Don't mistake Meta's selling of computing power as AI computing power surplus across the entire industry.

Author: Munger of Meiyan

Yesterday, the entire market was still collectively bullish on AI infrastructure, and today, optical modules, storage, and equipment were slashed by 7% to 15%.

The trigger was a rumor that "Meta is selling computing power."

In this article, I'll dive deep into why this logical chain is wrong, as well as two major flaws in most copycat analyses, and finally a counterargument and my conclusion.

My view is straightforward: the market is trading a stupid syllogism today. Believe nothing, especially not rumors.

The biggest mistake in today's trading is extrapolating one company's operational decision into a supply-demand signal for the entire industry.

Let's break it down layer by layer.

1️⃣ The Trigger: A Rumor That "Meta Is Selling Computing Power"

Bloomberg first leaked: Meta is planning a cloud business that may sell "excess AI computing power" to external parties, or even raw compute directly. Reuters then added a key phrase: the plan is still in development, the strategy may change, Meta declined to comment, and the news has not been independently verified.

With this unconfirmed rumor, the market immediately spun out a logical chain and began amplifying public sentiment from those who don't understand the AI computing industry:

Meta is selling computing power → Meta doesn't need it all → AI training/inference demand is weak → AI computing power is oversupplied → the scarcity rents originally belonging to Nebius and CoreWeave will be snatched → the AI hardware chain should fall.

As a result, almost every AI infrastructure stock you can see is down.

Why do I think this is wrong? Read on.

2️⃣ The Market Is Trading a Stupid Syllogism

The market is now trading a logical chain that has been constantly hyped:

Meta sells computing power = Meta doesn't need it all = AI is failing = oversupply of computing power = the entire industry chain should fall.

But the issue is: whether Meta sells computing power is a question of how Meta, as a single company, balances its own books. Meanwhile, "whether the world's AI computing power is actually scarce" is a completely different question.

Next, I'll dissect the two biggest flaws.

3️⃣ Flaw One: Meta Selling Computing Power ≠ AI Oversupply; Quite the Opposite

If Meta were really selling computing power because of "oversupply and AI's failure," then the following actions would make no sense.

In April, Meta raised its 2026 CapEx guidance to $125 billion–$145 billion. Which company would keep increasing capital expenditure when it believes demand is about to collapse?

In April, Meta signed an approximately $21 billion AI cloud agreement with CoreWeave, locked through 2032, including Vera Rubin GPU capacity. In March, it signed a multi-year deal with Nebius worth up to approximately $27 billion. That's just three months ago—has AI suddenly become oversupplied? Meta's management isn't that fickle, especially since large companies typically set budgets a year in advance.

Not long ago, there were reports that Google had restricted Meta's usage of Gemini precisely because Meta's demand was so high that it exceeded the capacity Google could provide. Demand is already so high that others can't supply enough, and now you're telling me Meta has excess computing power? (I've explained this point in my previous post.)

4️⃣ Flaw Two: Even If the Neocloud Logic Holds, Why Take Down the "Pick-and-Shovel" Sellers Together?

For CRWV and NBIS, I can barely understand the market's concern: if Meta goes from a super-customer to a potential competitor, that's indeed a topic worth discussing—provided you truly believe it will do so.

But extending this logic to optical modules, interconnects, storage, and equipment becomes completely untenable.

These companies are essentially the "pick-and-shovel sellers" in the AI infrastructure supply chain, and they have almost no direct relationship with whether Meta sells computing power.

The bigger the AI cluster, the more it requires high-speed interconnects, optical modules, and lasers. If Meta subleases part of its computing power, it won't make these physical requirements disappear, nor will it cause the entire cluster to use one less fiber optic cable.

The same goes for storage.

Equipment manufacturers (AMAT, etc.) are even more typical—they sit at the very bottom of the entire supply chain. Whether Meta sells computing power or not, and whether wafer fabs continue to procure etching, deposition, CMP, and other equipment, there is a vast gap between them; there is simply no direct transmission mechanism.

5️⃣ The Biggest Flaw: Why Did Oracle Barely Fall?

This point alone already says a lot.

If the market truly believed "AI is over" and "computing power is oversupplied," then Oracle—an AI cloud player holding massive GPU cloud resources, with negative free cash flow for FY and CapEx guidance already pulled to nearly $100 billion—should have fallen even harder.

Instead, Oracle held steady, Neocloud crashed, and the pick-and-shovel sellers declined along with them.

This combination itself shows that today's market is not targeting demand; "oversupply" looks more like a fig leaf.

6️⃣ The Real Problem Is Supply-Demand Mismatch, Not Oversupply

This is what I consider the core of the entire matter.

The overall market being in short supply is not at all contradictory to a particular company, at a particular time, in a particular region, or for a particular generation of GPU experiencing temporary surplus.

GPUs, electricity, data centers, and networks are all built ahead of time over multi-year cycles. Meanwhile, model training is project-based, API release dates may be delayed, and inference traffic grows non-linearly. The two sides are naturally not fully synchronized.

Thus, you get: strategically, there is still a shortage of computing power in the long run, but financially, there is a temporary resource gap in the short term.

This is called Mismatch, not Oversupply.

Conclusion

Meta is not selling NVIDIA because "AI is failing."

Quite the opposite: the more reasonable explanation is that inference business is too profitable and demand is too strong. Meta wants to monetize its temporary spare capacity and, incidentally, improve its profit statement.

The market, however, misinterprets a corporate-level resource optimization as the entire AI industry entering oversupply.

And in my opinion, this is the most critical and easily overlooked misjudgment in today's collective AI infrastructure sell-off.