Full moons wane.

Wu Wei, Investor Network

In July 2026, the tech and financial markets were rocked by a major announcement. On July 2nd, Meta Platforms (META.US), one of the world's largest buyers of computing power, declared its push to commercialize computing power, officially stepping into the market to 'sell computing power.' This shift from 'buyer' to 'seller' not only directly triggered violent fluctuations (which can be translated as 'severe turbulence' in context, but kept as is for precision to original tone if desired in final output) in the global semiconductor sector but also acted as a spotlight, revealing the deep-seated transformations hidden in the current AI infrastructure landscape.

Currently, the global computing power market stands at a delicate crossroads. On one hand, the construction of computing power centers struggles under the pressures of power depletion, stricter regulations, and high costs. On the other hand, the evolution and commercialization of large models face tests, forcing related companies to reevaluate their massive computing power expenditures. Thus, the entire industry seems to be bidding farewell to the initial stage of 'brute-force stacking' and moving into a deeper waters of refined operation (translated as 'refined operations') and Return on Investment (ROI) competition.

The market can't help but wonder: Is there already an oversupply of computing power?

The Giant's Turn: Capital Market Turmoil Triggered by Meta's 'Computing Power Sell-Off'

In early July 2026, news emerged that Meta was formally advancing an internal plan codenamed 'MetaCompute.' Specifically, the plan includes two business models: one is a 'Model-as-a-Service' similar to AWS, where Meta provides hosted access to its MuseSpark and Llama4/5 models for clients; the other is a 'Bare Metal Computing Power Rental' service, directly leasing idle GPUs.

The underlying driving force behind Meta's decision likely stems from the contradiction between its previously enormous capital expenditure pressures and the periodic idleness of its computing power. After completing Llama 4 training, Meta's massive computing power cluster experienced a downtime period before developing Llama 5.

Data shows that by the end of 2025, Meta's computing power scale was equivalent to approximately 2.5 million H100 graphics cards, and the company's capital expenditure guidance for 2026 reached a staggering $125 billion to $145 billion. High depreciation and investment costs compelled Meta's management to recoup funds and improve asset turnover through monetization.

This news triggered a chain reaction in the capital markets that could be described as a 'severe earthquake.' Firstly, funds affirmed Meta's decision, with Meta's (META.US) stock price surging 8%-10% in a single day after the news broke. Investors gave highly optimistic feedback on its shift from a single 'money-burning mode' to a 'revenue-generating mode.'

Conversely, the computing power sector suffered a 'heavy blow.' After the announcement, the Philadelphia Semiconductor Index plummeted by over 6%. The market began to worry: If even Meta is selling off computing power, does this imply a structural oversupply of AI hardware? Logically, this strike (translated as ' strike ' to 'hit') suppressed the valuation logic of core hardware manufacturers such as NVIDIA (NVDA.US), Micron Technology (MU.US), and Advanced Micro Devices (AMD.US).

Shares of emerging cloud providers like CoreWeave and Nebius, which rely solely on renting and selling GPUs for survival, plummeted by 10%-17%. Their once-major client instantly became a competitor with significantly stronger scale and cost advantages, posing a severe challenge to these startups' business models.

Regarding Meta's move to sell computing power, the market generally believes that computing power is transforming from a 'scarce resource' to a 'commodity.' The industry has begun to worry: If even the most financially robust giants need to resell computing power, does this mean that the industry's appetite for computing power can no longer keep pace with the expansion of hardware production capacity? This may also become a focal point of market speculation in the second half of 2026.

Demand Fog: Changing Mindsets of Buyers

The debate over whether 'demand for computing power is decreasing' does not have a simple black-and-white answer. Instead, 'nominal demand' is shrinking, while actual demand shifts towards 'efficiency as king.'

With technological maturity, the research and development paradigm for large models has undergone substantive changes. By 2026, inference computing power demand will account for over 70% of total societal computing power consumption. Thus, the market no longer craves massive bursts of training computing power but instead requires stable, cost-effective distributed inference networks.

Furthermore, besides enhancing large model capabilities, companies are also focusing on optimizing large models. Currently, Llama4/5 widely adopts a Mixture of Experts (MoE) architecture, significantly compressing the computational volume activated in a single instance. Meanwhile, the trend of 'training large models, deploying small models' is gradually becoming mainstream. Coupled with the proliferation of edge devices like AI phones and AI PCs, approximately 30% of daily AI interactions may reflow (translated as 'flow back') to local terminals, significantly reducing the marginal demand for cloud computing power.

At the commercialization level, companies' purchasing logic for computing power has shifted from 'PPT demonstrations' to comprehensive 'ROI calculations.' 2026 is seen as the 'Year of AI ROI Auditing.' Since killer consumer-side applications have not brought about the expected comprehensive paid explosion, companies have begun shutting down computing power expenditures that do not directly drive business growth. The 'premium' on computing power may be gradually disappearing. Most non-leading companies no longer blindly participate in the 'arms race' for general-purpose large models but instead turn to fine-tuning vertical industry models with lower computing power demands.

This return to rationality is directly reflected in the funding chains of AI companies. Wall Street's patience with tech giants is wearing thin, demanding clear demonstrations of positive correlations between AI investments and revenues in their financial reports.

Small and medium-sized AI startups are facing a financing winter. Some companies that blindly leveraged up to 'stockpile cards' in 2024 are now forced to resell computing power due to repayment pressures, leading to a surge in 'second-hand computing power' supply in the market. Even cash-rich leading companies like OpenAI have seen changes in their capital flows. Currently, these companies are allocating more funds towards 'power locking' and 'data purchasing' rather than pure hardware procurement.

High Costs: Supply-Demand Rebalancing Amid Computing Power 'Inflation'

While demand tends towards rationality, the computing power construction side faces insurmountable 'physical ceilings' and supply chain inflation pressures. In 2026, power supply is gradually replacing chips as the biggest constraint on the global expansion of computing power centers.

In North America, over 40% of computing power projects under construction in places like Virginia have been delayed due to lagging grid upgrades. In Europe, 'power depletion,' 'regulatory storms,' and 'resident protests' have become the three major obstacles to computing power construction. The queue time for obtaining large-scale power has extended to 7-10 years in places like London. Germany has even introduced stringent regulations requiring new data centers to achieve an extreme energy efficiency rating of PUE 1.2, leading to numerous project cancellations or 'houses without power' dilemmas.

Against this backdrop, NVIDIA's (NVDA.US) previously unbeatable closed-loop model of 'investment-procurement-revenue' has come under dual scrutiny from Wall Street and regulators.

NVIDIA previously secured large-scale procurement orders for its latest GPUs from computing power cloud startups through significant capital injections. However, with regulatory intervention from the U.S. SEC, this model has been accused by short-selling institutions as 'boomerang trade' manipulating market demand.

More severe (translated as ' severe ' to 'severe') is that, constrained by stalled data center construction, approximately 15%-20% of high-performance GPUs globally were 'unboxed but unpowered' in July 2026, directly severing this investment rolling chain. Currently, invested companies, facing excess inventory and difficulty in profitability, are being forced to sell chips at discounted prices on the second-hand market, unable to afford NVIDIA's next-generation chips.

As one of the world's core computing power markets, China is accelerating the construction of an independent computing power system. Industry chain ecosystems represented by companies like Hygon Information (688041.SH) and Cambricon Technologies (688256.SH) have begun to take shape. Meanwhile, leading Chinese tech giants like Alibaba and Tencent are also actively building their own computing power infrastructures. Driven by the dual forces of domestic substitution and self-research by large companies, the demand space for NVIDIA GPUs in the Chinese market undoubtedly faces significant compression.

However, the breakthrough of the independent computing power system also comes with short-term growing pains. On one hand, constrained by advanced packaging yield fluctuations and insufficient scale effects in the initial stages of HBM (High Bandwidth Memory) localization, the comprehensive procurement cost of domestically produced single-chip computing power rose by approximately 20%-30% in the first half of 2026.

On the other hand, the hidden costs of ecosystem migration are also high. When large model companies migrate from existing systems to domestic architectures, the software reconstruction and talent training expenses required account for over 35% of total computing power construction investments. These high software and hardware conversion costs are forcing the market towards rationality, leading large-scale infrastructure projects like China's 'East Data West Calculation' to gradually abandon extensive expansion and turn towards refined upgrades and transformations of existing computing power facilities. Under such circumstances, market expectations for computing power construction demand will undoubtedly adjust.

Furthermore, the significant inflation in global component costs, leading to prolonged investment return cycles, has further dampened companies' enthusiasm for investing in computing power. Spot prices for HBM4/4e memory have surged by 40% year-on-year, while copper prices have hit record highs, driving up data center civil engineering costs by 15%. Coupled with continued tight capacity in advanced packaging, both the construction cycle and capital costs of computing power centers have lengthened.

A single leaf heralds autumn. Meta's entry into selling computing power may foreshadow (translated as ' foreshadow ' to 'foreshadow') that the global computing power market in 2026 is undergoing a process of 'fever subsiding.' Purely commercially driven computing power rental companies are experiencing a shakeout, replaced by 'sovereign AI' infrastructures led by governments with strategic intentions.

Currently, the steep growth trajectory of computing power construction has slowed, and the industry is bidding farewell to its frenzied era. The performance forecast and stock price decline of optical module companies like Hangzhou Electric Cable Co., Ltd. (603618.SH) may be a signal. In the future, the market will no longer blindly pay for hardware scale but will instead grant true premiums to those companies capable of overcoming power bottlenecks, extremely lowering energy efficiency ratios, and achieving positive business cycles in vertical industries first.

What are your thoughts on Meta selling computing power? Feel free to leave comments, share, and forward. (Produced by Sui Si Finance) ■

Source: Investor Network