Meta is advancing plans for a cloud infrastructure business.

According to today's Bloomberg report (which the Cailian Press item summarizes), Meta is actively developing a new cloud offering under its "Meta Compute" initiative. The company intends to sell excess AI computing capacity as well as access to its AI models (such as those in the Llama family and Muse Spark) to external customers. Two main approaches appear to be under consideration: hosted model access (similar to AWS Bedrock, where Meta manages the infrastructure) and raw compute sales (comparable to specialist providers like CoreWeave).

This move aims to monetize the enormous infrastructure Meta has been building for its own AI workloads. The company has repeatedly raised its capital expenditure guidance, now expecting $125–145 billion in 2026, and launched the Meta Compute organization earlier in the year with plans for tens of gigawatts of capacity. In May, Mark Zuckerberg stated that entering the cloud business was “definitely on the table” if the company ended up with excess capacity after meeting internal needs.

Market reaction has been strongly positive. META shares rose approximately 8–10% today on the news, as investors view it as a path to better returns on the massive AI investments that had raised ROI concerns. Conversely, some pure-play AI cloud and GPU infrastructure stocks (e.g., CoreWeave-related names or neoclouds) saw pressure on expectations of increased competition.

In short, Meta is positioning itself as a potential fourth major hyperscaler with an AI-first focus. It will compete directly with Amazon AWS, Microsoft Azure, and Google Cloud while leveraging its open-source model advantage and vast internal infrastructure. The development is still in planning stages, but it represents a logical evolution of Meta’s aggressive buildout and could materially change the economics of its AI strategy.

Would you like the full English translation of the Cailian report, deeper competitive analysis, specific stock moves, or details from the original Bloomberg article?

知情人士表示，Meta公司在近年来一直在加速进行数据中心以及其他基础设施的建设工作，以便支撑自身人工智能（AI）战略。目前，该公司正在筹备建立一项新业务，借助向外部客户出售所富余算力的方式来实现收入增长。

其中一种方案是，Meta借助向客户开放部署在其现有AI基础设施之上的多种AI模型来实现这一目标，这种方式类似于AWS的Bedrock服务。知情人士表示，Meta将会负责运行那些支撑这些模型的数据中心以及芯片，其中包括其自研的Muse Spark模型，并且会向开发者收取调用所产生的费用。

Meta is advancing plans for a cloud infrastructure business under its internal “Meta Compute” initiative.[[1]](https://www.bloomberg.com/news/articles/2026-07-01/meta-is-building-a-cloud-business-to-sell-excess-ai-compute)

The paragraph you provided continues the Cailian Press summary of today’s Bloomberg report. It confirms that Meta is actively exploring two complementary approaches to monetize excess AI capacity generated by its massive infrastructure buildout: offering hosted access to its AI models (including the self-developed Muse Spark family) in a manner similar to AWS Bedrock, where Meta manages the underlying data centers, chips, and operations while charging developers based on usage; and directly selling raw computing capacity (“裸算力”) in a model akin to specialist neocloud providers such as CoreWeave.[[2]](https://finance.yahoo.com/technology/ai/articles/meta-sell-excess-ai-computing-125201412.html)

Full combined English translation of the excerpts shared so far:

“Insiders indicated that in recent years Meta has been accelerating the construction of data centers and other infrastructure to support its artificial intelligence (AI) strategy. The company is currently preparing to establish a new business to generate revenue growth by selling its surplus computing power to external customers.

One approach is for Meta to open up various AI models deployed on its existing AI infrastructure to customers, similar to AWS’s Bedrock service. Insiders say Meta will be responsible for running the data centers and chips supporting these models, including its self-developed Muse Spark model, and will charge developers for usage fees.

In addition, Meta is also considering directly selling ‘raw computing capacity’ (raw computing capacity), with a model similar to emerging AI cloud computing companies such as CoreWeave. According to insiders, the development of these new businesses is part of Meta’s internal ‘Meta Compute’ project.”

This aligns with earlier signals from Mark Zuckerberg in May 2026, when he noted that entering the cloud business was “definitely on the table” if the company generated excess capacity after meeting internal AI needs.[[3]](https://www.cnbc.com/2026/05/27/mark-zuckerberg-says-meta-starting-cloud-business-on-the-table.html) The Meta Compute organization, launched earlier in 2026 and led by infrastructure head Santosh Janardhan, Daniel Gross from Meta Superintelligence Labs, and President Dina Powell McCormick, oversees plans for tens of gigawatts of capacity this decade (with ambitions for hundreds longer-term).[[4]](https://www.theedgesingapore.com/news/tech/meta-building-cloud-business-sell-excess-ai-compute--bloomberg)

Market reaction has been strongly positive. META shares rose roughly 6–10% today, as investors interpret the move as a credible path to improve returns on the company’s sharply raised capital-expenditure guidance of $125–145 billion for 2026.[[2]](https://finance.yahoo.com/technology/ai/articles/meta-sell-excess-ai-computing-125201412.html)[[5]](https://www.facebook.com/KobeissiLetter/posts/breaking-meta-is-developing-a-cloud-infrastructure-business-that-will-sell-acces/1346841864257893/) Some pure-play AI cloud and infrastructure names faced selling pressure on expectations of heightened competition.

Strategic implications are significant. Meta has been one of the largest customers of neoclouds such as CoreWeave (with cumulative deals reportedly exceeding $35 billion in 2026), Google, Oracle, and others to supplement its own buildout. Transitioning portions of that capacity into an external offering could transform a major cost center into a diversified revenue stream beyond advertising, position Meta as a potential fourth hyperscaler with an AI-first and open-source (Llama family) advantage, and help address ongoing investor questions about the ROI of its aggressive AI infrastructure investments. Plans remain in development and could evolve, but the direction is now clear.

In short, this represents a logical evolution of Meta’s AI strategy—turning surplus scale into a commercial cloud business while leveraging its proprietary models like Muse Spark (Meta’s advanced multimodal reasoning system released in April 2026). It could materially reshape competitive dynamics in AI cloud infrastructure.

Would you like the original Bloomberg article browsed for additional quotes or timelines, a deeper competitive comparison (e.g., versus CoreWeave or AWS), specific analyst reactions, or further details on Muse Spark’s capabilities?

Meta Compute项目负责对公司的AI基础设施进行建设和管理工作，该项目由Meta基础设施负责人Santosh Janardhan、Meta Superintelligence Labs负责人之一Daniel Gross以及Meta总裁Dina Powell McCormick共同领导。

Meta is advancing plans for a cloud infrastructure business under its internal “Meta Compute” initiative, although these plans remain in the formulation stage and the future strategy may still be adjusted.

According to the Cailian Press summary of today’s Bloomberg report, Meta has been accelerating construction of data centers along with other infrastructure in recent years in order to support its artificial intelligence (AI) strategy. The company is currently preparing to establish a new business that can achieve revenue growth through selling surplus computing power to external customers.

One approach involves Meta opening up various AI models deployed upon its existing AI infrastructure to customers in a manner similar to the AWS Bedrock service. Insiders indicate that Meta will take responsibility for running the data centers as well as the chips that support these models—including its self-developed Muse Spark model—and will charge developers on the basis of usage fees.

In addition, Meta is also considering direct sales of raw computing capacity, employing a model similar to that of emerging AI cloud computing companies such as CoreWeave. According to insiders, the development of these new businesses forms part of the company’s internal “Meta Compute” project. The Meta Compute project carries responsibility for the construction and management of the company’s AI infrastructure work and is jointly led by infrastructure head Santosh Janardhan, one of the heads of Meta Superintelligence Labs Daniel Gross, and Meta President Dina Powell McCormick.

It is reported that Meta’s related plans are still in the formulation stage, and future strategies may still be adjusted.

This aligns with Mark Zuckerberg’s May 2026 statement that entering the cloud business was “definitely on the table” if excess capacity remained after internal needs were met. Meta has raised its 2026 capital expenditure guidance to $125–145 billion and is planning for tens of gigawatts of capacity this decade. The dual approach—hosted model access (Meta-managed, usage-based) and raw compute sales—would allow the company to monetize infrastructure that has largely been built for its own AI workloads, including the Llama family and Muse Spark multimodal reasoning model.

Market reaction has been positive. META shares rose roughly 6–10% on the news, as investors viewed the potential cloud offering as a path to improved returns on AI spending. Conversely, certain pure-play AI infrastructure and neocloud names faced pressure amid heightened competition expectations.

Strategic implications remain significant despite the early-stage disclaimer. Meta has been one of the largest customers of providers such as CoreWeave (with cumulative commitments reportedly exceeding tens of billions). Converting portions of its buildout into an external commercial service could transform a major cost center into diversified revenue beyond advertising, establish Meta as an AI-first hyperscaler with open-source advantages, and address ROI concerns. Execution risks around power, supply chains, and go-to-market remain, and the plans could still evolve.

Would you like the full original Bloomberg article details, a competitive comparison with AWS Bedrock and CoreWeave, deeper analysis of Muse Spark capabilities, or specific analyst reactions?

云业务可增加Meta收入

Meta公司已经将开发AI“超级智能”（superintelligence）的工作列为公司的最高优先级，并且承诺投入数千亿美元来进行数据中心建设以及高端AI芯片采购等基础设施方面的工作，从而有效支撑这一目标的实现。

Meta is advancing plans for a cloud infrastructure business to monetize its massive AI investments amid an industry-wide arms race that has triggered broad investor concerns over returns.

According to people familiar with the matter, Meta has accelerated construction of data centers and other infrastructure in recent years to support its artificial intelligence strategy. The company is preparing a new business that generates revenue growth by selling surplus computing power to external customers. One approach is to offer customers access to various AI models deployed on its existing infrastructure, similar to AWS Bedrock. Meta would manage the data centers and chips supporting these models—including its self-developed Muse Spark model—and charge developers usage-based fees. In addition, Meta is considering direct sales of raw computing capacity in a model similar to specialist providers such as CoreWeave. These efforts form part of the internal “Meta Compute” project, which is responsible for constructing and managing the company’s AI infrastructure and is jointly led by infrastructure head Santosh Janardhan, Meta Superintelligence Labs executive Daniel Gross, and President Dina Powell McCormick. It is reported that the related plans remain in the formulation stage and future strategy may still be adjusted. In addition to Meta, Google, Microsoft, and Amazon are also competing to invest enormous funds in the AI arms race. This has triggered investor concerns about how these companies will achieve returns.

This Bloomberg-reported development directly addresses the ROI worries highlighted across the hyperscalers. Meta has raised its 2026 capital expenditure guidance to $125–145 billion and is planning for tens of gigawatts of capacity this decade (with ambitions for hundreds longer-term) to support its superintelligence goals and models such as the Llama family and Muse Spark (its advanced multimodal reasoning system). By offering both hosted model access (managed like Bedrock) and raw compute (like CoreWeave), Meta can convert a major cost center into diversified revenue beyond advertising, position itself as an AI-first hyperscaler with open-source advantages, and improve capital efficiency. Plans are still evolving.

Market reaction has been strongly positive. META shares rose roughly 6–10% on the news as investors saw a credible path to better returns on AI spending, while certain pure-play AI infrastructure and neocloud names faced pressure on heightened competition expectations.

This represents a logical evolution of Meta’s strategy and could reshape competitive dynamics in AI cloud infrastructure. Would you like the full original Bloomberg article details, a competitive comparison with AWS Bedrock/CoreWeave/Azure/Google Cloud, deeper analysis of Muse Spark capabilities, specific 2026 capex figures across the hyperscalers, or analyst reactions?

Meta is advancing plans for a cloud infrastructure business under its internal “Meta Compute” initiative, even as it has already signed multiple major compute procurement agreements with providers such as CoreWeave, Google, and Oracle.

According to people familiar with the matter, Meta has accelerated the construction of data centers and other infrastructure in recent years to support its artificial intelligence (AI) strategy. The company is currently preparing to establish a new business to generate revenue growth by selling its surplus computing power to external customers.

One approach is for Meta to open up various AI models deployed on its existing AI infrastructure to customers, similar to AWS’s Bedrock service. Insiders say Meta will be responsible for running the data centers and chips supporting these models, including its self-developed Muse Spark model, and will charge developers for usage fees.

In addition, Meta is also considering directly selling ‘raw computing capacity’, with a model similar to emerging AI cloud computing companies such as CoreWeave. According to insiders, the development of these new businesses forms part of the company’s internal “Meta Compute” project. The Meta Compute project carries responsibility for the construction and management of the company’s AI infrastructure and is jointly led by infrastructure head Santosh Janardhan, one of the heads of Meta Superintelligence Labs Daniel Gross, and Meta President Dina Powell McCormick.

It is reported that Meta’s related plans are still in the formulation stage, and future strategies may still be adjusted. Meta has previously signed multiple computing power cooperation agreements with companies such as CoreWeave, Google, and Oracle. Meta has listed the development of AI “superintelligence” (superintelligence) as the company’s highest priority and has committed hundreds of billions of dollars to infrastructure work such as data center construction and high-end AI chip procurement in order to support the realization of this goal. In addition to Meta, Google, Microsoft, and Amazon are also competing to invest enormous funds in the AI arms race. This has triggered investor concerns about how these companies will achieve returns.

This Bloomberg-reported development (as summarized by Cailian Press) directly addresses the ROI worries that have surrounded the hyperscalers’ AI buildouts. Meta has raised its 2026 capital expenditure guidance to $125–145 billion and is planning for tens of gigawatts of capacity this decade (with ambitions for hundreds longer-term) to support its superintelligence goals and models such as the Llama family and Muse Spark (its advanced multimodal reasoning system). By offering both hosted model access (managed like Bedrock) and raw compute (like CoreWeave), while already being one of the largest customers of those same providers, Meta can convert a major cost center into diversified revenue beyond advertising, position itself as an AI-first hyperscaler with open-source advantages, and improve capital efficiency. Plans are still evolving.

Market reaction has been strongly positive. META shares rose roughly 6–10% on the news as investors saw a credible path to better returns on AI spending, while certain pure-play AI infrastructure and neocloud names faced pressure on heightened competition expectations.

This represents a logical evolution of Meta’s strategy and could reshape competitive dynamics in AI cloud infrastructure. Would you like the full original Bloomberg article details, a competitive comparison with AWS Bedrock and CoreWeave, deeper analysis of Muse Spark capabilities, specific 2026 capex figures across the hyperscalers, or analyst reactions?

Meta推出云计算业务这一举措，被视为其回收部分投资的一种有效方式。

AWS、Azure以及Google Cloud在经历了数十年的发展之后，已经成功构建起了借助互联网来出租算力、存储以及软件服务的平台，目前每个季度均能够贡献数百亿美元的收入。

Meta is advancing plans for a cloud infrastructure business under its internal “Meta Compute” initiative to monetize surplus AI capacity, even as the latest details highlight the substantial operational complexity of competing in this market.

According to people familiar with the matter, Meta has accelerated the construction of data centers and other infrastructure in recent years to support its artificial intelligence (AI) strategy. The company is preparing a new business that generates revenue growth by selling its surplus computing power to external customers.

One approach is to offer customers access to various AI models deployed on its existing infrastructure, similar to AWS Bedrock. Meta would manage the data centers and chips supporting these models—including its self-developed Muse Spark model—and charge developers usage-based fees. In addition, Meta is considering direct sales of raw computing capacity in a model similar to specialist providers such as CoreWeave. These efforts form part of the internal “Meta Compute” project, responsible for constructing and managing the company’s AI infrastructure. It is jointly led by infrastructure head Santosh Janardhan, Meta Superintelligence Labs executive Daniel Gross, and President Dina Powell McCormick. The related plans remain in the formulation stage and future strategy may still be adjusted.

Meta has previously signed multiple computing power agreements with companies such as CoreWeave, Google, and Oracle. It has listed the development of AI “superintelligence” as the company’s highest priority and has committed hundreds of billions of dollars to infrastructure work such as data center construction and high-end AI chip procurement to support this goal. In addition to Meta, Google, Microsoft, and Amazon are also competing to invest enormous funds in the AI arms race. This has triggered investor concerns about how these companies will achieve returns.

Meta launching this cloud computing business is seen as an effective way for it to recoup part of its investment. After decades of development, AWS, Azure, and Google Cloud have successfully built platforms that rent computing power, storage, and software services via the internet, currently contributing tens of billions of dollars in revenue each quarter. With the rapid growth of AI demand, these cloud service providers have also begun to rent out the specialized chips and computing resources required for training and running AI models. However, this business is very complex. It not only requires massive data center clusters but also a complete software platform, enterprise sales teams, and customer support systems.

Strategic implications remain significant. The latest paragraph underscores the high barriers to entry in the AI cloud segment—beyond raw infrastructure, success demands mature orchestration software, enterprise-grade sales motions, and 24/7 customer support that the hyperscalers have refined over decades. Meta already operates at hyperscale for its own workloads, possesses popular open-source Llama models plus the proprietary Muse Spark multimodal reasoning system (released April 2026), and has raised 2026 capital expenditure guidance to $125–145 billion with plans for tens of gigawatts this decade. By converting portions of that capacity into external offerings (starting potentially with easier hosted-model access before full raw compute), it can transform a major cost center into diversified revenue beyond advertising, address ROI concerns, and position itself as an AI-first fourth hyperscaler. Execution risks around the full stack noted in the report are real, and plans could still evolve. Mark Zuckerberg previously indicated in May 2026 that entering the cloud business was “definitely on the table” if excess capacity remained.

Market reaction has been strongly positive. META shares rose approximately 7–10% on the news, as investors viewed it as a credible path to better returns on AI spending, while certain pure-play AI infrastructure and neocloud names faced pressure on heightened competition expectations.

This Bloomberg-reported development (via Cailian Press) represents a logical evolution of Meta’s superintelligence strategy and could reshape competitive dynamics in AI cloud infrastructure.

Would you like the full original Bloomberg article details (via direct access for quotes and timelines), a competitive comparison with AWS Bedrock and CoreWeave, deeper analysis of Muse Spark capabilities, specific 2026 capex figures across the hyperscalers, or analyst reactions?

扎克伯格曾称愿出售富余算力

尽管这一业务进入门槛相对较高，Meta首席执行官马克·扎克伯格此前已经向投资者方面表示，公司愿意考虑出售富余算力，甚至推出按照API调用进行收费的AI服务。这类业务通常以“Token”（模型处理和生成的数据量）来作为计费单位。

Meta is advancing plans for a cloud infrastructure business under its internal “Meta Compute” initiative to monetize surplus AI capacity and address investor concerns about returns on massive infrastructure investments.

According to people familiar with the matter, Meta has accelerated the construction of data centers and other infrastructure in recent years to support its artificial intelligence (AI) strategy. The company is preparing to establish a new business to generate revenue growth by selling its surplus computing power to external customers.

One approach is for Meta to open various AI models deployed on its existing AI infrastructure to customers, similar to AWS’s Bedrock service. Meta would be responsible for running the data centers and chips supporting these models—including its self-developed Muse Spark model—and would charge developers usage-based fees.

In addition, Meta is considering directly selling raw computing capacity in a model similar to specialist AI cloud providers such as CoreWeave. The development of these new businesses forms part of the company’s internal “Meta Compute” project, which is responsible for the construction and management of the company’s AI infrastructure. It is jointly led by infrastructure head Santosh Janardhan, Meta Superintelligence Labs executive Daniel Gross, and President Dina Powell McCormick.

The related plans are still in the formulation stage, and future strategies may still be adjusted. Meta has previously signed multiple computing-power cooperation agreements with companies such as CoreWeave, Google, and Oracle. The company has listed the development of AI “superintelligence” as its highest priority and has committed hundreds of billions of dollars to infrastructure work such as data center construction and high-end AI chip procurement to support this goal. In addition to Meta, Google, Microsoft, and Amazon are also competing to invest enormous funds in the AI arms race. This has triggered investor concerns about how these companies will achieve returns.

Meta launching this cloud computing business is seen as an effective way for it to recoup part of its investment. After decades of development, AWS, Azure, and Google Cloud have successfully built platforms that rent computing power, storage, and software services via the internet; each currently contributes tens of billions of dollars in revenue per quarter. With the rapid growth of AI demand, these cloud service providers have also begun to rent out the specialized chips and computing resources required for training and running AI models. However, this business is very complex. It not only requires massive data center clusters but also a complete software platform, enterprise sales teams, and customer support systems.

Despite the relatively high barriers to entry for this business, Meta CEO Mark Zuckerberg previously told investors that the company is willing to consider selling surplus computing power and even launching AI services charged on an API-call basis (typically billed by “Tokens” — the volume of data the model processes and generates).

Zuckerberg said at this year’s May shareholder conference call: “This is absolutely one of our options. Almost every week, external companies contact us hoping that we will launch API services, or inquire whether they can purchase our computing power, and are even willing to pay a price higher than our procurement cost.”

This direct quote underscores strong external demand and the potential for attractive pricing, strengthening the economic case for Meta’s move.

Market reaction has been strongly positive. META shares rose roughly 7–10% on the news, as investors see a credible path to better returns on the sharply increased 2026 capital expenditure guidance of $125–145 billion. Some pure-play AI infrastructure and neocloud stocks faced selling pressure amid expectations of new competition.

Strategic implications are significant. By leveraging its massive self-built infrastructure (tens of gigawatts planned this decade), popular open-source Llama models, and proprietary systems like Muse Spark (advanced multimodal reasoning model released in April 2026), Meta could evolve from a major buyer of cloud capacity into a fourth AI-first hyperscaler. This could diversify revenue beyond advertising, improve ROI in the superintelligence race, and capitalize on the very demand Zuckerberg highlighted. Execution challenges around building the full software stack, sales organization, and support systems remain substantial, as the article notes, and plans could still evolve.

Would you like the original Bloomberg article browsed for additional quotes or timelines, a competitive comparison with AWS Bedrock and CoreWeave, deeper analysis of Muse Spark capabilities, specific 2026 capex figures across the hyperscalers, or analyst reactions?

不过，他当时也明确表示，我们目前还没有这样做，这是因为我们认为这些算力还具备其自身的用途。但是，如果未来我们认为算力建设出现了过剩的情况，那么这将会是我们可以采取的一项选择，这同时也是我们有信心持续扩大AI基础设施投资的部分原因。

Meta is advancing plans for a cloud infrastructure business under its internal “Meta Compute” initiative to monetize potential surplus AI capacity, while Zuckerberg has repeatedly stressed that computing power supply remains the industry’s primary bottleneck — leading the company to stockpile aggressively first before deciding on future allocation.

According to people familiar with the matter, Meta has accelerated the construction of data centers and other infrastructure in recent years to support its artificial intelligence strategy. The company is preparing a new business to generate revenue growth by selling its surplus computing power to external customers. One approach is for Meta to open various AI models deployed on its existing AI infrastructure to customers, similar to AWS’s Bedrock service. Meta would be responsible for running the data centers and chips supporting these models — including its self-developed Muse Spark model — and would charge developers usage-based fees. In addition, Meta is considering directly selling raw computing capacity in a model similar to specialist AI cloud providers such as CoreWeave. The development of these new businesses forms part of the company’s internal “Meta Compute” project, which is responsible for the construction and management of the company’s AI infrastructure. It is jointly led by infrastructure head Santosh Janardhan, Meta Superintelligence Labs executive Daniel Gross, and President Dina Powell McCormick. The related plans are still in the formulation stage, and future strategies may still be adjusted. Meta has previously signed multiple computing-power cooperation agreements with companies such as CoreWeave, Google, and Oracle. The company has listed the development of AI “superintelligence” as its highest priority and has committed hundreds of billions of dollars to infrastructure work such as data center construction and high-end AI chip procurement to support this goal. In addition to Meta, Google, Microsoft, and Amazon are also competing to invest enormous funds in the AI arms race. This has triggered investor concerns about how these companies will achieve returns. Meta launching this cloud computing business is seen as an effective way for it to recoup part of its investment. After decades of development, AWS, Azure, and Google Cloud have successfully built platforms that rent computing power, storage, and software services via the internet; each currently contributes tens of billions of dollars in revenue per quarter. With the rapid growth of AI demand, these cloud service providers have also begun to rent out the specialized chips and computing resources required for training and running AI models. However, this business is very complex. It not only requires massive data center clusters but also a complete software platform, enterprise sales teams, and customer support systems. Despite the relatively high barriers to entry for this business, Meta CEO Mark Zuckerberg previously told investors that the company is willing to consider selling surplus computing power and even launching AI services charged on an API-call basis (typically billed by “Tokens” — the volume of data the model processes and generates). He stated that this is absolutely one of their options because almost every week external companies contact them hoping that Meta will launch API services or inquire whether they can purchase its computing power, and are even willing to pay a price higher than its procurement cost. However, they have not done so yet because they see internal uses for the capacity. If they later determine there is surplus, this will be an option they can take — and this possibility is part of what gives them confidence to continue expanding AI infrastructure investment. As the AI competition continues to heat up, Zuckerberg has repeatedly stated that he believes the biggest bottleneck facing the entire industry is still the supply of computing power. Therefore Meta should reserve as many computing resources as possible first, and then decide in the future how to utilize them.

This “hoard-now, decide-later” philosophy directly underpins Meta’s sharply raised 2026 capital expenditure guidance of $125–145 billion and plans for tens of gigawatts of capacity this decade (with ambitions for hundreds longer-term). It prioritizes internal needs for superintelligence and models such as the Llama family and Muse Spark (its advanced multimodal reasoning system released in April 2026), while the emerging cloud offering provides flexible upside through hosted model access or raw compute sales. The approach converts a major cost center into potential diversified revenue beyond advertising and helps address investor ROI concerns amid the broader AI arms race. Execution challenges around building the full software stack, sales organization, and support systems remain substantial, and plans could still evolve.

Market reaction has been strongly positive. META shares rose approximately 7–10% on the news, as investors viewed it as a credible path to better returns on AI spending, while certain pure-play AI infrastructure and neocloud names faced pressure on heightened competition expectations.

This Bloomberg-reported development represents a logical evolution of Meta’s superintelligence strategy and could reshape competitive dynamics in AI cloud infrastructure. Would you like the original Bloomberg article details for additional quotes or timelines, a competitive comparison with AWS Bedrock and CoreWeave, deeper analysis of Muse Spark capabilities, specific 2026 capex figures across the hyperscalers, or analyst reactions?

马斯克的SpaceX已入局

值得一提的是，世界首富埃隆·马斯克所旗下的太空与AI公司SpaceX，近期也得以成为这一领域的重要参与者。

Meta is advancing plans for a cloud infrastructure business under its internal “Meta Compute” initiative to monetize surplus AI capacity, even as SpaceX (following its February 2026 acquisition of xAI) has already become an active player by leasing large-scale computing power in its Memphis data center to Anthropic while partnering with Google, with estimates projecting that this approach could drive xAI revenue above $50 billion by 2028 and to $100 billion by 2030.

According to people familiar with the matter, Meta has accelerated the construction of data centers and other infrastructure in recent years to support its artificial intelligence (AI) strategy. The company is preparing to establish a new business that generates revenue growth by selling its surplus computing power to external customers. One approach is for Meta to open various AI models deployed on its existing AI infrastructure to customers, similar to AWS’s Bedrock service. Meta would be responsible for running the data centers and chips supporting these models—including its self-developed Muse Spark model—and would charge developers usage-based fees. In addition, Meta is considering direct sales of raw computing capacity in a model similar to specialist AI cloud providers such as CoreWeave. The development of these new businesses forms part of the company’s internal “Meta Compute” project, which is responsible for the construction and management of the company’s AI infrastructure. It is jointly led by infrastructure head Santosh Janardhan, Meta Superintelligence Labs executive Daniel Gross, and President Dina Powell McCormick. The related plans are still in the formulation stage, and future strategies may still be adjusted. Meta has previously signed multiple computing-power cooperation agreements with companies such as CoreWeave, Google, and Oracle. The company has listed the development of AI “superintelligence” as its highest priority and has committed hundreds of billions of dollars to infrastructure work such as data center construction and high-end AI chip procurement to support this goal. In addition to Meta, Google, Microsoft, and Amazon are also competing to invest enormous funds in the AI arms race. This has triggered investor concerns about how these companies will achieve returns. Meta launching this cloud computing business is seen as an effective way for it to recoup part of its investment. After decades of development, AWS, Azure, and Google Cloud have successfully built platforms that rent computing power, storage, and software services via the internet; each currently contributes tens of billions of dollars in revenue per quarter. With the rapid growth of AI demand, these cloud service providers have also begun to rent out the specialized chips and computing resources required for training and running AI models. However, this business is very complex. It not only requires massive data center clusters but also a complete software platform, enterprise sales teams, and customer support systems. Despite the relatively high barriers to entry for this business, Meta CEO Mark Zuckerberg previously told investors that the company is willing to consider selling surplus computing power and even launching AI services charged on an API-call basis (typically billed by “Tokens” — the volume of data the model processes and generates). He stated that this is absolutely one of their options because almost every week external companies contact them hoping that Meta will launch API services or inquire whether they can purchase its computing power, and are even willing to pay a price higher than its procurement cost. However, they have not done so yet because they see internal uses for the capacity. If they later determine there is surplus, this will be an option they can take — and this possibility is part of what gives them confidence to continue expanding AI infrastructure investment. As the AI competition continues to heat up, Zuckerberg has repeatedly stated that he believes the biggest bottleneck facing the entire industry is still the supply of computing power. Therefore Meta should reserve as many computing resources as possible first, and then decide in the future how to utilize them. It is worth mentioning that SpaceX, the space and AI company under the world’s richest man Elon Musk, has recently also become an important participant in this field. After completing the acquisition of xAI in February this year, SpaceX has leased the large data center computing power located in Memphis, USA, to the AI company Anthropic, and has reached cooperation with Google. It is estimated that this strategy is expected to drive xAI to achieve revenue exceeding $50 billion by 2028 and reach $100 billion by 2030.

This “hoard-now, decide-later” philosophy directly underpins Meta’s sharply raised 2026 capital expenditure guidance of $125–145 billion and plans for tens of gigawatts of capacity this decade (with ambitions for hundreds longer-term). It prioritizes internal needs for superintelligence and models such as the Llama family and Muse Spark (its advanced multimodal reasoning system released in April 2026), while the emerging cloud offering provides flexible upside through hosted model access or raw compute sales. The SpaceX/xAI example in the report serves as a live case study of successful execution at scale (leasing the Colossus 1 Memphis supercluster to Anthropic for multi-billion annual revenue alongside Google partnerships), validating the market demand Zuckerberg highlighted and illustrating the transformative revenue potential in AI infrastructure. The approach converts a major cost center into potential diversified revenue beyond advertising and helps address investor ROI concerns amid the broader AI arms race. Execution challenges around building the full software stack, sales organization, and support systems remain substantial, and plans could still evolve.

Market reaction has been strongly positive. META shares rose approximately 7–10% on the news, as investors viewed it as a credible path to better returns on AI spending, while certain pure-play AI infrastructure and neocloud names faced pressure on heightened competition expectations.

This Bloomberg-reported development represents a logical evolution of Meta’s superintelligence strategy and could reshape competitive dynamics in AI cloud infrastructure. Would you like the original Bloomberg article details for additional quotes or timelines, a competitive comparison with AWS Bedrock, CoreWeave and SpaceX/xAI’s Memphis deals, deeper analysis of Muse Spark capabilities, specific 2026 capex figures across the hyperscalers, or analyst reactions?

来源：Meta据悉布局AI云业务 拟对外出售算力 挑战亚马逊、微软和谷歌 | 财联社