当 Meta 準備販售算力，AI 牛市的「鬼故事」要來了嗎？

What is the biggest fear for the AI bull market?

It's not that a particular company's model has temporarily fallen behind, nor that a specific generation of chips has underperformed expectations. It's that the market begins to doubt whether the capital expenditures of tech giants, which were considered the most certain variable over the past two years, can continue to grow forever.

On July 1st, according to Bloomberg, Meta is preparing a new cloud computing business, planning to sell its potential surplus AI computing power to external customers, while also considering offering managed model services similar to AWS Bedrock.

After the news broke, Meta's stock price rose over 10% during trading before closing up 8%, while CoreWeave and Nebius closed down 13% and 17% respectively. On the other side, selling pressure spread to AI hardware during the Asian trading session, with South Korea's KOSPI falling about 7% intraday, and both Samsung Electronics and SK Hynix dropping over 8%.

Overnight, Meta transformed from one of the most aggressive super-buyers in the computing market into a potential seller.

This sudden industry upheaval has also, for the first time, created a clear crack in a fundamental belief that has underpinned the entire AI bull run over the past two years: Does this mean that AI infrastructure has shifted from scarcity to surplus, and that the two-year-long computing power arms race among giants is about to hit an inflection point?

Or does Meta's move reveal another, more brutal reality: What the market is truly short of—is it GPUs, or the ability to turn GPUs into models, products, and revenue?

1. Everyone is short on computing power, but you, Meta, have too much?

Over the past two years, the most fundamental logic driving this AI rally can be summed up in two words: "Scarcity".

More accurately, it has been a structural bull market driven by a surge in demand, a shortage of supply, and a frenzy of capital expenditure expansion by tech giants.

Initially, the shortage was in high-end GPUs and advanced packaging capacity. Then the bottleneck spread outward—HBM, high-speed optical modules, and networking equipment became in short supply. This was followed by data center space, power capacity, gas turbines, electrical equipment, and high-density cooling. Today, the supply-demand tension has extended to general DRAM, NAND, enterprise SSDs, and even hard disk drives, which were once regarded as "legacy assets" by the market.

It can be said that the speculation across the entire AI industry chain over the past two years is like an ever-lengthening list of out-of-stock items, clearly demonstrating the 'barrel effect' and sector rotation. This also means that as long as model training and inference demand continues to grow, and new computing power capacity, electricity, and data centers cannot be delivered in time, each scarce link bottlenecked in between will have an opportunity to gain stronger pricing power. Conversely, upstream manufacturers can raise prices, secure long-term contracts, and have the incentive to continue expanding production.

For this reason, tracing further up the chain reveals that the true engine of this bull run is not just Nvidia, SK Hynix, or power equipment companies themselves, but precisely the ever-growing AI demand expectations and capital expenditures of tech giants like Microsoft, Meta, Amazon, and Google.

How much money the upstream giants are willing to spend determines how many GPUs, storage devices, and networking equipment they will buy, how many data centers they will build, and how much third-party cloud computing power and long-term electricity resources they will lock in, directly impacting the upper limit of the prosperity of the entire AI supply chain.

According to Bridgewater's estimates, the combined investments of Alphabet, Amazon, Microsoft, and Meta in expanding AI infrastructure in 2026 are projected to be around $650 billion, up nearly 60% from approximately $410 billion in 2025. Reuters, citing estimates from Goldman Sachs and Morgan Stanley in May, suggests that global AI-related capital expenditure encompassing data centers, power, equipment, and software could reach around $800 billion in 2026.

In a sense, this is a "super-sized" version of the food delivery wars in the AI world.

Among this, Meta, far from pulling back, actually stepped on the gas.

It had previously raised its 2026 capital expenditure guidance from $115-135 billion to $125-145 billion. As of the end of the first quarter this year, Meta held approximately $237.7 billion in non-cancellable contractual commitments (albeit contract obligations to be fulfilled over multiple future years), a significant portion related to servers, data centers, network infrastructure, and third-party cloud computing power.

So, strictly speaking, Meta's current consideration of selling some computing power externally does not mean it suddenly believes the entire industry is no longer short on computing power, nor does it signal an exit from the AI arms race. On the contrary, because data center construction cycles often span several years, Meta must prepare capacity in advance based on more aggressive demand scenarios. However, once the infrastructure is built ahead of time, internal model, product, and traffic demands may not perfectly align in the same period, leading to a temporary supply-demand mismatch.

To put it bluntly, Meta is building computing power on a large scale for future years, but at a time when its in-house models are temporarily lagging and internal products haven't fully taken off, some of the capacity that has come online may not be immediately absorbed. Rather than letting these expensive GPUs sit idle in data centers and depreciate continuously, it makes more sense to push them to the external market, improve utilization rates, and recover some costs.

Theoretically, Meta is not the first AI company to sell its self-built computing power. In May this year, xAI partnered with Anthropic, granting it access to its Colossus 1 supercomputing cluster, equipped with over 220,000 Nvidia GPUs, for a monthly fee of $1.25 billion.

The economic logic behind this isn't complicated: resources eventually flow to the enterprises that can best utilize their value. When a company cannot fully leverage its own computing power, and another company is willing to pay a high enough price, the most rational choice is not to let the GPUs gather dust but to rent them out for monetization.

However, Meta's symbolic significance far exceeds that of xAI.

Because Meta does not lack user entry points. Facebook, Instagram, WhatsApp, Messenger, and Threads together form one of the world's largest consumer internet product matrices. Theoretically, it should be one of the most capable enterprises of embedding AI models into existing products, creating a user flywheel, and absorbing computing power.

Yet, at least at this stage, Meta hasn't smoothly connected models, products, cloud services, and user entry points like Google has. This creates a rather paradoxical contradictory mirror—while Meta massively builds its own AI infrastructure, it still needs to procure external models and computing services like Gemini. Just a few days ago, reports indicated that Meta's demand for Gemini models and computing resources was so large that Google couldn't fully satisfy it, and some internal AI projects were even affected.

At first glance, this seems contradictory, but in essence, it's a mismatch between long-cycle supply and short-cycle demand. This is mainly because its current large model applications and real-time inference needs still rely on suppliers like Google, as its self-developed models cannot yet fully replace external solutions.

Therefore, it's not contradictory for Meta to simultaneously "procure external computing power" and "sell some of its own computing power." The real issue is whether the computing power it possesses can be matched with truly competitive models and products at the right time and in the right form. This means Meta was previously too optimistic about its own capabilities, built too much computing power, and now its own models/products cannot fully utilize it, leading to the need to sell the surplus.

2. Is the Shortage Computing Power, or Models/Products That Can Effectively Use It?

After Meta announced its intention to sell computing power, the market's reaction was quite telling.

Meta's stock rose over 10% intraday, eventually closing up 8%. On the other hand, CoreWeave and Nebius plunged 13% and 17%, respectively. The next day during Asian trading, selling pressure spread to AI hardware, with South Korea's KOSPI falling about 7% intraday, and both Samsung Electronics and SK Hynix dropping over 8%.

"Cloud drops, hardware drops, software rises" became the most intuitive market snapshot of the moment.

At first glance, this reaction seems perfectly logical:

• For Meta, it's definitely a short-term positive: Since its self-developed models and internal products cannot fully absorb all the computing power, renting out some resources or offering managed model services similar to AWS Bedrock allows Meta to recover some costs from infrastructure that would otherwise purely generate depreciation. This adds a safety net to hundreds of billions in capital expenditure. In the worst case, it can follow Apple's playbook, leverage its traffic, and partner with the best external models. After all, Zuck isn't a stranger to the "survival by cutting losses" strategy.
• For CoreWeave and Nebius, however, it's a bolt from the blue: Meta was a major client. Just in April, CoreWeave added about $21 billion to its long-term computing agreement with Meta, extending the contract term to 2032. Nebius's related agreements with Meta were valued at up to $27 billion. Yet, in the blink of an eye, the huge customer who was signing contracts across the table has moved its chair to the same side and is now competing with them for sub-tenancy business. This is, naturally, bad news.

The panic across the hardware supply chain stems from a deeper, more reasonable market extrapolation: If even a giant like Meta is starting to sell off computing power, doesn't that mean supply is nearing surplus? Are the giants about to cut their capital expenditures?

However, we must clarify a core fact: Meta's internal surplus of computing power does not equate to the entire tech industry hitting a ceiling on computing power. In fact, it could be a massive misinterpretation.

If we widen the lens to the ultra-long cycle of the next 3 to 5 years, we see that the expansion plans of major hyperscale cloud providers are still advancing at an almost insane compound growth curve. To more intuitively see the endgame of this arms race, let's do a quantitative comparison of the future computing power capacity of global key players.

First, look at Meta itself. By the end of 2025, external agencies estimate Meta's AI computing power to be roughly equivalent to 2 million to 2.5 million H100 GPUs (corresponding to about 2 GW). Based on its 2026 capital expenditure guidance, it will add 2-3 GW of computing power throughout the year, meaning by the end of 2026, Meta's total computing power base could reach around 5 GW.

5 GW sounds significant, but it immediately pales in comparison to the appetite of the entire industry. The market's true demand anchors are planning on a completely different scale:

• Google: In May, The Information reported a bombshell: Anthropic committed to spending a whopping $200 billion on Google Cloud for TPU computing power over the next 5 years. This portion alone corresponds to a computing power level of 5 GW. If we conservatively assume Anthropic accounts for 25% of Google Cloud's demand, then Google Cloud alone would be aiming for a total computing power of 20 GW by 2028, and Google overall could be looking at 25 GW.
• Amazon: Similarly backed by the 5 GW order from Anthropic and a 2 GW order from OpenAI, combined with their internal plan to double computing power capacity from 6.5 GW in 2025 by 2027, the projected overall demand is also around the 20 GW level.
• Microsoft: Tied to its $250 billion Azure contract with OpenAI, estimated using the same methodology, this corresponds to an exposure of roughly 20 GW. Not to mention OpenAI's own independent deployment plans like Stargate, its 10 GW deal with Nvidia, and its 10 GW deal with Broadcom (though far from realized). None of these are fully accounted for in the cloud providers' computing pool yet.

Putting these figures together, the conclusion is clear and somewhat harsh—even if Meta were to open up all of its 5 GW of computing power by the end of 2026 to the outside world, compared to the 10 GW, 20 GW, or larger new computing power plans over the next three years, it's just a drop in the bucket.

Zuck surely knows this too. The locomotive driving the industry's computing power construction has long been the ultra-large model demand parties like Google, Anthropic, and OpenAI. Whether Meta's models remain at the table or not doesn't affect the direction this train is heading.

Since the industry doesn't lack demand, why does Meta have surplus computing power? This reveals a deeply uncomfortable question: How can a company with billions of top-tier global traffic users not fully utilize its 5 GW of computing power? What the market truly lacks—is it computing power, or the models and products that can effectively use it?

From this perspective, we could even argue that Meta's computing power rental isn't necessarily a leading indicator of a surplus. Instead, it might lay bare the extremely hungry supply-demand state of the current computing power market:

Just look at the price xAI charged Anthropic for renting computing power: $1.25 billion per month for 500 MW of capacity. This translates to $30 billion per GW per year. This shows that even if a player temporarily 'leaves the table,' their idle computing power will be instantly snapped up by leading players with stronger models and shorter monetization paths.

Therefore, it's too early to conclude whether Meta's move is the first alarm bell for loosening computing power supply. What truly needs observation is whether this released computing power will be snatched up immediately, and whether the transaction price remains sufficiently high. If things proceed as expected, it might prove that AI computing power is still extremely tight.

At this point, the deeper logic behind "cloud drops, hardware drops, software rises" begins to truly surface. The market isn't trading on "computing power surplus"; instead, it's trading on the restructuring and migration of computing power's value within the industry chain.

3. What is the Real 'Horror Story' the Market Should Fear?

Here's a point that's easy to misinterpret: Meta selling computing power absolutely does not mean Zuck is completely giving up on the AI arms race.

On the contrary, the more Meta relies on external models like Google's and Anthropic's, the more its product ecosystem and high-margin advertising profits become constrained by others. The game between AWS and Anthropic already illustrates this point—once model companies truly command users and core demand, even cloud providers with vast infrastructure might be forced to cut into their profit margins to keep them.

Zuck couldn't possibly miss this point. Otherwise, why would he be restructuring his management team this year, aggressively releasing the closed-source MuseSpark model to build a moat, while simultaneously raising capital expenditure guidance again and continuing massive procurement and deployment?

Since Meta hasn't surrendered, why did the market experience such violent sector shifts? Because it has unveiled the beginning of a shift in the underlying pricing logic of the industry—this is the real 'horror story' the market should be afraid of.

As mentioned earlier, the valuation logic for the entire AI bull market over the past two years was: AI's Return on Investment (ROI) is uncertain, but AI's Capital Expenditure (CapEx) is absolutely certain. As long as the giants are still