170 million in March at a $1.1 billion valuation and filed to launch 88,000 satellites.
This vision is central to SpaceX's reported pursuit of a staggering $1.75 trillion valuation and a confidential IPO filing for a $75 billion raise. The company's vertical integration—acting as both the launch provider and the operator—creates a powerful, self-reinforcing business model. However, scientists and engineers highlight profound technical hurdles: managing waste heat in a vacuum requires about 1,200 square meters of radiator per megawatt, radiation damages chips, communication delays hinder AI model training, and costs are exorbitant. IEEE Spectrum estimates a one-gigawatt space facility would cost over $50 billion, roughly triple a similar ground facility. Even Sam Altman has called the idea "ridiculous" for this decade.
The push for orbital infrastructure intersects with an uncertain AI compute demand curve. Some AI labs are reassessing projected needs, potentially undermining the economic urgency for massively expensive orbital alternatives. Analysts suggest space-based compute would likely serve only as a specialized supplement for specific workloads, not a replacement for terrestrial clouds.
