Meta Plans Up to $95 Billion Stock Sale to Fund AI Data Center Buildout
Meta plans up to $95B in stock sales, starting with $35B, to fund AI data centers including a 5GW Louisiana campus.
The first phase includes about two-thirds of the $35 billion from common stock and $12.5 billion from securities Meta calls Series A depositary shares. Each depositary share corresponds to one-twentieth of a Series A mandatory convertible preferred share, which will convert into Meta common stock on Aug. 15, 2029. Until then, the depositary shares will trade under a separate ticker symbol on Nasdaq. Meta said Nvidia Corp., Advanced Micro Devices Inc. and Broadcom Corp. had expressed interest in buying up to $5 billion of the shares. The banks underwriting the transactions will have options to buy 37.5 million additional depositary shares and 3.375 billion additional common shares.
Meta intends to follow the initial deals with an at-the-market stock sale expected to raise as much as $60 billion. The company will sell shares incrementally over a relatively long period and will begin issuing them early next year.
The fundraising is intended to support a costly data center construction program headlined by four AI superclusters, each expected to provide more than 1 gigawatt of computing capacity. The largest, a Louisiana campus called Hyperion, is set to provide 5 gigawatts. Meta will commission 10 new power plants to meet Hyperion's energy requirements. Hyperion is a joint venture between Meta and investment firm Blue Owl Capital, which has committed $7 billion to the project. Meta plans to build its Texas supercluster through a similar partnership with BlackRock Inc.
The data center network powers Meta's Muse family of large language models. The newest model in the series, Muse Wonder Full, outperformed GPT-6 Astra and Claude Opus 5.5 on multiple popular benchmarks, including the DeepSWE collection of coding challenges, according to the report. Meta said its AI engineering roadmap emphasizes making its models more cost efficient by lowering their hardware footprint, enhancing the underlying inference infrastructure and making better use of off-peak computing capacity and energy.