Author: Chao Xiang Research

On July 1, Bloomberg broke a major piece of news: Meta is setting up a cloud computing business unit and is preparing to sell its excess AI computing capacity to external customers.

After the news went out, the market reaction was immediate, but the divergence was extremely severe. Meta itself surged more than 8% in pre-market trading, while the two banner companies in the "new cloud computing" sector—CoreWeave and Nebius—plunged 6% and 10%, respectively. Amazon also reversed into declines in pre-market trading. On one side there was celebration, on the other panic—this news was like a scalpel, precisely cutting along the line that separates interests across the AI compute power industry chain.

The most ironic part is this: Meta recently signed a $21 billion compute procurement agreement with CoreWeave, and also signed a partnership worth as much as $27 billion with Nebius. Now it's turning around to compete with its own suppliers for business.

What exactly is going on here?

According to a Bloomberg report citing people familiar with the matter, a department inside Meta called "Meta Compute" is spearheading this effort. The department is co-led by three executives: Santosh Janardhan, head of Meta infrastructure; Daniel Gross of Meta's superintelligence lab; and Dina Powell McCormick, president of Meta.

There are two paths for the business models currently being considered:

The first is "Model-as-a-Service" (Model-as-a-Service), letting external developers pay to call AI models hosted on Meta's infrastructure, including Meta's in-house Muse Spark model. The benchmark product for this path is AWS's Bedrock service—in essence, it takes the AI model inference capabilities Meta built at great expense and opens them up as a paid API.

The second one is even more aggressive: renting out bare GPU compute power directly. This is what CoreWeave and Nebius are doing—and it is this path that caused the stocks of these two Neocloud companies to collapse instantly. When your biggest customer announces it's going to do exactly the same business as you, your investors will likely run for cover first.

Meta has not yet issued an official comment on this matter.

A $145 billion compute bet needs an "insurance policy"

To understand the core logic of this news, you need to look at a set of numbers first.

In 2026, Meta's guidance for AI-related capital expenditures is $125 billion to $145 billion, up again from the prior guidance of $115 billion to $135 billion. What does this mean? It's only slightly below Google's parent Alphabet's $175 billion to $185 billion, Microsoft's $190 billion, and Amazon's $200 billion. The four tech giants' combined spending on AI infrastructure this year will exceed $700 billion.

More importantly, Meta has a unique position. Among these four companies, Amazon has AWS, Microsoft has Azure, and Google has Google Cloud—they all have mature cloud computing businesses to absorb AI infrastructure investment and can sell compute capacity to customers directly. Only Meta doesn't. All its data centers and GPU clusters are, in theory, only for its own social platforms, advertising systems, and AI R&D services.

This creates a huge risk exposure: if Meta's internal demand for AI compute doesn't grow as expected, the $145 billion in capital expenditures will become sunk costs. Data centers are built, GPUs are purchased, and long-term power contracts are signed—these are "fixed" investments that can't be cut anytime like adjusting a marketing budget.

At the May 27 shareholders meeting, Zuckerberg addressed this concern directly. He said the cloud computing business is "definitely on the table," and revealed that "almost every week, external companies come to us—either asking whether we can open up API services, or asking whether we can sell them compute capacity at a higher price."

Translate Zuckerberg's implied message: We don't mind spending this much money. If all the AI compute gets used, the return on this investment will show up through our products. If there is excess capacity, we can also sell it out to make money. Either way, we don't lose.

Wall Street's biggest worry has always been: "Meta spends too much on AI but doesn't see returns." A cloud computing business gives investors a cushion: the $145 billion in capital expenditures is no longer purely risky spending—it becomes a double bet that can be attacked and also defended.

Neocloud's survival crisis

But Meta investors' frenzy is a nightmare for CoreWeave and Nebius.

To understand this relationship, start with a key fact: the entire business model of Neocloud companies is to manage GPU compute on behalf of tech companies that don't do cloud computing themselves. CoreWeave and Nebius are able to land those billion-dollar contracts because the core logic is: "Meta has massive AI compute demand, but it doesn't have a cloud business to absorb excess capacity, so it needs to rent from outside."

Now Meta says I'm ready to do it myself.

This shock is structural. CoreWeave currently has nearly $100 billion in revenue backlog orders, a large portion of which comes from Meta and other large AI companies. Nebius' $27 billion contract with Meta includes reserving $12 billion worth of GPU capacity for Meta starting in early 2027. If Meta's own buildout begins to replace external leasing, the conversion rate of these orders would be thrown into question.

A deeper issue is that Neocloud companies are already in a fragile position along the AI industry chain. What they provide is essentially a "compute intermediary" service: buying GPUs from Nvidia, building data centers, then marking them up and selling them to AI companies. This model can generate outsized profits when GPU supply is tight, but when supply bottlenecks ease and major customers begin building in-house, the value of the intermediary gets rapidly squeezed.

CoreWeave's Q1 revenue this year was $2.078 billion, up 168% year over year, but it posted a net loss of $740 million—its losses doubled on a year-over-year basis. Its total debt has exceeded $25 billion. Nebius' Q1 revenue was $399 million, up a whopping 684% year over year and performed impressively, but it also lost more than $100 million, with total debt exceeding $9.5 billion. Both companies are using high leverage to buy high growth. If the biggest customer turns into a competitor, this "borrow money to expand capacity" model becomes especially dangerous.

The market has already been voting with its feet. In June, CoreWeave's short interest ratio reached 14%, and Nebius was even higher at 20%. Investors' confidence in the Neocloud space is wavering.

The traditional three cloud giants also can't afford to be complacent

Meta's entry into cloud computing is bad news for AWS, Azure, and Google Cloud as well.

The global cloud infrastructure market reached a quarterly scale of $129 billion in Q1 2026, with annualized growth of 35%, as it moves toward annual revenue of $500 billion. This market is long dominated by AWS, Azure, and Google Cloud, which together hold more than 60% market share.

If Meta officially enters the arena, it would break this triopoly landscape. And Meta has several unique advantages: it operates one of the world's largest social network platforms, with vast real-world experience across AI model and application use cases; it has built a strong developer ecosystem in open-source AI (Llama series models); and its scale of AI infrastructure investment is already approaching that of the three major cloud giants.

Of course, cloud computing is not just hardware. AWS dominates the market not only because it has data centers, but because it spent nearly 20 years building a whole product ecosystem: from compute, storage, and databases to machine learning, security, and the Internet of Things—more than 200 services in total. For Meta to build an equivalent product matrix from scratch, it needs to invest enormous engineering resources and time.

But Meta's strategy may not be a full replication of AWS. A more realistic path is to focus on the hottest, fastest-growing niche: AI compute. If Meta only does those two things—"AI computing + model services"—its entry barriers would be much lower, and it would be entering precisely the most profitable part of the cloud market.

Seeking Alpha noted that after a Bloomberg report circulated, Amazon's stock price flipped from rising to falling in pre-market trading. AWS is Amazon's most profitable business: Q1 2026 cloud revenue grew 28% year over year, the fastest pace in 15 quarters. Any new entrant that slices up the cloud market will make AWS investors nervous.

You can also sell "leftover" worth $600 billion: the deep logic behind Meta's compute strategy

There is one detail worth savoring again and again.

In January this year, Meta announced a "Meta Compute" plan, aiming to accumulate "tens of GW" of compute capacity over this decade, and in the long term to look toward "hundreds of GW or even more." Meta currently operates more than 30 data centers, and the AI-optimized capacity being built ranges from 1GW to 5GW. In June, it also signed a compute procurement contract for 1.6GW with data center company Crusoe.

What do these numbers add up to? Meta is building AI infrastructure on the scale of a "supercomputing nation."

This also brings in another broader backdrop: in 2026, the real bottleneck in the AI industry is not chips, nor money, but electricity. Just a few days ago, reports said Google can't provide enough Gemini compute power, so it had to limit Meta's access to its models. Google Cloud itself has more than $460 billion in signed contracts that have not yet been delivered. Even the richest tech companies on Earth can't get enough compute—not because they lack money or chips, but because they lack power.

Against this backdrop, Meta doing both as a buyer and a seller has another layer of strategic meaning: whoever locks in power and data center capacity first will have a structural advantage in the AI race. The infrastructure Meta built with $145 billion is both a weapon to catch up in superintelligence and a "strategic reserve" that can be monetized externally.

Several key judgments

The market impact of this news will continue to percolate over the coming months, and there are several dimensions worth关注:

For Meta itself, if a cloud computing business becomes operational, it would open up an entirely new source of revenue and reduce overreliance on advertising. Currently, more than 99% of Meta's revenue comes from ads. Even if cloud business initially accounts for only a few percentage points of revenue, its symbolic significance and valuation impact are not to be underestimated.

Uncertainty for the Neocloud track has risen significantly. The investment logic for CoreWeave and Nebius rests on the premise that "big tech companies need to rent external GPU compute." If Meta's own cloud business succeeds, other tech giants may follow suit, shrinking the long-term survival space for Neocloud. Of course, in the short term, the gap between supply and demand for AI compute is still huge, and the signed contracts already provide some revenue certainty. But valuations need more margin of safety.

The bigger picture is that the AI industry is moving from the stage of "burning money like crazy to build infrastructure" into the stage of "making infrastructure investments pay off." Behind seemingly unrelated events—Meta selling compute capacity, Open Standard issuing OUSD, and major banks rolling out stablecoins—lies the same logic: when the scale of investment reaches a certain level, capital will look for every possible path to monetize itself.

For the AI arms race, Meta's move is essentially telling the whole world: $145 billion isn't a gamble—it's infrastructure investment. The characteristic of infrastructure is that once it's built, you can charge everyone for it.

Disclaimer: This article is for informational purposes only and does not constitute investment advice. Technology stocks and related investments carry a high level of risk—readers should make their own judgments.