Has AI infrastructure hit the brakes?

(This article was authored by Semiconductor Industry Insights and published by TMTPost with authorization)

By Semiconductor Industry Vertical

Recently, the global AI infrastructure sector has undergone dramatic shifts. First, Meta announced plans to lease out some of its idle AI computing capacity, triggering sharp volatility in financial markets. Shortly afterward, SoftBank Group officially launched a new company, SB Neo, making a major push into the U.S. computing capacity leasing market. Meanwhile, the veteran private equity firm Blackstone abruptly halted its previously planned $100 billion-plus investment in what would have been the world’s largest data center project, and Microsoft walked away from a $3 billion cloud computing leasing agreement with Oracle due to security concerns.

On one side, tech giants are entering the market to 'sell computing power'; on the other, trillion-dollar infrastructure projects are slamming on the brakes.This series of seemingly contradictory moves has left the market stunned: Could AI computing capacity already be in oversupply? Is the investment bubble in computing infrastructure about to burst?

A deeper analysis from the perspective of the semiconductor industry reveals not a simple case of 'computing oversupply,' but rather a profound restructuring of the logic underpinning AI industry development. The past two years—characterized by reckless, cost-agnostic expansion and land grabs—are coming to an end. Competition in AI infrastructure is now entering a new era where 'efficiency reigns supreme.'

In July 2026,Meta announced plans to launch a cloud infrastructure business, offering external customers access to its AI computing capacity and models.The news sent Meta’s stock soaring nearly 9% in a single day, adding roughly $127 billion to its market capitalization. However, the broader AI computing supply chain came under pressure—newcomers in computing leasing like CoreWeave and NEBIUS saw their shares plunge more than 13%, while memory chip giants Micron, SK Hynix, and Samsung Electronics all posted significant declines.

The market’s initial reaction was panic: even Meta can’t fully utilize its own GPUs, which must mean computing capacity is oversupplied.

Yet this linear thinking overlooks the unique nature of AI computing assets and the true intentions behind the giants’ moves.Meta's move to rent out computing capacity is fundamentally an upgrade in asset operational efficiency, not a signal that demand has peaked.

For 2026, Meta has guided capital expenditures to a staggering $125–145 billion, with the vast majority allocated to data centers and GPU procurement. To date, Meta has already committed a total of $183 billion to AI infrastructure investments. As a company whose revenue is 98% advertising-driven, its annual investments—amounting to hundreds of billions of dollars—are being converted into massive computing clusters. However, its open-source Llama models do not directly generate revenue. Monetizing previous-generation computing capacity or temporarily idle resources externally not only directly offsets depreciation and operational costs but also represents a critical step in transforming GPU clusters from 'pure cost centers' into 'revenue-generating assets.' Morgan Stanley estimates that if Meta rents out 250 megawatts (MW) of computing capacity for one year, it could generate approximately $10 billion in revenue.

This is not unique to Meta. Previously, Elon Musk’s xAI successfully leased out capacity from its Colossus supercomputing cluster on a massive scale. According to multiple media reports, Anthropic leased the entire capacity of Colossus 1—approximately 220,000 NVIDIA GPUs—for $1.25 billion per month, under a contract running until May 2029, valued at roughly $40 billion in total. Google also pays $920 million monthly to lease bridging capacity to compensate for delays in its own data center construction. These two deals alone generate over $2.1 billion in monthly cash flow for SpaceX. Institutional analyses estimate that, at these rental rates, the implied return on investment allows full recovery of capital expenditures in about two years.

SoftBank Group’s entry further validates the attractiveness of this sector.On July 2, SoftBank announced the formation of SB Neo, planning to launch cloud services based on NVIDIA’s latest GPUs for U.S. enterprises in fiscal year 2027. The initiative aims to build AI data center infrastructure with a total capacity of 10 gigawatts (GW), starting with an initial 800 MW deployment site in Ohio. To support this expansion, SoftBank is securing a $10 billion loan using its OpenAI equity stake as collateral.

From Meta to xAI and now SoftBank, tech giants are increasingly becoming 'computing landlords'—not because they no longer need computing power, but because, amid soaring capital expenditures on AI infrastructure, they must find new pathways to generate returns. As TF Securities pointed out: 'Meta entering the AI cloud business does not mean GPUs are universally oversupplied. This is not the end of AI capex deals but rather an evolution of the business model—from pure infrastructure spending toward monetizable platform assets.'

Notably, Synergy Research data shows that revenue in the neocloud (next-generation computing cloud) market exceeded $25 billion in 2025, more than doubling year-over-year. Gartner forecasts that by 2030, neocloud providers will capture 20% of the AI cloud market share. However, McKinsey simultaneously cautions that this business model faces commoditization risks—as GPU supply gradually eases, models relying purely on GPU availability for competitive advantage will see margin compression. The entry of hyperscalers like Meta undoubtedly intensifies this competitive pressure.

While the computing capacity leasing market heats up, physical-world data center construction repeatedly runs into real-world roadblocks.

In early July, Blackstone’s data center operator QTS officially halted the Digital Gateway project in Virginia. Spanning 2,100 acres, the project was originally slated for over $100 billion in investment to construct 37 data center buildings, which would have made it the world’s largest data center campus upon completion. However, after five years of local resident opposition, a state court ruling invalidating zoning approvals, and the prior exit of partners, Blackstone ultimately chose to cut its losses and withdraw. Just days earlier, Blackstone had already sold three mature Virginia data center assets for $3.5 billion, sending a clear signal of strategic retrenchment.

In a similar vein, computing infrastructure company Crusoe announced in June that it had 'paused' a massive 1.8-gigawatt data center project in Wyoming. Reports indicated this move came under pressure from its major client, Google, which raised 'serious concerns' about the project. The facility’s power consumption would have been sufficient to supply a mid-sized city.

The collapse of these mega-projects reveals multiple real-world challenges underlying the AI computing infrastructure boom.

First is the physical bottleneck of power supply.Data centers are genuine 'power hogs.' According to 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 world’s most densely concentrated region for data centers, this share already exceeds 25%. Existing power grids simply cannot keep pace with the scale and growth rate of AI infrastructure demands. JPMorgan analysis shows that over 60% of data center projects slated for completion by 2027 have not yet broken ground, with power supply constraints being a primary reason. In Q1 2025 alone, delayed data center projects across the U.S. amounted to approximately $130 billion in total value.

Second is public resistance and tightening regulations.A Gallup poll shows that 70% of Americans oppose building AI data centers near their communities. Practical concerns—such as high energy consumption, noise, water usage, and rising living costs—have repeatedly stalled grand AI narratives at the local level. In Q1 2026, opponents nationwide blocked or delayed at least 75 data center projects. Active grassroots opposition groups targeting data centers surged from 396 at the end of 2025 to 833 by March 2026, spanning 49 states. In 2025 alone, canceled data center projects quadrupled to 25, with $18 billion worth of projects halted and $46 billion delayed.

In addition,Compliance and security requirements have also become constraining factors.Microsoft abandoned a $3 billion cloud computing lease agreement with Oracle precisely because Oracle lacked the federal security certifications required to manage U.S. government data and was unwilling to undertake the extensive engineering upgrades needed to obtain them. This incident highlights that, amid increasingly abundant computing capacity, security and compliance are becoming hard prerequisites in compute transactions.

Power shortages, water scarcity, and permitting delays are now replacing 'chip shortages' as the biggest hard constraints on computing infrastructure. Blackstone’s exit and Crusoe’s pause signal that investor sentiment toward AI infrastructure is shifting from frenzy back to rationality. These bottlenecks won’t kill demand for computing power, but they will slow the pace at which that demand materializes—already secured orders won’t vanish, but the timeline for new projects to come online will significantly lengthen.

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

First, it must be clarified that high-end AI computing power is not oversupplied. Industry research indicates thatthe current computing power market suffers from a 'structural mismatch'—while some low-end general-purpose computing capacity lacking practical applications remains idle, there is still a 40% shortfall in high-end AI computing power needed for large model training, leaving it in a state of undersupply.

This assessment has been strongly corroborated by financial reports from semiconductor giants. NVIDIA reported record annual revenue of $215.9 billion for fiscal year 2026, up 65% year-over-year, with data center revenue reaching $193.7 billion—nearly 90% of total revenue. Its most recent quarterly results were even more impressive, with data center revenue surging 92% year-over-year. C. C. Wei, CEO of Taiwan Semiconductor, explicitly stated in June that global demand for AI chips remains robust, and despite the company’s aggressive capacity expansion efforts, supply will still fall short of demand for the next few years. Taiwan Semiconductor’s May revenue jumped 30% year-over-year, and the company forecasts capital expenditures for 2026 to range between $52 billion and $56 billion, leaning toward the upper end. According to recent reports, AI chipmakers like NVIDIA continue to face shortages, and Taiwan Semiconductor’s advanced process and advanced packaging capacities remain tight.

In the memory sector, competition for HBM remains fierce. SK Hynix, leveraging its leading position in the HBM market, has surpassed Samsung Electronics in market capitalization to become South Korea’s most valuable company. Both Samsung and SK Hynix have moved forward the mass production timeline for their next-generation HBM4 to early 2026 to meet soaring AI demand.

However,the growing adoption of computing power leasing models is indeed reshaping procurement logic across the industry chain.When giants like Meta and xAI open their computing resources to external clients, they are effectively boosting overall societal utilization of computing power. Smaller and medium-sized AI companies no longer need to purchase expensive hardware but instead turn to leasing. A report from Apollo notes that GPU prices have risen approximately eightfold since early 2025, making leasing significantly more attractive for smaller enterprises. This resource-sharing model has somewhat slowed the absolute growth rate of total computing power demand, prompting cloud providers to prioritize cost-effectiveness and energy efficiency when procuring hardware.

This is also the underlying reason why major AI firms are increasingly investing in custom-designed chips. On June 24, OpenAI and Broadcom jointly unveiled the Jalapeño chip, specifically optimized for large-model inference—OpenAI’s first in-house chip, developed from design to production in just nine months. Meanwhile, Anthropic is in talks with Samsung to co-develop customized AI chips, and Meta’s fourth-generation in-house chip, codenamed 'Iris,' is scheduled to enter production in September, aiming to double its computing capacity.Faced with soaring GPU costs, major AI companies are turning to specialized, custom-designed chips to reduce per-unit inference costs and lessen their heavy reliance on NVIDIA.This trend significantly benefits chip design companies like Broadcom, but poses a potential threat to NVIDIA's long-term market share in the inference segment.

From a broader industrial chain perspective, the rise of the compute capacity leasing model is fostering a new supply-demand equilibrium mechanism. In the past, the AI computing supply chain was linear: chip designers shipped products to cloud providers, who either used them internally or resold them to end customers. Today, tech giants are simultaneously the largest buyers of chips and providers of leased computing power. This dual role of 'self-use plus external leasing' has substantially improved the efficiency of computing resource allocation. For semiconductor equipment suppliers, this means downstream customers’ procurement behavior will become more rational—shifting from panic-driven stockpiling to precision purchasing based on actual utilization rates and return on investment. In the short term, this could slow the pace of some orders; however, over the long run, a healthier demand structure will support sustainable growth across the entire industry chain.

58K Views