What Meta, Oracle moves say about data center economics

Hyperscalers continue to invest astronomical amounts in data infrastructure for AI workloads, even as they concede that the massive projects may not pan out as planned.

Oracle, for instance, detailed in its latest 10-K report to the US Securities and Exchange Commission (SEC) all of the ways that these investments could be adversely impacted, potentially making the projects more expensive and extending build-out times.

And Meta, as well as building infrastructure, is purportedly planning to create a new line of business that would generate revenue from excess compute power. An infusion of cash from that venture could return some of its massive investments in data centers and chips to bolster its efforts to develop “super intelligence.”

These developments and others indicate that hyperscalers are beginning to optimize their investments for return on AI infrastructure, not just focusing on a data center land grab.

“Both companies recognize that there is huge potential in the buildout, but both have a legal obligation to share material risks that are significant when an industry is spending hundreds of billions of dollars on a capital outlay,” noted Jeremy Roberts, a senior director at Info-Tech Research Group.

He pointed to reports that the top four hyperscalers (Alphabet, Microsoft, Amazon, and Meta) are planning to invest up to $725 billion this year on AI infrastructure, including data centers, which is “mind-boggling.”

“The datacenter buildout we’re seeing now is unprecedented,” he said.

Oracle’s challenges

In its 10-K report, Oracle detailed its data center investment risks, as required by US federal law. In it, the company concedes that data center construction may cost more or take longer due to supply chain risks, local restrictions on development, or because of failure by third parties to complete on schedule, deliver required equipment and components as promised, fail to pay their bills, or opt not to renew contracts.

Oracle has recently invested heavily in AI and data center infrastructure, notably the $300 billion Stargate contract with OpenAI. As part of that deal, Oracle is building enormous data centers across the US. The company says it will continue to enter into long-term lease commitments with data center companies and chip and infrastructure suppliers.

However, it warns, “if we underestimate customer demand or our data center capacity needs, we may face shortages of available infrastructure, limiting our ability to support customer growth.”

New customers may not even consider Oracle offerings at all if the company can’t get what it needs for build-outs, the company says. This could leave it with expensive assets that may not be able to be re-leased, re-purposed, or assigned on “acceptable terms, if at all,” it said.

And while Oracle says it has continued to bolster its cloud products and services, it must continue to be able to develop new or “sufficiently differentiated” offerings and enhance those offerings in a timely manner at a competitive price to meet growing demand, it says. The company also must be able to secure data center capacity at affordable rates, and accurately plan for and manage capacity requirements.

Meta’s infrastructure plans

Meta, meanwhile, is continuing its spending spree on AI infrastructure, anonymous sources told Bloomberg. The company is purportedly developing plans for new cloud infrastructure business lines that would sell access to AI computing power and models, putting it in competition with other data center giants.

One potential scenario would have Meta selling access to models, including its new Muse Spark, hosted on its own AI infrastructure, as well as running the underlying data centers. This model is similar to AWS’ Bedrock offering. Another possibility is Meta selling access to “raw” computing capacity, as do neocloud businesses such as CoreWeave. This move is part of the company’s internal Meta Compute initiative, the sources said.

Like Oracle, Meta has been investing hundreds of billions of dollars in data centers and expensive AI chips. And, according to its latest 10-K: “We plan to continue to significantly expand the size of our infrastructure primarily through data centers, subsea and terrestrial fiber optic cable systems, and other projects.”

Also like Oracle, Meta conceded that these projects are complex and generate exposure to various compliance issues and political challenges. As such, it has “changed, suspended, and terminated” certain projects due to these factors, and will continue to do so if necessary.

The view from the enterprise

For enterprise IT buyers, more competition is probably a good thing in the long term, Roberts noted.

“Advances in data center tech that come about when companies are forced to find every efficiency to stay on top will probably be beneficial,” he said.

However, prices will likely to continue to rise as buildout continues, and anyone making purchasing decisions for components, adjacent products or “basically all hardware” will “feel that pain,” Roberts said.

Those trying to build right now will be competing with hyperscalers who can “pour essentially unlimited funds into construction,” he noted. They will likely have to pay more than they ever have before if they need Nvidia GPUs or high-bandwidth memory from SK Hynix, or even for a crew to plumb and wire a facility.

“I don’t know if this frenzy will cool off any time soon, but now is probably the worst time in recent history to embark on a project like this if you’re a smaller fish in the ocean of AI capital,” Roberts said.

Thus, for this and other reasons, most organizations likely won’t lead their own AI buildouts, he predicted. The successful ones will be those who “effectively optimize AI use, leverage cheaper models, and engage in thoughtful buildouts versus trying to go head-to-head against the richest companies in history in a spending contest.”