Meta made a decision, and the storage sector plummeted.

On July 1st Beijing Time, foreign media reported that Meta is building a cloud computing business and will sell AI computing power to external customers.

This matter is not without signs. Five weeks ago, when Mark Zuckerberg was asked at Meta's annual shareholders' meeting if it would compete with Amazon and Microsoft in the cloud computing field, he gave a clear response: "It's definitely on the table."

He also revealed a detail: "Almost every week, external companies come to us, either asking if we can open an API or if they can pay a premium to buy Meta's computing power."

It only took five weeks from "under consideration" to "under construction." After the news spread, Meta's stock price soared, while the AI infrastructure stocks in the U.S. stock market tumbled.

At the close of the U.S. stock market in the early hours of July 2nd, Meta rose 8.81%. The Philadelphia Semiconductor Index plunged more than 6%. Micron Technology fell 10.57%, SanDisk fell more than 11%, Intel fell more than 7%, and ASML, AMD, TSMC, and ARM fell more than 5%. Independent cloud computing providers suffered even more severe selling. Nebius tumbled more than 14.5%, and CoreWeave fell more than 13%.

01 How much has Meta invested in AI?

In 2026, the total capital expenditure of the world's four major technology giants (Meta, Microsoft, Alphabet, and Amazon) is approximately $725 billion, a 77% increase from about $410 billion in 2025. Among them, Meta's CapEx guidance is between $125 billion and $145 billion. This figure was revised upward by $10 billion from the previous $115 billion - $135 billion when the Q1 financial report was released at the end of April.

In addition to building its own data centers, Meta has also signed several large external contracts this year: a five - year strategic agreement with AMD worth $60 billion to purchase 6 gigawatts of customized Instinct GPUs; a $21 billion AI computing power infrastructure contract with CoreWeave; and a computing power procurement agreement of up to $27 billion with Nebius. The total of these three external contracts exceeds $100 billion.

However, there is a fundamental difference between Meta's investment situation and the other three. Microsoft has Azure, Google has GCP, and Amazon has AWS. Their huge CapEx investments are directly offset by cloud service revenues. Meta doesn't have such an offset. Every penny it previously invested in infrastructure was a pure cost item, all used to serve its own advertising recommendation system and AI applications, with no part being a product for external sales.

Sherwood News pointed out directly in its analysis in May that compared with other technology giants making large - scale investments, Meta doesn't have the high - profit cloud business and enterprise - level revenues to buffer the impact.

This also explains an abnormal phenomenon: Meta exceeded Wall Street's profit expectations for two consecutive quarters in 2026, but its stock price has still fallen about 4% since the beginning of the year. The core question from the market is where the return is for spending $135 billion a year to build data centers.

02 Zuckerberg's calculation: Buying insurance

Zuckerberg's exact words at the shareholders' meeting were: "We're not doing this right now because we think this computing power can still be put to good use. But obviously, if one day we feel that we've overbuilt, then this is also an option; to some extent, this also strengthens our confidence to continue investing and building."

There are two keywords. "If we have overbuilt," he is leaving a way out for the possibility of over - construction. "Partially what gives us confidence," the existence of the option to do cloud business itself is his confidence to continue spending money. In other words, Meta didn't build data centers because it wanted to do cloud business, but because it built too many data centers and thus needs cloud business to provide a safety net.

Datafloq, a technology content platform that has long focused on big data, AI, and cloud computing, pointed out in its analysis in early June that this makes it easy for investors to understand Meta's capital expenditure bet as an either - or judgment - either the internal AI investment is successful or it fails.

But in fact, doing cloud business is an option. If the internal monetization of AI is successful, all the computing power will be used internally, and the cloud business can be abandoned. If the internal consumption fails to meet expectations, the surplus computing power doesn't have to depreciate on the books and can be turned into revenue. It turns "losing everything if the bet fails" into "still getting rent even if the bet fails."

But reading the same statement in reverse, one can also sense anxiety. Foreign media comments were sharp: "If you can't use it all up yourself, then shift the cost to others. This is not something someone who is confident about the future of AI would say. If Zuckerberg really believes that the internal demand can consume all the computing power, he has no reason to share the precious GPU resources with external competitors."

03 What does Meta lack to do cloud business? It's not just about having GPUs

Having a GPU cluster doesn't mean you can do cloud business.

Meta lacks a lot of things, which can be listed as follows: an enterprise - level multi - tenant isolation architecture, security and compliance certifications (such as SOC 2, HIPAA, ISO 27001, etc.), a fine - grained billing and SLA guarantee system, global multi - regional deployment and network access nodes, and most importantly, an enterprise sales team and a customer success system.

Since its establishment, Meta has been a pure to - C company. It has never sold anything to enterprise customers and has no muscle memory for B2B sales.

Datafloq's analysis made a judgment on Meta's possible path: "Trying to build a full - stack cloud platform is a strategic mistake. The right way is to make a narrow cut."

The article listed four possible product forms: First, renting bare computing power, priced by the hour, with no long - term contracts, and scheduling the GPU cluster through API. Second, hosting Llama model inference, allowing enterprises to run Llama without building their own GPU infrastructure. Third, enterprise model fine - tuning services, fine - tuning open - source models with private data on Meta's hardware. Fourth, Agent infrastructure, providing dedicated tool calls, credential management, and audit logs for AI Agent workloads.

This means that in the short term, Meta's cloud business will most likely be in the form of "wholesale" computing power sales, targeting a small number of large customers and signing long - term contracts, similar to CoreWeave's model. Instead of providing a complete cloud platform like AWS, which offers self - registration, on - demand use, and hundreds of services. The organizational capabilities and customer ecosystem required for the latter cannot be developed in just two or three years.

Meanwhile, on May 28th, Meta also did two other things: it announced the launch of paid subscription tiers for Instagram, WhatsApp, and Facebook, and according to The Information, it established a new "Enterprise Solutions" department, sending engineers and product managers directly into large enterprise customers to help deploy AI tools.

These three things form a complete narrative: Meta is systematically looking for revenue sources other than advertising to support its CapEx bill. Doing cloud business is just the boldest step among them.

04 Earthquake in the industrial chain: Meta up 6%, CoreWeave and Nebius down 9%

After this news spread, Meta's stock price rose more than 6%, while the AI computing power leasing companies CoreWeave and Nebius both fell more than 9%.

The sharp decline of CoreWeave and Nebius indicates that the market believes this is a re - pricing of the moat of the entire neocloud business model.

The blow is three - fold.

The first layer is the direct competitive threat. The business models of CoreWeave and Nebius are essentially "bulk purchasing GPUs → building clusters → reselling to AI companies at a premium." The premise of high gross profit margins is that the supply of GPU computing power in the market is tight, and customers have few alternative options.

Once Meta, the most aggressive investor in computing power, enters the market, there will be an additional player with a huge supply of computing power. Moreover, Meta's GPU procurement cost is lower than that of neocloud companies because it directly signs strategic deals worth tens of billions of dollars with Nvidia and AMD and gets the best prices. It can sell at a lower price than CoreWeave and still make a profit.

The second layer is more fatal: the identity conflict. One of CoreWeave's largest customers at present is Meta. In April 2026, CoreWeave announced the expansion of its AI infrastructure agreement with Meta, with a total scale of $21 billion and a service period until 2032.

Now that Meta wants to do the same thing itself, it's like your client announcing that it will become your competitor. The natural reaction of the market is to question whether this $21 billion contract will be renewed after it expires. Is Meta just buying time with money and will no longer need CoreWeave after its cloud business is established?

The third layer is the collapse of the valuation narrative. The story CoreWeave told when it went public in March 2025 was "the explosive growth of AI computing power demand, extremely scarce supply, and we are the scarce suppliers." This narrative supported its rocket - like growth from zero to a market value of tens of billions of dollars.

But Meta's entry into the market to sell computing power directly shakes the core premise of "scarce supply." If even the biggest spender on AI computing power in the world thinks it may have surplus computing power to sell, is the supply - demand relationship in the entire market really as tight as previously described?

This doesn't mean that CoreWeave's business will collapse immediately. Its revenue in Q1 2026 was about $2.1 billion, and it has a large backlog of contracts, so its short - term revenue is guaranteed. But the capital market prices based on expectations rather than the current situation. When the largest customer is also the potential biggest competitor, the long - term growth story needs to be rewritten.

05 Is it good news or a warning?

Regarding Meta's move to do cloud business, is it really good news?

Those who are bullish believe that this is an upgrade of Meta's investment logic. Previously, Meta's CapEx was a pure one - way bet, betting that AI could significantly increase advertising revenue and user engagement. If it won the bet, the return would be huge; if it lost, it would be an astronomical sunk cost. Now, with the cloud business option, the investment has become "offensive when advancing and defensive when retreating."

The quarterly revenue of the global cloud infrastructure service market in Q1 2026 reached $128.6 billion (data from Synergy Research Group), with an annualized figure of over $455 billion, and AI compute is the fastest - growing sub - sector. Meta only needs to capture a small share to generate considerable revenue. From the perspective of portfolio theory, this turns Meta's CapEx from a "high - risk single bet" into a "two - way option with a hedge."

Those who are bearish believe that this is precisely an "early warning signal" of the AI CapEx bubble. The logic is simple: if Meta really believes that the internal AI demand can consume all the computing power and generate corresponding returns, why would it share the precious GPU resources with external competitors? The very act of doing cloud business is a hedge against the possibility that the internal AI monetization speed fails to meet expectations.

The total CapEx of the four giants in 2026 is about $725 billion, but the incremental revenue directly brought by AI may only be in the tens of billions. The input - output ratio is seriously mismatched. Meta's move to do cloud business may be the most aggressive player preparing for the possibility of computing power surplus.

There is also a technical concern. The efficiency of AI inference has been improving rapidly in the past year, and the unit inference cost has been cut by a large margin every few months. If the speed of efficiency improvement continues to exceed the speed of demand growth, the data centers built today may not be "insufficient" but "too many." Meta's move to do cloud business is an insurance policy against this possibility.

On the same day, the U.S. stock storage sector tumbled. Micron, SanDisk, etc., all fell about 10%. The core logic behind the sharp rise of these companies in the past year was "the wave of AI data center construction driving the explosive demand for HBM and enterprise - level SSDs." Micron's revenue in the previous quarter increased by 196% year - on - year, and the story it told was "unlimited demand and insufficient supply."

But Meta's news directly shakes the underlying assumption of "insufficient construction." If the data center construction pace of technology giants slows down in the future, it means that the growth expectation of HBM and enterprise - level storage procurement will be revised downward.

This is also a story about the second half of the AI arms race. In the past two years, everyone has been competing in terms of who dares to spend more money, who can grab more GPUs, and who can build larger data centers.

But even Zuckerberg, who has been the most daring in spending, is "afraid." "After building so much infrastructure, we need to ensure that we won't lose everything whether AI monetization is fast or slow."

When the biggest buyer starts to prepare to become a seller, who will be the real buyers?

This article is from the WeChat official account "Tencent Technology", author: Worth Paying Attention To. It is published by 36Kr with authorization.