Don't Misread Meta Selling Compute Alone as Industry-Wide AI Compute Oversupply
TechFlow Selected TechFlow Selected
Don't Misread Meta Selling Compute Alone as Industry-Wide AI Compute Oversupply
Inference business is highly profitable and demand is surging; Meta hopes to monetize periodic surplus capacity and improve its income statement in the process.
Navigating Web3 tides with focused insightsAuthor: US Research Munger Jun
Yesterday the entire market was collectively bullish on AI infrastructure, yet today optical modules, storage, and equipment were slashed by 7%~15% all at once.
The trigger was an unverified rumor stating "Meta Plans to Sell Compute Power."
In this article, I will deeply analyze why this logic chain is incorrect, as well as two major loopholes in most follow-up analyses, followed by a counterargument and my conclusion.
My view is direct: The market is trading on a stupid syllogism today. Believe anything but unverified rumors.
The most erroneous part of today's trading lies in extrapolating one company's operational decision directly into supply-demand signals for the entire industry.
Let's break it down layer by layer below.
1️⃣ The Cause of the Incident: An Unverified Rumor That "Meta Plans to Sell Compute Power"
Bloomberg broke the wind first: Meta is planning a cloud business that might sell "excess AI compute power" externally, or even sell raw compute directly. Reuters subsequently added a key statement: the plan is still in the development stage, strategies may change, Meta declined to comment, and the news has not been independently verified.
Based on just this unconfirmed rumor, the market immediately deduced a logic chain and began amplifying public sentiment among those who do not understand the AI compute industry:
Meta wants to sell compute → Meta can't use it all itself → AI training/inference demand is weak → AI compute oversupply → Scarce rents originally belonging to Nebius, CoreWeave will be taken away → AI hardware chain should fall.
Thus, almost all visible AI infrastructure stocks are falling.
Why do I think this is wrong? Read on.
2️⃣ The Market Is Trading on a Stupid Syllogism
What the market is trading now is a logic chain that has been continuously sensationalized:
Meta wants to sell compute = Meta can't use it all = AI is failing = Compute oversupply = The entire industry chain should fall.
But the problem lies in this: whether Meta sells compute is a matter of how one company, Meta, accounts for its balance sheet; whereas "whether the world actually lacks AI compute" is a completely different question.
Next, I will dissect the two biggest loopholes.
3️⃣ Loophole One: Meta Selling Compute ≠ AI Compute Oversupply, Quite the Opposite
If Meta were really selling compute because of "compute oversupply and AI failing," then the following actions would not make sense at all.
In April, Meta continued to raise its 2026 CapEx guidance to $125 billion~$145 billion. Which company would continue to increase capital expenditure when judging that demand is about to collapse?
In April, Meta signed an AI cloud agreement with CoreWeave worth about $21 billion, locked until 2032, including Vera Rubin GPU capacity; in March, it signed a multi-year agreement with Nebius worth up to about $27 billion. It has been only three months since then, has AI suddenly become oversupply? Meta's management is not that fickle, moreover, large companies' budgets are usually set one year in advance.
Not long ago, there were reports that Google limited Meta's quota for using Gemini, precisely because Meta's demand had grown so high that it exceeded the capacity Google could provide. Demand is already so high that others cannot supply it, and now you tell me it has compute oversupply? (I have already interpreted this view in my previous posts.)
4️⃣ Loophole Two: Even If the Neocloud Logic Holds, Why Kill the "Pick-and-Shovel Sellers" Too?
For CRWV, NBIS, I can barely understand the market's concerns: if Meta transforms from a super customer into a potential competitor, this is indeed a question worth discussing, provided you truly believe it will do so.
But extending this logic to optical modules, interconnects, storage, and equipment makes absolutely no sense.
These companies are essentially "pick-and-shovel sellers" in the AI infrastructure industry chain; they have almost no direct relationship with whether Meta sells compute.
The larger the AI cluster, the more high-speed interconnects, optical modules, and lasers are needed. Meta subletting part of its compute power will not make these physical demands disappear, nor will it reduce the fiber optic cables needed for the entire cluster by a single strand.
Storage is the same.
Equipment manufacturers (AMAT, etc.) are even more typical; they are located at the very bottom of the entire industry chain. Whether Meta sells compute and whether wafer fabs continue to purchase etching, deposition, CMP, and other equipment are worlds apart; there is simply no direct transmission mechanism.
5️⃣ The Biggest Flaw: Why Did Oracle Hardly Fall?
This point actually explains a lot.
If the market truly believed "AI is finished" or "compute oversupply," then Oracle, as an AI cloud player holding massive GPU cloud resources, with negative FY free cash flow, and CapEx guidance already pulled to nearly tens of billions of dollars, should have fallen much harder.
The result, however, was Oracle remained stable, Neocloud collapsed, and pick-and-shovel sellers fell along with them.
This combination itself indicates that what the market is killing today is not demand; "compute oversupply" is more like a fig leaf.
6️⃣ The Real Problem Is Supply-Demand Mismatch, Not Supply Surplus
This is the point I consider most core to the whole incident.
The entire market being in short supply and a certain company, at a certain point in time, in a certain region, or with a certain generation of GPU experiencing a temporary surplus are not contradictory at all.
GPU, electricity, server rooms, and networks are inherently built ahead of cycle by several years; whereas model training is project-based, API release times may be delayed, and inference traffic grows non-linearly; the two sides naturally will not be fully synchronized.
Thus, the situation arises: strategically long-term there is still a lack of compute, but short-term financially there are gaps in local resources.
This is called Mismatch (Supply-Demand Mismatch), not Oversupply (Supply Surplus).
Conclusion
Meta is not selling NVIDIA because "AI is failing."
Quite the opposite, a more reasonable explanation is: inference business is too profitable, demand is too vigorous, Meta hopes to monetize temporary surplus capacity, and improve the income statement along the way.
The market, however, misinterpreted a company-level resource optimization as the entire AI industry having entered supply surplus.
And I believe this is precisely the most core and most easily overlooked misjudgment in today's collective decline of AI infrastructure.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
@Kay2289123
