Meta's $182.9B AI stack has slack — «Meta Compute» is coming for AWS / Azure / GCP
TL;DR
Bloomberg reported on July 1 that Meta is forming «Meta Compute» to sell excess AI training capacity to outside customers, taking direct aim at AWS, Azure and Google Cloud.
Meta is about to flip from cloud buyer to cloud seller. Bloomberg reported on July 1 that Meta is spinning up a new business line codenamed «Meta Compute» to sell its unused internal AI training capacity to outside customers, taking direct aim at AWS, Azure and Google Cloud. Meta stock jumped in pre-market trading on the report.
Scale is the pitch. As of Q1 2026 Meta had committed $182.9 billion in AI infrastructure spend, with 2026 alone landing between $125 and $145 billion in capex. Two giant clusters are going up in Louisiana and Ohio; the Ohio site — «roughly the size of Manhattan» — comes online this year. In May, on the earnings call, Zuckerberg said a cloud business was «definitely on the table». Two months later, it landed.
Meta Compute has three names on top: infrastructure chief Santosh Janardhan, Meta Superintelligence Labs head Daniel Gross, and newly promoted president Dina Powell McCormick. Two models run in parallel internally: copy CoreWeave and sell raw GPU capacity by the slice, or copy AWS Bedrock and put Meta's newly released closed-weight model Muse Spark on the shelf behind an API meter.
xAI ran the same play weeks earlier — leasing capacity on the Colossus 1 cluster to Anthropic, Google and Reflection AI. The read underneath: in this AI cycle, who owns the racks collects more rent than who tops the leaderboard. Meta's Llama family has never disclosed material revenue, and Muse Spark is not booking numbers yet — selling power and GPUs directly is faster cash than waiting on the model layer.
Meta declined to comment.
If it works, $182.9 billion of AI stack gets sold under the umbrella of AWS's ~30% cloud margins, ad-revenue concentration breaks, and the story reprices from «social platform» to «fourth cloud». If it doesn't, AWS / Azure / GCP push prices to cost to lock the new entrant out, and half of $182.9 billion sits on the balance sheet depreciating.
via Bloomberg / TechCrunch / Reuters (via Yahoo) / Seeking Alpha
Scale is the pitch. As of Q1 2026 Meta had committed $182.9 billion in AI infrastructure spend, with 2026 alone landing between $125 and $145 billion in capex. Two giant clusters are going up in Louisiana and Ohio; the Ohio site — «roughly the size of Manhattan» — comes online this year. In May, on the earnings call, Zuckerberg said a cloud business was «definitely on the table». Two months later, it landed.
Meta Compute has three names on top: infrastructure chief Santosh Janardhan, Meta Superintelligence Labs head Daniel Gross, and newly promoted president Dina Powell McCormick. Two models run in parallel internally: copy CoreWeave and sell raw GPU capacity by the slice, or copy AWS Bedrock and put Meta's newly released closed-weight model Muse Spark on the shelf behind an API meter.
xAI ran the same play weeks earlier — leasing capacity on the Colossus 1 cluster to Anthropic, Google and Reflection AI. The read underneath: in this AI cycle, who owns the racks collects more rent than who tops the leaderboard. Meta's Llama family has never disclosed material revenue, and Muse Spark is not booking numbers yet — selling power and GPUs directly is faster cash than waiting on the model layer.
Meta declined to comment.
If it works, $182.9 billion of AI stack gets sold under the umbrella of AWS's ~30% cloud margins, ad-revenue concentration breaks, and the story reprices from «social platform» to «fourth cloud». If it doesn't, AWS / Azure / GCP push prices to cost to lock the new entrant out, and half of $182.9 billion sits on the balance sheet depreciating.
via Bloomberg / TechCrunch / Reuters (via Yahoo) / Seeking Alpha
