Meta is strategizing to monetize its massive AI computing power assets—a move that both represents the embryonic stage of a new business line and serves as a strategic signal to address investor concerns over the returns on hefty capital expenditures.
According to reports, Meta is formulating plans to launch a cloud infrastructure business that would sell access to AI computing power and model services to external customers. Following the announcement, Meta's stock surged approximately 10% in a single day, far outpacing the S&P 500's roughly 0.25% gain over the same period, reflecting a positive market response to this potential new business line.
Analysts Justin Post and Nitin Bansal of BofA Securities stated in a research report released on July 1 that the advancement of the cloud business would help highlight the underlying value of Meta's computing assets and model development, thereby alleviating, to some extent, investor doubts about the company's continued heavy investment in AI infrastructure with no visible returns in sight. BofA maintained a Buy rating on Meta with a price target of US$835.
As reported by Bloomberg, citing insiders, Meta's cloud business plan currently has two directions: first, offering AI model hosting services that allow developers to access various models running on Meta's existing AI infrastructure, including its Muse Spark series models, with charges based on usage—a model similar to Amazon AWS's Bedrock product; second, directly selling raw computing power, positioning itself closer to emerging cloud service providers like CoreWeave.
These plans fall under Meta's internal strategic initiative known as "Meta Compute," which focuses on the construction and operational management of AI infrastructure. Meta's CEO has previously publicly hinted at commercial opportunities in the enterprise market and stated that the company expects to be able to sell computing power at prices above construction costs.
BofA noted in its report that from a broader perspective, if Meta's capital expenditure scale in 2026 can support the construction of up to 3GW of computing capacity (estimated at roughly US$40–45 billion per GW), establishing a cloud business platform in the near term would grant the company greater strategic flexibility—in the event of surplus computing power, it could be leased out at an estimated US$10–15 billion per GW annually, providing a positive support to the company.
Despite the enthusiastic market reaction, BofA also candidly pointed out potential skepticism. Meta's progress in developing its own chips appears to lag behind established hyperscalers such as Amazon, Microsoft, and Google. At the same time, the company is still actively procuring computing power through third-party agreements—including a recent 1.6 GW procurement deal with Crusoe.
This situation raises questions about Meta's strategic logic: Can a company that still needs to buy computing power from external sources build a credible resale business for computing power? And how will it position itself competitively in the hyperscale cloud market?
BofA believes that whether Meta can gain stronger market recognition in this arena will depend, to some extent, on the frontier capabilities of its large language models (LLMs): the higher the model performance, the stronger the external demand for Meta's computing power, and the more solid the commercial rationale for the cloud business will become.
Beyond Meta's cloud plans, there are also noteworthy signals on the cost side of AI computing. According to The Information, OpenAI has reportedly discovered a system-level optimization that reduces inference costs for certain models by about half. The optimization is achieved through more efficient utilization of existing server infrastructure, without the need for additional hardware or new model architectures.
According to the report, OpenAI has already applied this optimization to unauthenticated ChatGPT traffic, with such traffic requiring only a few hundred Nvidia GPUs to support operations. It remains unclear what the specific mechanism of this method is, nor whether it can be extended to logged-in users, API workloads, or computationally intensive inference products.
BofA believes that improvements in computing cost efficiency are broadly positive for large internet companies: if this technology can be generalized across the industry, it would expand the effective output of existing computing power without increasing hardware investments, reduce the urgency for additional capital expenditures, and improve the unit economics of AI businesses. As Agentic application scenarios drive a significant surge in token consumption, the strategic value of computing optimization will become increasingly prominent.
However, for hyperscale cloud service providers, declining inference costs also carry some risk of pricing pressure; but at the same time, better gross margin structures and a broader addressable market are expected to drive continued growth in demand for AI workloads, making the overall outlook still positive.
Related News:
Deepline | The hidden eye in router: When Wi-Fi knows who you are
Deepline | The 'are you sure' benchmark: Why AI can't stop people-pleasing
Comment