Meta made a decision, storage plummeted.

Author: Xiaojing, Tencent Technology

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

This is not without warning. Five weeks ago, at Meta’s annual shareholder meeting, when Zuckerberg was asked whether the company would compete with Amazon and Microsoft in the cloud computing space, 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 whether they can get an API, or asking whether they can pay more to buy Meta’s computing power.”

From “under consideration” to “under construction” took only five weeks. After the news broke, Meta surged, but it “crashed” the AI infrastructure stocks in the U.S. market.

At the close of trading in the U.S. on the early morning of July 2, 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 all fell more than 5%. Independent cloud computing providers were hit with even more severe sell-offs: Nebius plunged more than 14.5%, and CoreWeave fell more than 13%.

How much has Meta actually invested in AI?

In 2026, the combined capital expenditures of the world’s four tech giants (Meta, Microsoft, Alphabet, Amazon) are expected to total about $725 billion, up 77% year over year from about $410 billion in 2025. Of this, Meta’s own CapEx guidance is $125 billion to $145 billion. This figure was revised upward by $10 billion from the previous $115–135 billion when the Q1 financial results were released in late April.

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

But Meta’s investment situation has a fundamental difference from the other three companies. Microsoft has Azure, Google has GCP, and Amazon has AWS—meaning their massive CapEx spending is directly offset by cloud service revenue. Meta does not have that. Every dollar Meta previously spent on infrastructure was purely a cost item, all used to serve its own advertising recommendation systems and AI applications, with no portion sold as an external product.

In its May analysis, Sherwood News directly pointed out: Compared with other tech giants making similarly large investments, Meta does not have the high-profit cloud business and enterprise revenue to cushion the impact.

This also explains an abnormal phenomenon. In 2026, Meta exceeded Wall Street’s earnings expectations in two consecutive quarters, yet its stock price is still down about 4% year-to-date. The market’s core doubt is: if it spends $135 billion a year building data centers, where exactly is the return?

Zuckerberg’s calculation: buying insurance

Zuckerberg’s exact words at the shareholder meeting were: “We’re not doing this right now because we believe these computing resources still have room to be used. But obviously, if one day we think we’ve overbuilt, then this is also an option; and this, to some extent, strengthens our confidence to continue investing in building.”

Two key phrases. “If we have overbuilt”—he himself is leaving an escape route for the possibility of overbuilding. “Partially what gives us confidence”—the very existence of the option to do cloud is the basis for his confidence to keep spending. In other words, Meta is not building data centers because it wants to do cloud; it needs to do cloud because it has built too many data centers, so it has to provide a backstop.

In early June, Datafloq, a technology content platform focused on big data, AI, and cloud computing, pointed out in an analysis: This makes it easy for investors to interpret Meta’s capital expenditure as a binary bet—either its internal AI investment succeeds, or it fails.

But in reality, doing cloud is an option. If internal monetization of AI works and all computing power is used internally, then there’s no need to do cloud; if internal consumption falls short of expectations, the extra computing power doesn’t have to sit on the books depreciating—it can be turned into revenue. Turning “if the bet loses, you lose everything” into “if the bet loses, you can still collect rent.”

However, reading the same sentence the other way also reveals anxiety. Comments from foreign media were sharp: “If you can’t use it all, shift the costs onto other people. That’s not something someone who is confident about the future of AI would say. If Zuckerberg truly believed that internal demand could consume all computing power, he would have no reason to allocate precious GPU resources for external competitors to use.”

What Meta still lacks to do cloud isn’t having GPUs—it’s the ability to sell

Having a GPU cluster doesn’t mean you can run cloud business.

Meta is missing things that can be listed as a checklist: enterprise-grade multi-tenant isolation architecture; security and compliance certifications (SOC 2, HIPAA, ISO 27001, etc.); fine-grained billing and an SLA guarantee system; global multi-region deployment and network access nodes; and, most importantly, an enterprise sales team and a customer success system.

From its founding to today, Meta has been a pure-to-consumer-to-C company. It has never sold anything to enterprise customers, and it has no muscle memory for B2B sales.

In its analysis, Datafloq made a judgment about Meta’s possible path: “Trying to build a full-stack cloud platform is a strategic mistake; the right approach is a narrow cut.”

The article listed four possible product formats. First, renting bare computing power—hourly pricing, no long-term contracts, and scheduling GPU clusters via API. Second, hosting Llama model inference—so enterprises can run Llama without building their own GPU infrastructure. Third, enterprise model fine-tuning services—fine-tuning open-source models using private data on Meta’s hardware. Fourth, agent infrastructure—providing dedicated tool-calling, credential management, and audit logs for AI agent workloads.

This means that Meta’s likely short-term cloud shape is probably “wholesale” computing power sales, targeting a small number of large customers, signing long-term contracts, and operating in a model similar to CoreWeave. It will not be like AWS, offering self-service registration, on-demand use, and a complete cloud platform with hundreds of services. The organizational capabilities and customer ecosystem required for that kind of full platform cannot be built up in just two or three years.

Meanwhile, on the same day, May 28, Meta also did two other things. It announced paid subscription tiers for Instagram, WhatsApp, and Facebook. And, according to The Information, it established a brand-new “Enterprise Solutions” department, sending engineers and product managers directly into large enterprise customers to help deploy AI tools.

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

Industry earthquake: Meta up 6%, CoreWeave and Nebius down 9%

After this news broke, Meta rose by more than 6%, while AI computing power leasing companies CoreWeave and Nebius both fell by more than 9%.

The magnitude of the drops in CoreWeave and Nebius indicates that the market believes this is a repricing of the moat of the entire neocloud business model.

The blow is threefold.

The first layer is a direct competitive threat. CoreWeave and Nebius’s business model is essentially “bulk buy GPUs → build clusters → mark up and sell to AI companies.” High gross margins depend on tight GPU supply and customers having few alternative options.

If Meta—the most aggressive spender on computing power—enters the game, it adds another player with a computing power supply on an enormous scale. Moreover, Meta’s GPU procurement costs are lower than those of neocloud companies because it directly signs strategic deals worth hundreds of billions of dollars with Nvidia and AMD, securing the best prices. Its selling price can be cheaper than CoreWeave’s while still being profitable.

The second layer is more lethal: identity conflict. One of CoreWeave’s largest customers is currently Meta. In April 2026, CoreWeave announced it would expand its AI infrastructure agreement with Meta, with a total value of $21 billion and a service term through 2032.

Now Meta wants to do the same thing itself. That’s equivalent to your buyer announcing it will become your competitor. The market’s natural reaction is to question whether that $21 billion contract will be renewed after it expires. Is Meta buying time—waiting for its own cloud business to be built, so it no longer needs CoreWeave?

The third layer is the collapse of the valuation narrative. In March 2025, when CoreWeave went public, the story it told was “explosive growth in AI computing demand, extremely scarce supply, and we are the scarce supplier.” That narrative supported its rocket-like growth from zero to a market value of hundreds of billions.

But Meta entering the business of selling computing power directly undermines the core premise of “supply scarcity.” If even the largest AI computing buyer in the world believes it might have excess computing power that needs to be sold externally, then is the market’s supply-demand relationship really as tight as it was previously portrayed?

This is not to say CoreWeave’s business will collapse immediately. Its Q1 2026 revenue is about $2.1 billion, with a massive backlog of contracts, so short-term revenue is secured. But the capital market prices expectations rather than the present reality. When your biggest customer is also your potential biggest competitor, the long-term growth story needs to be rewritten.

Good news or warning?

As for whether Meta doing cloud is truly good news, the answer depends.

Those who are bullish think this is an upgrade to Meta’s investment logic. Previously, Meta’s CapEx was a purely one-way bet—betting that AI would significantly boost ad revenue and user engagement, betting that if it won the returns would be enormous, and betting that if it lost it would be a sky-high sunk cost. Now that there is a cloud business route, the investment becomes “attackable and defensible”—it can go on the offensive while also serving as risk protection.

In Q1 2026, the global cloud infrastructure services market reached quarterly revenue of $128.6 billion (Synergy Research Group data), with annualized revenue exceeding $455 billion, and AI compute being the sub-segment with the fastest growth. Meta only needs to carve out a small slice to generate meaningful revenue. From the perspective of portfolio theory, this turns Meta’s CapEx from a “high-risk single bet” into a “hedged two-way option.”

Those who are bearish, however, believe this is precisely an “early warning signal” of an AI CapEx bubble. The logic is simple: If Meta truly believes internal AI demand can absorb all computing power and generate the corresponding returns, why would it allocate valuable GPU resources to external competitors? Taking the cloud step itself is hedging against the possibility that the pace of internal AI monetization will fall short of expectations.

The combined CapEx of the four giants in 2026 is about $725 billion, but the incremental revenue directly brought by AI may only be on the order of several tens of billions—an extreme mismatch between input and output. If Meta does cloud, it may be the most aggressive player, preparing first for the possibility of computing power oversupply.

There is also a technical concern. AI inference efficiency has improved rapidly over the past year, with unit inference costs cut every few months. If the speed at which efficiency improves continues to outpace demand growth, then the data centers built today may not be “not enough,” but “too much.” Meta doing cloud is essentially buying insurance against that possibility.

On the same day, U.S. stocks in the storage sector crashed. Micron and SanDisk, among others, all fell by around 10%. The core logic that drove these companies’ surging performance over the past year was “the AI data center construction boom driving explosive demand for HBM and enterprise SSDs.” In the prior quarter, Micron reported year-over-year revenue growth of 196%, and its story was “demand is infinite, but supply cannot keep up.”

But Meta’s news directly shakes the underlying assumption that “if you build it first, it won’t be enough.” If the future pace of data center construction by tech giants might slow down, then it means expectations for procurement growth of HBM and enterprise storage need to be revised downward.

This is also a story about the AI arms race entering its second half. Over the past two years, everyone has been competing over who is willing to spend the most, who can secure GPUs first, and whose data center scale is bigger.

But even Zuckerberg—the biggest spender—is “afraid.” “After building so much infrastructure, we need to ensure that whether AI monetizes quickly or slowly, we won’t lose everything.”

When the biggest buyer starts preparing to become a seller, who will still be the real buyer?