Meta Platforms has aspirations to be one of the big AI model builders and to push the state of the art to “superintelligence,” under the impression that it can sell access to personal AI agents to its more than 1 billion users worldwide. It is a reasonable business model – and a lot more sensible than the metaverse that was the obsession of company founder Mark Zuckerberg a few years and a few tens of billions of dollars ago.

While the social network and AI model maker has a fondness for its own server, storage, switch, and datacenter designs, even after the launch of the Open Compute Project and its open strategy firmly charted, the company has not been shy about renting datacenter capacity when and where it needed to. And in the GenAI boom, Meta Platforms has not shied away from renting compute capacity from others.

Back in September 2025, Meta Platforms inked a supplier deal with king of the neoclouds, CoreWeave, that saw the former buy $14.2 billion in AI processing capacity through December 2031. This was a big deal for CoreWeave, which had just gone public and which needed to show that it has customers other than Microsoft renting its GPU hours. With the deal announced today in an 8-K filing with the US Securities and Exchange Commission, CoreWeave said that Meta Platforms has signed up for an additional $21 billion in AI processing capacity running out through December 2032.

No one knows the ramp of that capacity usage, and it is no doubt nailed down pretty hard because the one thing CoreWeave can’t have is a GPU burning capital and electricity and a customer not paying for the privilege of using it. It is not safe to assume that all $21 billion of spending happens in 2032, the incremental year, but for all we know that is exactly what will happen.

This deal will bring CoreWeave’s revenue backlog up to $87.8 billion, and Meta Platforms will represent 40.1 percent of that backlog. Back in March 2025, OpenAI inked a five year deal with CoreWeave for $11.9 billion in capacity, and another $10.5 billion in capacity was added last year, bringing the total to $22.4 billion in capacity. Assuming that none of that OpenAI capacity has been activated as yet, that would represent 25.5 percent of that extended backlog. So these two customers – Meta Platforms and OpenAI – represent nearly two-thirds of the guaranteed revenue on the books. In 2024, Microsoft represented 62 percent of the company’s $1.92 billion in sales, and CoreWeave’s top three customers accounted for 77 percent of revenues.

CoreWeave only has two dozen named customers as far as we can tell, and no doubt will have more, especially as its big clients roll off old iron and onto new stuff. In fact, that is the whole neocloud business model. Keep the GPUs humming for seven or eight years and sell the old capacity to enterprises, governments, startups, and others as the tech titans pay to rent the new GPU systems as they come hot out of the factories.

In all of 2025, CoreWeave had $5.13 billion in sales, up 2.7X from 2024, and it had an operating loss of $46 million, but due to stock compensation and other factors, it posted a net loss of $1.17 billion.  The losses are at least growing slower than the revenue, unlike some fast-growing startups.

The company has just shy of $4 billion in cash and marketable securities, which is partly due to going public last March. Thanks to hefty outside investment, however, CoreWeave was able to spend $14.9 billion on capital equipment expenses, building out its datacenter fleet to 43 centers. It ended 2025 with 850 megawatts of active power – we think CoreWeave has maybe 600,000 GPUs, mostly Nvidia “Hopper” H100s and H200s with a rising number of “Blackwell” B200s and B300s. The company has 3.5 megawatts of total power under contract, which is a measure of how big it can build itself without having to scrounge for more juice. By the way, assuming there has not been much in the way of depreciation on those GPUs, they are worth somewhere between $15 billion and $20 billion. That's a fair amount of collatoral.

It is going to take more money than CoreWeave wants to tap from its balance sheet to start building out that 2.25 gigawatts of unused contracted power – at the $50 billion per gigawatts figure that Nvidia co-founder and chief executive officer Jensen Huang cites, and assuming that CoreWeave continues to be an Nvidia-only shop (a fair assumption), we are talking about CoreWeave needing on the order of $113 billion to build out that unused capacity, which might rent for 4X to 5X that amount on a one of the big cloud builders over a bunch of years and maybe half that on a neocloud, which exists to be cheaper and focused on AI.

To help cover its datacenter expansion costs, CoreWeave said this morning that it was doing a private offering of $1.25 billion in senior notes that would come due in 2031, and by the afternoon, that amount was raised to $1.75 billion in notes. The company also said that it was going to offer $3 billion in senior notes that could be convertible to its stock. These notes would come due in 2031 as well.

Back in February, CoreWeave said that it would spend somewhere between $6 billion and $7 billion on capital expenses, and for the full 2026 year expected to invest somewhere between $30 billion and $35 billion. That’s a tall order for a company with just under $4 billion in the bank, but CoreWeave has some very rich friends – think the green eye of Horus – who want there to be alternatives to Amazon Web Services, Microsoft Azure, and Google Cloud, all of whom buy lots of GPUs from Nvidia but who also spend a lot of time designing and deploying their own custom AI accelerators.