Has the computing power infrastructure hit a sudden brake?

Recently, the global AI infrastructure landscape has undergone dramatic shifts. First, Meta announced plans to lease out some of its idle AI computing power, triggering sharp volatility in the capital market. Shortly afterward, SoftBank Group officially established a new company, SB Neo, to aggressively enter the U.S. computing power rental market. Meanwhile, established private equity firm Blackstone suddenly halted its planned global largest data center project with an investment of over $100 billion, and Microsoft also abandoned a $3 billion cloud computing power lease agreement with Oracle due to security concerns.

On one hand, tech giants are stepping into the "computing power sales" arena; on the other hand, hundred-billion-level infrastructure projects are hitting emergency brakes. This series of seemingly contradictory moves has left the market wondering: Is there already a surplus in AI computing power? Is the investment bubble in computing infrastructure about to burst?

An in-depth analysis from the semiconductor industry perspective reveals that what we are seeing is not a simple "computing power surplus," but a profound restructuring of the logic behind the AI industry's development. The era of unconstrained, reckless "growth at all costs" over the past two years is drawing to a close, and competition in AI infrastructure is fully entering a new phase where "efficiency is king."

Giants Choosing to Monetize Computing Power?

In July 2026, Meta announced plans to launch a cloud infrastructure business, selling AI computing power and model access rights to external customers. Right after the news broke, Meta's stock price surged nearly 9% in a single day, with its market cap jumping by approximately $127 billion. However, the entire AI computing power industry chain came under pressure—shares of emerging computing power rental players such as CoreWeave and Nebius plummeted by over 13%, while storage chip giants Micron, SK Hynix, and Samsung Electronics all recorded significant declines.

The market's immediate reaction was panic: If even Meta cannot fully utilize its own GPUs, that must mean there is a computing power surplus.

Yet this linear thinking overlooks the unique nature of AI computing power assets and the true intentions of the giants. Meta's leasing move is essentially an upgrade in asset operation efficiency, not a signal that demand has peaked.

In 2026, Meta's capital expenditure guidance reached as high as $125–145 billion, with the vast majority allocated to data centers and GPU procurement. To date, Meta has committed a total of $183 billion to AI infrastructure investments. As a company that derives 98% of its revenue from advertising, hundreds of billions of dollars in annual investment has been accumulated into massive computing clusters, yet the open-source Llama model itself does not generate direct revenue. Monetizing previous-generation computing power or temporarily idle resources externally can not only directly dilute depreciation and operation and maintenance costs, but also serves as a key step in transforming the GPU cluster from a "pure cost center" to a "revenue-generating asset." Morgan Stanley estimates that if Meta leases 250MW of computing power for one year, it could generate approximately $10 billion in revenue.

This is not an original idea from Meta. Earlier, Elon Musk's xAI had already successfully leased out the computing power of its Colossus supercomputing cluster on a massive scale. According to multiple media reports, Anthropic has leased the full capacity of Colossus 1—roughly 220,000 Nvidia GPUs—paying as much as $1.25 billion per month in rent under a contract running until May 2029, with a total value of about $40 billion. Google is also paying $920 million per month to lease transitional computing power to make up for delays in its own data center construction. These two transactions alone have brought SpaceX a monthly cash flow of over $2.1 billion. According to institutional calculations, based on this monthly rent level, the implied return on investment would allow full recovery of all capital expenditures in roughly two years.

SoftBank Group's entry further confirms the appeal of this sector. On July 2, SoftBank announced the establishment of SB Neo, which plans to launch cloud services based on Nvidia's latest GPUs for U.S. enterprises in the 2027 fiscal year. Its goal is to build 10GW of AI data center infrastructure, with an initial deployment at an 800MW site in Ohio. To support this expansion, SoftBank is securing a $10 billion loan using its stake in OpenAI as collateral.

From Meta to xAI and then to SoftBank, the fact that giants are transforming into "computing power landlords" is not because they no longer need computing power, but because amid high capital expenditures for computing power, they must find new paths for return on investment. As Tianfeng Securities pointed out: "Meta entering the AI cloud business does not equate to admitting that there is a general GPU surplus. This is not the end of AI capital expenditure transactions, but an evolution of the business model from pure infrastructure spending to a monetizable platform asset."

Notably, data from Synergy Research shows that in 2025, revenue in the neocloud (new-type computing cloud) market exceeded $25 billion, growing by over 200% year-on-year. Gartner predicts that by 2030, neocloud providers will capture 20% of the AI cloud market share. However, McKinsey also warns that this business model faces commoditization risks—when GPU supply gradually becomes more abundant, models that compete solely on GPU availability will face compressed profit margins. The entry of hyperscalers like Meta has undoubtedly intensified this competitive pressure.

Data Center Construction Hits Bottlenecks

While the computing power rental market is booming, physical data center construction has repeatedly run into harsh real-world obstacles.

Earlier in July, QTS, a data center operator under Blackstone, officially halted the Digital Gateway project in Virginia. The project spanned 2,100 acres, with originally planned investment of over $100 billion to build 37 data center buildings, which would have become the world's largest data center park upon completion. However, after five years of local resident protests, legal setbacks from a state court ruling invalidating zoning approvals, and the early exit of partners, Blackstone ultimately chose to cut its losses. A few days prior, Blackstone had sold three mature data center assets in Virginia for $3.5 billion, sending a clear signal of strategic contraction.

Similarly, computing infrastructure firm Crusoe announced in June that it would "pause" a massive 1.8GW data center project in Wyoming. Reports indicate this was due to pressure from its major customer Google, which raised "serious concerns" about the project. The power consumption of this facility would be enough to supply electricity to a medium-sized city.

The collapse of these super projects exposes multiple real-world dilemmas behind the frantic expansion of AI computing infrastructure.

First is the physical bottleneck in power supply. Data centers are veritable "power hogs." According to data from the Electric Power Research Institute, data centers currently account for 5% of U.S. electricity demand, a figure that could triple by 2035. In Virginia—the region with the densest concentration of data centers globally—this proportion already exceeds 25%. Existing power grids simply cannot keep up with the scale and growth rate of AI infrastructure demand. JPMorgan Chase's analysis shows that over 60% of data center projects scheduled for completion in 2027 have not yet started construction, and one of the core reasons is power supply bottlenecks. In the first quarter of 2025 alone, the total value of delayed data center projects across the U.S. reached approximately $130 billion.

Second comes public pushback and tightening regulations. A Gallup poll shows that 70% of Americans oppose building AI data centers near their homes. Practical issues such as high energy consumption, noise, water usage, and rising living costs have caused grand AI narratives to repeatedly face setbacks at the community level. In the first quarter of 2026, opponents across the U.S. blocked or delayed at least 75 data center projects. The number of active grassroots opposition groups targeting data centers nationwide surged from 396 at the end of 2025 to 833 in March 2026, covering 49 states. According to statistics, the number of canceled data center projects quadrupled in 2025 to 25, with $18 billion worth of projects blocked and $46 billion delayed.

Additionally, compliance and security requirements have also become constraints. Microsoft's abandonment of the $3 billion cloud computing power lease agreement with Oracle stemmed from the fact that Oracle lacked the federal security certifications needed to manage U.S. government data and was unwilling to undertake large-scale engineering modifications to obtain them. This incident demonstrates that amid increasingly abundant computing power supply, security compliance is becoming a hard threshold for computing power transactions.

Shortages of power, water, and permits are replacing "chip shortages" as the biggest hard constraints for computing infrastructure. Blackstone's exit and Crusoe's pause mark that capital's attitude toward AI infrastructure investment is shifting from frenzy to rationality. These bottlenecks will not eliminate computing power demand, but they will slow down the pace of demand realization—existing locked-in orders will not disappear, but the implementation timeline for new projects will be significantly extended.

Computing Power Supply and Demand Enter a New Landscape

What does the rise of computing power leasing and the slowdown in infrastructure projects mean for the semiconductor industry chain?

First and foremost, it must be clarified that there is no surplus of high-end AI computing power. Industry research indicates that the current computing power market is facing "structural mismatch"—while some low-end general-purpose computing power without clear application scenarios sits idle, the gap in high-end intelligent computing power supporting large model training remains as high as 40%, with demand outstripping supply.

This assessment is fully validated in the financial reports of semiconductor giants. Nvidia's full-year revenue for FY2026 reached a record $215.9 billion, up 65% year-on-year, with data center revenue hitting $193.7 billion, accounting for nearly 90% of total revenue. Its latest quarterly results were even more impressive, with data center revenue surging 92% year-on-year. TSMC CEO C.C. Wei explicitly stated in June that global demand for AI chips remains strong, and although the company is working to expand production, supply will still not be able to meet demand in the coming years. TSMC's May revenue jumped 30% year-on-year, and its 2026 capital expenditure is projected to be $52–56 billion, with internal plans leaning toward the upper end of that range. According to recent reports, AI chip manufacturers like Nvidia are still facing supply shortages, and TSMC's advanced process and advanced packaging capacity remains tight.

In the memory sector, competition over HBM is still intense. SK Hynix, leveraging its leading position in the HBM market, has surpassed Samsung Electronics in market capitalization to become the most valuable company in South Korea. Both Samsung and SK Hynix have moved up mass production of their next-generation HBM4 to early 2026 to cope with the skyrocketing AI demand.

However, the widespread adoption of the computing power rental model is indeed reshaping the procurement logic of the industry chain. When giants like Meta and xAI open their computing power to external customers, they are essentially improving the overall utilization rate of computing power across society. Small and medium-sized AI companies no longer need to purchase expensive hardware, and instead are shifting to leasing. A report from Apollo points out that GPU prices have risen by roughly 8 times since early 2025, making the leasing model more attractive to SMEs. This resource-sharing model, to a certain extent, slows down the absolute growth rate of total computing power demand, pushing cloud providers to focus more on cost-effectiveness and energy efficiency when purchasing hardware.

This is also the underlying reason why major AI players are investing in in-house developed chips. On June 24, OpenAI, in partnership with Broadcom, officially launched the Jalapeño chip, specifically optimized for large model inference—OpenAI's first in-house developed chip, completed from design to production in just 9 months. Meanwhile, Anthropic is in talks with Samsung to develop custom AI chips; Meta's fourth-generation in-house "Iris" chip is scheduled to go into production in September, with the goal of doubling computing power. Faced with high GPU costs, major AI companies are using custom specialized chips to reduce per-unit inference costs and reduce over-reliance on Nvidia. This trend is a major positive for chip design firms like Broadcom, but it poses a potential threat to Nvidia's long-term market share in the inference space.

From a broader industry chain perspective, the rise of the computing power leasing model is fostering a new supply-demand balancing mechanism. In the past, the AI computing power supply chain was linear: chip designers shipped products to cloud providers, who used them for their own operations or resold them to end customers. Today, giants are not only the largest buyers of chips, but also providers of leased computing power. This dual identity of "in-house use plus external leasing" has greatly improved the efficiency of computing power resource allocation. For semiconductor equipment manufacturers, this means downstream customers' procurement behavior will become more rational—no longer panic-driven hoarding, but refined purchasing based on actual utilization rates and return on investment. In the short term, this may lead to a slowdown in the pace of some orders; but in the long run, a healthier demand structure will instead benefit the sustainable growth of the entire industry chain.

This article originates from the WeChat public account "Semiconductor Industry Insights" (ID: ICViews), authored by Jun Xi, and published with authorization from 36Kr.