Meta's Computing Power Sales Ignite Wall Street Debate: Will the AI Infrastructure Bubble Burst?

According to Bloomberg, Meta is establishing a new business unit called "Meta Compute" with plans to sell its excess computing capacity to external customers. Potential options include offering access to AI models and directly leasing out underlying computing power.

This pivot is not a sudden whim. As early as May this year, Meta CEO Mark Zuckerberg hinted during a shareholder call that selling excess computing power or API services was "definitely on the table."

Public disclosures reveal a clear generational divergence in Meta's computing power allocation. On one hand, the company continues to aggressively expand high-end computing capacity, having recently signed a deal with Crusoe in mid-to-late June to secure 1.6 GW of AI computing capacity, while raising its full-year capital expenditure guidance to $125 billion to $145 billion in the first quarter. On the other hand, Meta plans to lease out its previous-generation GPU computing power.

This seemingly contradictory move is actually a fine-grained management of computing resources—reserving the latest-generation computing clusters for frontier model training, while monetizing idle capacity from previous-generation or non-core workloads to improve asset turnover.

Morgan Stanley ( MS )'s estimates show that if Meta leases 250 MW of computing capacity, priced at $40 per watt, it could add approximately $2.97 to its 2028 EPS, representing about an 8% upside.

For Meta shareholders, this move effectively establishes a short-term monetization channel for its massive AI investments, easing market concerns that 2027 EPS could flatline or even contract.

Meta's move has sparked a fierce market debate over the supply and demand dynamics of AI computing power.

Bears argue that this marks a shift in corporate objectives from "scrambling for chips to secure supply" to "monetizing existing assets," which will lead to a contraction in expected procurement demand for upstream chips and memory.

D.A. Davidson analyst Gil Luria even suggested that Meta's move implies the company is "abandoning frontier AI R&D" in favor of monetizing computing power for short-term gains.

He pointed out that since the establishment of Meta's super-intelligence lab last year, despite the release of the Muse Spark model, it has lagged behind Anthropic and OpenAI in terms of technological iteration speed.

Luria noted in his report that if Meta indeed cuts AI R&D spending to focus on monetizing computing power, the company's revenue and cash flow could indeed see a significant boost, but this would also mean choosing to take a step back in core technological competition.

On the other hand, bulls emphasize that this is merely a dynamic reallocation of computing resources by Meta and does not represent an industry-wide surplus of computing power.

Looking at the development of the global AI industry, the demand for computing power is still growing rapidly. Google Cloud's backlog has reached nearly $460 billion, and Microsoft Azure's commercial remaining performance obligations (RPO) grew 99% year-over-year to $627 billion. Leading model company Anthropic not only signed a 10-year, $100 billion partnership agreement with AWS but also leased 300MW of computing capacity from SpaceX's Colossus 1, paying $1.25 billion monthly through 2029. These figures indicate that high-end computing power remains in short supply, and Meta is only leasing out idle, previous-generation capacity, which represents a structural mismatch rather than an all-out surplus.

An analysis by Bernstein pointed out that Meta currently has about 20GW of global data center capacity, with plans to add approximately 14GW over the next few years. However, not all of this capacity is available for AI computing rental, as it includes resources of different generations and for different purposes.

More importantly, recent rumors suggest that Google has restricted Meta's computing usage due to its own capacity constraints, which indirectly indicates that Meta itself still faces a shortage of computing power.

Jefferies analyst Brent Thill believes that market worries about Meta's "overbuilding" are completely "putting the cart before the horse." He emphasized that demand for computing power currently continues to outstrip supply, and Meta's new cloud business will increase infrastructure utilization, improve return on invested capital (ROIC), and boost cash flow, ultimately "funding more capital expenditure, not less."

According to Jefferies' research, Meta's current internal infrastructure utilization rate is around 65%, with the remaining 35% of idle capacity providing the company with significant monetization potential.

Mark Zuckerberg stated last year that Meta is confident it can lease out this idle capacity at "a premium above the acquisition cost." Thill believes that Meta is "not exiting the AI race, but rather turning its early aggressive capacity layout into options that generate strategic value."

Meta's entry into the compute sales market has had varying impacts on different types of companies. For CoreWeave ( CRWV ), Nebius ( NBIS) and other emerging cloud service providers, this is undoubtedly a major headwind—Meta is not only their key customer but will also become a direct competitor in the future.

Data shows that Meta's current contract value with CoreWeave stands at $35.2 billion, accounting for more than one-third of its order backlog. Over the long term, the dynamics of a customer turning into a competitor will weaken the pricing power of these emerging providers.

For the chip hardware industry, the short-term impact stems more from sentiment than fundamentals. JPMorgan's trading desk pointed out that semiconductor and memory positioning is close to historic highs, and the news from Meta was merely a trigger for profit-taking. Most analysts believe that demand for AI compute, storage, and communication is still poised to benefit from the rising penetration of large language models, and short-term volatility does not necessarily signal the end of the upcycle.

For Meta itself, selling compute is merely a temporary strategy rather than a long-term core business. Morgan Stanley emphasized that the core of Meta's valuation still lies in its primary businesses, such as advertising revenue, Reels monetization, and AI engagement growth; selling compute only serves as a short-term supplement to EPS and cannot automatically lift valuation multiples. If Meta truly intends to build a comprehensive AI cloud service, it will need to invest more resources in model capabilities, software stacks, and enterprise customer service.

This content was translated using AI and reviewed for clarity. It is for informational purposes only.