Behind Meta's sale of computing power, has the inflection point of AI capital expenditure arrived?

Overnight, the core bull thesis for AI stocks was punctured by an insider.

On July 1, Meta's share price skyrocketed nearly 9% in a single day, adding hundreds of billions of dollars to its market value. Meanwhile, Micron and SanDisk plummeted more than 10%, while leading computing power rental firms including CoreWeave and Nebius saw their stock prices plunge 15% in a straight line.

Behind this sharply divided market performance lies a single piece of news: Meta is set to start selling excess computing power to external parties. What sounds like a simple story of "revitalizing idle assets" has unexpectedly sparked the biggest divergence in this year's AI market rally:

From the bulls' perspective, this is just a normal tiering between "old and new computing power" — renting out some older GPUs will not affect purchases of new chips. But from the bears' perspective, Meta entering the computing power sales market is proof of a computing power glut, and marks the inflection point for declining AI capital expenditure.

While this divergence remains unresolved, at least for now, the market has begun to reward cloud vendors that are "pulling back" on aggressive expansion.

Right after the news of selling computing power broke, Meta's share price surged sharply. In the month prior, the stock prices of the four major CSPs (Cloud Service Providers) all underperformed the Nasdaq index. Among them, Microsoft fell 17.15%, Amazon dropped 11.93%, Meta declined 10.86%, and even the best-performing Google posted a 6.18% loss.

Why has this divergence emerged? And how should we understand Meta's move to sell computing power? Today, Silicon-Based Observer will break this topic down for you.

The Market Begins to Reward Cloud Vendors That "Pull Back"

Behind this market performance of starkly contrasting fortunes is a shift in Wall Street's pricing logic for cloud vendors' capital expenditure.

In 2024, "computing power scarcity" was still the most powerful driving force in the market.

Google's Q1 financial report showed that capital expenditure surged 91% year-on-year to $12 billion, exceeding analysts' expectations of $10.3 billion, and its after-hours share price jumped nearly 12%.

Amazon's Q3 2024 financial report recorded capital expenditure of $22.6 billion, an 81% year-on-year surge that was roughly $3 billion higher than market expectations, driving its after-hours share price up more than 5%.

The prevailing logic back then could be summed up in four words: Spending is good news. But from the start of 2025, the tide gradually turned.

In July 2025, Google raised its full-year capital expenditure guidance to $85 billion — far exceeding market expectations — and its after-hours share price initially dipped briefly before turning positive, closing with a 2% gain. In October 2025, Microsoft posted quarterly capital expenditure as high as $34.9 billion, a 74.5% year-on-year surge that far outstripped market forecasts, yet its after-hours share price fell nearly 4%.

By 2026, the market's view on capital expenditure shifted completely.

At the end of April, Meta delivered strong results with $56.3 billion in revenue and a 61% year-on-year surge in net profit, but just because it raised the midpoint of its full-year capital expenditure guidance range by 8%, its after-hours share price plummeted nearly 7%.

Amazon CEO Andy Jassy mentioned during the February 2026 earnings call that 2026 capital expenditure was expected to reach as high as $200 billion, far exceeding Wall Street's consensus forecast of $146.6 billion. Right after the call ended, Amazon's after-hours share price crashed, plunging more than 11%.

Do you see the pattern? The year before last, expanding production drove share prices up; last year, differentiation began to appear; this year, even expansion exceeding expectations is no longer acceptable — as long as the market sees that "spending will continue", it will penalize valuations.

The valuation pressure brought by this pricing logic eventually evolved into collective weakness across the entire sector.

Especially in the most recent month, the stock prices of the four major CSPs all underperformed the Nasdaq. Among them, Microsoft fell 17.15%, Amazon dropped 11.93%, Meta declined 10.86%, and even the best-performing Google posted a 6.18% loss, while the Nasdaq Composite Index only fell 2.81% over the same period.

The market's anxiety is very straightforward: it is not afraid of companies spending money, but of spending more and more money without seeing an end to the payback period.

Precisely because of this, the news that Meta is renting out idle computing power has triggered such a strong market reaction. Just announcing that it will monetize its idle capacity externally has brought about a nearly 9% single-day gain and a market value increase of hundreds of billions of dollars.

In fact, this is not an isolated case for Meta. In reality, all the top players in AI capital expenditure have now quietly retreated from the first front of "stockpiling all available shovels" to the second front of "using shovels for themselves while also selling them to others".

Behind this collective shift among major tech companies lie two unavoidable practical pressures.

First, the scale of capital expenditure has grown so large that it is beginning to erode the company's cash flow security.

Amazon plans $200 billion in capital expenditure for 2026, and Morgan Stanley predicts its full-year free cash flow will be -$17 billion. Even for Meta, which has relatively solid cash flow, JPMorgan forecasts its 2026 free cash flow will reach -$4 billion.

Even a powerhouse like Google, which is set to post capital expenditure of $175–185 billion in 2026, saw Morgan Stanley cut its free cash flow per share forecast by 58%.

Among the four major CSPs, only Microsoft has managed to maintain positive free cash flow growth thanks to its diversified business portfolio, but even so, its capital expenditure has already consumed 63% of its operating cash flow.

Second, the overall rollout speed of AI applications is falling short of the market's initial high expectations, and internal use cases cannot absorb the rapidly expanding computing power capacity.

Take Meta as an example: although its open-source Llama series models have significant influence, none of its own AI commercial products have achieved breakout success. Looking across the entire industry, only two top vendors, OpenAI and Anthropic, have truly achieved large-scale commercialization. The hundreds of billions of dollars in computing power that CSPs have invested in have not been met with an internal demand boom of the same magnitude, making temporary idle capacity inevitable.

The Biggest Divergence Since the AI Wave Began

This debate over "Meta selling computing power" is not just a simple market fluctuation — it represents the first systematic shift in the underlying logic of the AI narrative.

Over the past two years, the market's only consensus was that computing power would always be scarce, and capital expenditure would only continue to rise. But now, this belief has developed its first crack.

First, industry management teams have begun to incorporate "overbuild" into their decision-making frameworks, which in itself is a structural change.

When Mark Zuckerberg mentioned in discussions about the computing power business that "if one day we feel we have overbuilt", it means that top-tier industry management has started to account for the risk of excess capacity in their expectations. This shift in expectations will directly transmit upstream: since there is a possibility of overbuilding, companies will no longer lock in production capacity at all costs, and the rigidity of orders will weaken first.

Analyst forecasts compiled by Bloomberg show that the capital expenditure growth rate of the five major cloud computing giants will see a notable slowdown in 2027, plummeting from a roughly 100% year-on-year growth rate in 2026 to 22%, before further falling to single digits. Even the most optimistic JPMorgan only raised its 2027 growth forecast to 50%, without denying the overall trend of slowing growth.

Second, computing power now has "secondary supply" capabilities. The bears' logic is simple: if idle computing power can circulate freely, the formula "new demand = new purchases" no longer holds.

Recently, major industry players have been systematically opening up this channel. xAI long-term leased its Colossus cluster of 220,000 GPUs to Anthropic for $1.25 billion per month, with a contract running to 2029 that totals over $40 billion; Meta has now also joined the track of renting out idle computing power externally.

Put simply, the computing power market now has an "existing stock supply pool". Going forward, when assessing supply and demand, we cannot only focus on new GPU production capacity — we must also account for the revitalization of existing stock resources.

From the bulls' perspective, Meta selling computing power does not mean the end of the AI computing power cycle. Their reasoning rests on two points:

First, there is no sign of any slowdown in the industry's expansion pace. By the end of 2025, Meta had in-hand computing power equivalent to roughly 2.5 million H100 GPUs, totaling around 2GW. Its 2026 full-year Capex midpoint is $135 billion, corresponding to 2–3GW of new computing power, which will bring its total in-hand capacity to roughly 5GW by year-end.

If we look at the entire industry, the scale of future expansion is even more staggering. Assume that over the next three years, all cloud vendors' computing power construction will be anchored to the demand of three model companies.

Huayuan Securities conducted a calculation: Google's $200 billion contract signed with Anthropic corresponds to 5GW of computing power. Assuming Anthropic accounts for 25% of GCP's demand, GCP alone will need 20GW of total computing power by 2028, putting Google's overall capacity target at 25GW.

Amazon is anchored to the 5GW contracted with Anthropic and 2GW contracted with OpenAI, coupled with its goal of doubling computing power capacity in 2027 compared to 2025 (6.5GW). Assuming AWS supplies 40% of its computing power to these two firms, this corresponds to roughly 18GW of demand, putting its overall target at 20GW.

Microsoft is anchored to its $250 billion Azure contract with OpenAI. Assuming OpenAI accounts for 25% of Microsoft's total demand, its overall target also reaches 20GW.

Add to that OpenAI's own Stargate project, its 10GW partnership with NVIDIA, and its 10GW partnership with Broadcom.

The conclusion is clear: the industry's total planned new computing power construction exceeds 85GW in the future. Even if Meta rented out all 5GW of its current in-hand capacity, it would only be a tiny fraction of this incremental market.

Second, only old and outdated capacity is being rented out, and new GPU purchases are completely independent of this.

Morgan Stanley's industry chain survey confirmed the computing power tiering rules: Meta only rents out previous-generation inference cards such as the H100 and A100, while 100% of its GB200, GB300, and new-generation Rubin training chips are kept in-house, all dedicated to iterative upgrades of the Llama large language model.

Just in June, Meta signed a new 1.6GW contract with Crusoe — signing the deal in June and announcing its computing power rental business in July. This very short time gap is the strongest evidence: it is renting out existing stock capacity, while purchasing new incremental capacity, with the two operations running in parallel.

While both sides have well-reasoned arguments, the current market debate over "whether Meta's move to sell computing power marks a cycle inflection point" cannot be resolved by a single event. At least based on current information, the market narrative is indeed shifting toward the view that "AI cloud vendors are marginally tightening their spending". Whether this judgment holds true will likely need to wait for cloud vendors' August financial reports to show clearer verification signals.

This article is from the WeChat public account "Silicon-Based Observer Pro", written by Yuanyuan, and republished with authorization from 36Kr.