Meta (META.US) wants to turn AI capital expenditures into "computing power balance sheets"! From the AI arms race to entering the cloud computing world, Meta wants to take full advantage of the benefits of the AI reasoning era.

Facebook and Instagram's parent company, Meta Platforms Inc. (META.US), is developing a strategic plan for a groundbreaking cloud computing infrastructure business, which will sell access rights to Meta's large-scale deployment of AI computational power infrastructure and exclusive AI model resources, creating a new dimension of competition with industry leaders such as Amazon.com, Inc. cloud services (Amazon.com, Inc.AWS), Microsoft Corporation Azure cloud platform, and Alphabet Inc. Class C cloud platform (Google Cloud). According to sources cited by the media, Meta has been rapidly locking in the expensive process of building artificial intelligence data centers and massive underlying AI computational power infrastructure resources, to support its artificial intelligence ambitions. At the same time, the company is setting up an important new business, creating revenue from selling surplus AI computational power resources to external customers. The sources requested anonymity as the details have not been publicly disclosed. Meta's foray into cloud computing is essentially focusing on transforming the substantial AI capital expenditure that has been repeatedly questioned by the market into a "AI computational power asset balance sheet" with external revenue elasticity, at a time when global demand for AI computational power is rapidly expanding. Meta is building cloud businesses to sell surplus AI computational power, and the recent reports of Alphabet Inc. Class C restricting Meta's use of the Gemini model are also proving that the demand for AI computational power is so strong that even large tech companies internally face shortages in cloud-based AI computational power. Google Cloud's revenue reached 20 billion dollars in the first quarter, but computational constraints still limit higher growth and increase backlog orders. This also means that Meta's strategy is not simply "crossing boundaries into the cloud," but striving to lead in the development of AI superintelligence, ad recommendations, video generation, AI assistants, and model training/cloud-based large-scale AI inference clusters with GPU/ASIC/TPU resources, network infrastructure, HBM/DRAM/NAND storage components, power, and data center capacity; once surplus capacity is reached, this capacity is packaged into model APIs or raw computational power leases, comparable to AWS Bedrock or CoreWeave-like new cloud leaders, providing downward protection for "excessive AI computational infrastructure construction" and providing a pathway for AI capital expenditure recovery for investors. At a deeper level, Meta is betting on long-term AI computational power scarcity as AI workloads transition from training centers to inference centers. McKinsey estimates that by 2030, global data centers will require approximately $6.7 trillion in investments to meet computational power demands, with about $5.2 trillion related to AI inference processing load data center capital expenditures; the International Energy Agency predicts that global data center electricity use will double to about 945 terawatt-hours by 2030, with AI-driven accelerated server electricity use growing at an annual rate of approximately 30%; Goldman Sachs Group, Inc. also projects that U.S. data center power demand will increase from 31 gigawatts in 2025 to 66 gigawatts in 2027. Therefore, Meta's strategic focus is ultimately on the most scarce resource in the AI era, "available inference capacity": GPU/ASIC/TPU AI computational power resource clusters, low-latency networks, model hosting, token billing, and efficient scheduling capabilities of super large-scale data centers. Wall Street's valuation anchor for Meta may switch from "advertising platform + metaverse expenditure discount" to "advertising cash cow + AI model entry + computational power infrastructure option"; however, the key to success is not just acquiring chips, but ensuring supply for enterprise sales, AI developer ecosystem, reliable cloud-based AI computational power leasing services, cloud platform software stack, and unit token economics. From the AI superintelligence arms race to challenging AWS, Azure, and Alphabet Inc. Class C cloud, Meta aims to transform "surplus computational power" into a cloud business engine Meta's leading potential plan involves selling access rights to various AI models hosted on Meta's existing AI computational power infrastructure; sources say this approach is similar to Amazon.com, Inc.'s cloud service Bedrock product. Meta will operate data center components and AI chips that support these models, as well as its own recently developed Muse Spark model, and charge developers access fees. Sources also say that the company is considering selling access rights to "raw" cloud-native computational power capacity, similar to so-called new cloud business leaders like CoreWeave Inc. Sources say the development of these new business lines is part of Meta Compute; Meta Compute is an internal program aimed at building and managing the company's AI computational power infrastructure deployment work. Meta Compute is led by Meta's infrastructure head, Santosh Janardhan, Meta's internal AI department leader Daniel Gross, and Meta President Dina Powell McCormick. A Meta spokesperson declined to comment. The company's plans are still in development, and related strategies may change. Meta's stock price initially rose nearly 10% in early trading on US stocks on Wednesday, before partially retracting some of the gains. CoreWeave's stock price plunged over 10% in early trading. Meta has listed the development of AI "superintelligence" as one of its top priorities, pledging to invest billions of dollars in building data centers and other critical AI computational power infrastructure, such as expensive AI chips and a range of AI computational power-related hardware systems believed to be essential to achieving this goal. This investment has made investors anxious about how Meta will gain returns from these billion-dollar-level expenditures, which include major computational power transactions with CoreWeave, Alphabet Inc's Alphabet Inc. Class C, and cloud computing leaders such as Oracle Corporation. Cloud computing businesses indeed offer an important way to recover some of the investments in AI computational power infrastructure. Amazon.com, Inc.'s cloud services, Azure, and Alphabet Inc. Class C's cloud have spent decades building supercloud computing platforms, renting out computational power, large storage capacities for enterprise clients, and certain software access permissions through their internet platform ecosystemthese businesses now generate hundreds of billions of dollars in revenue each quarter. As AI demand increases, these service providers have expanded to renting out dedicated AI chips and a wide range of AI computational power infrastructure resources and capacities required for training and running AI models. This is a complex business that requires a large ecosystem built up of massive data center clusters, software platforms, enterprise sales teams, and customer support operations to operate. SpaceX, led by Elon Musk, has recently become a key player in this cloud-based computational power resource leasing space after acquiring its AI startup xAI in February; the company recently rented out its large AI data center underlying computational resources in Memphis to AI applications leader Anthropic PBC, and entered into a cloud-based AI computational power leasing deal with Alphabet Inc. Class C. According to a forecast by Bloomberg Intelligence, this strategy could help xAI generate over $50 billion in revenue by 2028, and at least $100 billion by 2030. Despite its complexities, Meta CEO Mark Zuckerberg has indicated to investors that he is willing to sell surplus computational power infrastructure and even launch so-called large model API services, allowing customers to pay for their AI usagethe business is usually measured in "token scale," where customers pay for the vast amounts of data queries they use and generate. Zuckerberg said during a shareholder conference call in May: "This is definitely within our consideration. Nearly every week, different external companies come to us, both requesting us to build a professional API service system and inquire about whether we have computational power they can purchase from us, and are willing to pay a premium above our acquisition costs." Zuckerberg added: "We haven't done this yet because we believe we have uses for this computational power. But obviously, if we reach a stage where we believe we have built too much, then this is an option we have, which also gives us more confidence in investing in building this infrastructure." In the rapidly evolving arms race of AI computational power infrastructure, Zuckerberg has indicated multiple times that he believes the industry is facing computational power constraints and Meta should accumulate as much computational power as possible before deciding its use. Behind Meta's transformation of AI cloud into a "computational power recovery engine": Agent Runtime eruption, semiconductor shortages resonating with cloud-based AI inference computational power demand on the rise The core logic behind Meta's development of an AI cloud business is to transform the significant ASIC/TPU/GPU, data center CPU, data center power chain, optical interconnect network, storage component, and other AI computational power assets formed in the arms race for "superintelligence" into an AI computational power infrastructure platform that can be externally leased. For the capital markets, this means installing a "residual value recovery mechanism" on Meta's multibillion-dollar AI capital expenditures: serving ad recommendations, video generation, AI assistants, and superintelligence training when in use; and converting into a significant revenue category of APIs, model hosting, inference clusters, or GPU rentals when idle, alleviating market anxiety about the return on AI capital expenditures. The increasingly leveraged theme of AI semiconductor trading and position crowding, coupled with pressure from industry leaders such as Apple Inc. to raise prices, has seen the Philadelphia Semiconductor Index fall by as much as 7.9% in one day, with multiple instances of sharp fluctuations exceeding 5% in a month, highlighting that the AI computational power industrial chain associated with semiconductors has entered a high volatility, leverage, extreme position crowding, and high expectations phase of cash out pressure. This is why institutional investors have recently begun to focus on emphasizing narratives such as "the peak of the semiconductor market has been seen," and "the AI bubble is gradually bursting," which are overly bearish bear market narratives. However, a recent research report from Nomura, a well-known Wall Street investment institution, refutes the "semiconductor peak" thesis. The key counterargument to the "semiconductor peak" thesis by Nomura is not simply stating that AI chip prices will continue to rise, but pointing out that the demand for AI cloud infrastructure is transitioning from a single point GPU shortage to system-wide component mismatches. According to Nomura's research framework, AI server revenue is expected to grow by 78% and 76% respectively in 2026 and 2027, with the number of global data center projects expected to increase from 240 to 280, with about 50 gigawatt projects, and an additional 32 gigawatts of compute power deployment expected in 2027, with visibility for 23 gigawatts in 2028; but the real bottleneck is transitioning from GPU capacity, Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR CoWoS advanced packaging to wafer-level substrates, AI PCBs, copper-clad laminates (CCL), electronic fabrics, MLCCs, glass substrates/ABF substrates, IC carrier boards, high-end capacitors, power management chips, data center optical high-speed optical interconnect components, and overflows. It is reported that McKinsey's medium- to long-term projections also support Nomura's emphasis on this direction: by 2030, the global computational power value chain will require approximately $5.2 trillion to meet AI-related demands, corresponding to approximately 156 gigawatts of AI-related data center capacity demand. This means that the main storyline of semiconductor trading is not about reaching a peak, but rather about "rotating shortages": from GPUs to HBMs, from advanced packaging to substrate materials, and then to power, liquid cooling, networking, and cloud scheduling software, profit upgrades and price hikes may still be the most potent catalysts for the core hardware chain of AI computational power. The popular "Agent Runtime" theory currently sweeping the globe further explains why the demand for cloud-based AI inference may be close to endless. Traditional model hosting is a stateless service of "input-inference-output," while intelligent AI is a cyclical workflow: planning, calling models, calling tools, observing, retrying, correcting, until the task is completed. The latest market use cases are crucial: even with a single call success rate of 95%, the success rate of a task after 15 consecutive calls is only about 46.3%, meaning that what businesses are really purchasing is not cheap tokens but a runtime system that is programmable, observable, traceable, cost-controllable, and auditable. The emphasis from Nebius on moving from running over 200MW of computational power to reaching 800MW by the end of the year and locking in over 3GW of reserved capacity reflects how cloud infrastructure is upgrading from "selling GPU hours" to "selling deterministic results." The International Energy Agency also predicts that global data center electricity consumption will double to about 945TWh by 2030, indicating that Agent Runtime is not just a simple software architecture change, but will also lead to investments in GPU clusters, power capacity, network switches, storage retrievals, model routing, and observational platforms for a full-stack transformation. Behind Meta's entry into cloud computing is the explosive breakthrough of the Agent Runtime, resonating with the demand for cloud-based AI inference computational power and shortages in the semiconductor market. The demand side is amplified by intelligent task cycles that increase token consumption, while the supply side is constrained by advanced packaging and small components limiting AI server deliveries, and the business side is dominated by Meta, CoreWeave, Nebius, AWS, Azure, and Alphabet Inc. Class C cloud competing for the monetization entry of "available AI computational power." AI cloud computing is no longer just the story of traditional cloud providers' IaaS CPU and scaling software, but has evolved into a composite cycle of "computational power asset securitization + inference economics + semiconductor price chain"; the assets that benefit the most are usually not from a single GPU narrative but are nodes that have supply bottlenecks, pricing power, and cash flow visibility. --- This translation was done by a professional translator.