As Silicon Valley's recognized cash-burning machine, Zuckerberg suddenly decided not to be a pure buyer of computing power.
With the exposure of a secret new business at Meta, hardware giants across Silicon Valley collectively broke out in a cold sweat. According to a Bloomberg report, Meta is preparing to publicly sell its internally "excess" AI computing power to external customers.
This move directly triggered a chain reaction of sell-offs in the AI infrastructure sector on July 2. Nvidia fell, TSMC fell, AMD fell, Micron fell, CoreWeave plunged.
The entire AI infrastructure sector collectively lost trillions in market value overnight. Meta itself, conversely, rose 8%.
In hindsight, this is just a very ordinary business decision. A company selling off idle resources. In any industry, such a thing wouldn't even make the news.
But in the AI industry, it almost equates to saying publicly: Here, I have some left.
The market instantly fell silent. For the past two years, the only sentence holding up the entire AI industry's valuation has been: Computing power will never be enough.
I. The AI Bull Market is a Russian Doll Story
To understand why the market reacted so strongly, you first need to understand how AI actually made money over the past two years.
To be precise, how it was valued.
The entire AI industry really has only one commercial closed loop:
AI needs GPUs → GPUs sell out → HBM sells out → PCBs sell out → switches sell out → optical modules sell out → electricity sells out → data centers sell out → keep buying GPUs
And so the cycle continues.
Capital markets have bet countless real gold and silver on this chain.
Thus, a very simple and crude formula formed in the market: AI growth = GPU growth = computing power forever scarce
Why did Nvidia once touch a $6 trillion market cap? Supporting this figure is the belief that everyone held: GPUs would forever be in short supply.
This assumption of extreme supply-demand mismatch is the core pricing model for the entire AI bull market over the past two years.
It has even become a recognized faith.
Relying on this faith, HBM turned from a cyclical stock into a growth stock; optical modules transformed from industrial goods into core AI assets, creating wealth myth after myth.
But the starting point for all these stories is the same sentence: Computing power is insufficient.
As long as this sentence holds true, every link in the chain has a story to tell.
Once this sentence is shaken, the entire chain must re-calculate.
Goldman Sachs' Delta One head, Rich Privotsky, put it bluntly in a recent internal meeting:
The market's core premise was that computing power was in a state of scarcity. If this premise is shaken, the hardware sector will be the first to bear the brunt.
He didn't say it *would* be shaken, he said *if*.
And Meta's news turned this *if* from a hypothesis into a question.
II. Why is Meta Selling Now?
Regarding Meta selling computing power, the most discussed point externally is making extra money.
This isn't wrong, but it doesn't get to the root.
Meta has at least three motives for doing this.
First layer, utilization.
Meta's capital expenditure for 2025 was already close to $70 billion, with guidance for 2026 continuing to rise. The market widely expects it to be in the hundreds of billions.
This money buys hundreds of thousands of GPUs, over a hundred megawatts of power, and teams of thousands for operations and maintenance.
The characteristic of large model development is its non-linearity. The training phase requires pouring all computing power to peak capacity, but once it enters intermittent periods like model fine-tuning, alignment, or waiting for evaluations, computing power demand plummets into a trough.
During these troughs, idle time is just depreciation.
Large numbers of GPUs idling in data centers consume expensive depreciation costs and baseline power every second. Rather than letting these graphics cards gather dust in warehouses, it's better to rent them out.
Second layer, strategic choice.
This is the key.
Over the past two years, Silicon Valley giants' paths in AI have already begun to diverge.
OpenAI sells APIs, Anthropic sells APIs, Google sells APIs + models. Microsoft sells APIs + cloud.
Meta? Llama has no barriers, the model code is fully open-source, the Agent framework doesn't charge extra either, hardly making any money from APIs.
The outside world has always interpreted this as open-source idealism.
Zuckerberg is not a philanthropist. He just doesn't want to become OpenAI; he wants to become AWS.
In the Facebook era, Meta made money through social networks.
In the AI era, Meta's bet is:
It's okay if models don't make money; computing power makes money.
Opening Llama's doors wide is to draw developers into Meta's ecosystem. Not charging for the Agent framework is to let AI teams run their tasks on Meta's cloud.
What it ultimately wants to earn is that underlying compute money.
Looking back at internet history, when Amazon first created AWS, it wasn't to make extra money either. Bezos just wanted to rent out the excess capacity of Amazon's e-commerce servers, which were built to handle "Black Friday."
As a result, AWS later became Amazon's most profitable division, with gross margins far higher than retail.
Meta is betting on the same story.
Third layer, redefining AI infrastructure.
This is the deeper layer.
In the past, mentioning AI infrastructure, the first reaction was GPUs. But what Meta is launching now is essentially a full suite of services: GPU + training framework + open-source models + inference optimization + Meta's cloud.
Not selling a single card, but selling an "AI factory."
If we follow this logic further, what Meta wants to capture isn't the GPU market, but the AI cloud market. GPUs are just the entry point.
When AWS sold EC2, it didn't position itself as selling servers either; it sold not having to buy your own servers.
What Meta is doing today is essentially the same thing, just with the target audience changed from e-commerce customers to AI teams.
III. What is Capital Afraid Of?
Capital wouldn't be afraid of Meta selling a few more cards by itself.
What truly unsettles capital is the signal behind Meta's statement—it might be valid.
Meta is saying, "I have computing power to sell."
The actual signal transmitted is that GPUs can be shared. If GPUs can be shared, then the entire demand model built over the past two years must be recalculated.
The old logic was: either buy cards yourself, build your own data center, or hoard supplies in advance.
Each new AI company meant another GPU order.
Under this highly anxious hoarding logic, computing power became moats that major companies built defensively against each other.
But what if you can rent? What if a giant, holding hundreds of thousands of GPUs, rents them out during idle times?
New AI entrants wouldn't necessarily need to buy their own cards; they could simply call them directly.
Then the demand model for GPUs changes, from linear growth following the number of companies to dynamic adjustment following actual network-wide usage.
These are two completely different pricing logics.
It's comparable to a piece of history. Everyone used to buy servers, then AWS came along.
Now nobody buys their own servers; Capex (capital expenditure) turned into Opex (operating expense), heavy assets turned into subscriptions.
What Meta wants to do today is replay this story in the AI era.
The market reacted most violently towards cloud infrastructure tech company CoreWeave. CoreWeave's business model can be explained in one sentence: I built GPU clusters and rent them to AI companies.
But CoreWeave's commercial foundation was built during the vacuum period when giants used their own computing power and the market was severely short of cards. Once Meta really starts doing this, CoreWeave's role becomes awkward.
In terms of scale, CoreWeave lacks Meta's scale advantage. In terms of stack completeness, CoreWeave lacks Meta's models and software.
When enterprise clients can directly run native Llama models on Meta's cloud and enjoy the underlying framework optimized by Meta's engineers, they have no reason to pay a premium to sublet CoreWeave's pure bare-metal servers.
The battlefield CoreWeave lost has already moved up from the layer of selling computing power to the layer of selling complete AI services.
This is not competition on the same level.
Capital saw this layer, hence the fiercest sell-off.
IV. Why Did Prices Rebound the Next Day?
The stocks that fell most sharply mostly rebounded the next day.
Many interpretations attribute it to market correction.
But I believe the real reason is that capital conducted a re-pricing within 24 hours.
Everyone figured out one thing: What Meta said is true, but not today.
The current reality is that the Scaling Law for large models hasn't hit its ceiling yet. The main clusters of major companies are still running non-stop, training the next generation of multimodal models.
Inference demand is still growing rapidly year-on-year.
The trough computing power Meta can release at this stage is still only a very limited supplement compared to the vast throughput of the entire network.
Short-term, the sentiment release is over. Medium-term, the capital expenditure story for AI remains unchanged. Long-term, the only thing shaken is one thing: The supply-demand relationship for GPUs might slowly change from absolute shortage to structural surplus.
This change won't happen overnight; it might take two years, maybe three.
But the direction has changed.
With one statement, Meta pushed the entire market's bull run from being driven by shortage to being driven by efficiency.
And the valuation logic for an efficiency-driven bull market is a different set. Nvidia's market cap was driven by shortage. OpenAI's high valuation was driven by shortage.
Most companies on the entire AI infra chain were driven by shortage.
Once the market starts re-pricing based on efficiency, the valuation anchors for the entire sector must be adjusted. Meta just pushed this adjustment's timetable forward.
V. What's Really Changing Isn't GPUs
Looking back, the story of the AI industry over the past two years was simple.
GPUs were in short supply.
Just these few simple words were enough to support trillions in market cap.
But at this stage, the marginal effect of this logic has already begun to diminish sharply.
GPUs are still in short supply, but "shortage" can no longer fully explain valuations.
Moving forward, the AI industry has only one new question to answer: How to utilize the GPUs that have already been bought.
These are two completely different eras.
One era competes on procurement; whoever has more money, more cards, has a higher valuation.
One era competes on utilization; whoever has lower TCO (Total Cost of Ownership), lower PUE (Power Usage Effectiveness), cheaper per-Token inference costs, survives in the market.
One era is driven by Capex (capital expenditure), one era is driven by ROA (Return on Assets).
If ChatGPT changed AI, this move by Meta is the first to begin changing *how* AI makes money.
Words from 【Beyond the Headlines】:
The true inflection point of an industry is never when demand first appears. It's when someone first starts discussing how supply can make money.
For the past two years, the whole world has been asking one question: How many more GPUs does AI need?
With this move, Meta asked another question for the first time: Why can't the GPUs already bought start making money?
This is the same industry, two different eras.
The earlier era is called shortage; the later era is called operations.
Zuckerberg's move is essentially a starting gun.
What he's fighting for is never about how many cards he can sell, but about making everyone realize that the AI industry's红利期 (dividend period) of simply buying and hoarding cards to sit on soaring valuations is already over.
This article is from the WeChat public account "Beyond the Headlines," author: Hua Hua
