The sharp decline in A-share technology stocks on July 2nd stemmed from a market misinterpretation of Meta Platforms, Inc.'s plan to sell computing power as a signal of oversupply. Industry insiders believe this move does not signal an end to AI capital expenditure but rather an evolution of the business model from "pure cash-burning infrastructure" to "chargeable platform assets," with demand for computing power still in its early growth stage. Short-term volatility may present a buying opportunity, with technology stocks poised to reach new highs in late July.
On July 2nd, A-share technology stocks experienced a significant drop, with sectors like semiconductors, computing hardware, and memory chips seeing substantial declines. Taking the ChiNext and STAR Market indices, which are heavily weighted with tech stocks, as examples, they fell by 5.71% and 5.64% respectively. The market downturn was triggered by news of Meta Platforms, Inc. selling its computing power, which was interpreted as a sign of oversupply. However, multiple industry professionals consider this a misinterpretation. They argue that Meta Platforms, Inc.'s sale of computing power is not the end of AI capital expenditure but signifies the maturation of the AI infrastructure business model.
"Selling computing power is positive for Meta Platforms, Inc.. Previously, every time the company raised its AI CapEx, the market worried about how to recoup these investments. Now, if Meta Platforms, Inc. can sell part of its AI computing power externally, CapEx is no longer just a cost but could become a direct revenue source," stated one securities firm. "The market previously penalized Meta Platforms, Inc. for high CapEx; now it may revalue this asset."
Chief Strategy Analyst Zhang Qiyao from another securities firm also believes that Meta Platforms, Inc. is unique among hyperscalers. Its primarily consumer-facing business means its AI monetization relies heavily on advertising, and exploring cloud services could enhance shareholder returns and cash flow. As one of the global companies most capable of internally utilizing AI computing power, Meta Platforms, Inc. has numerous application scenarios like advertising and recommendations. Its decision to lease or sell part of its computing power inevitably raises investor concerns: have major tech firms engaged in phased over-procurement in the past two years? Have some H100/H200 GPUs transitioned from being scarce training resources to inference assets needing external monetization?
Regarding this, analysts argue that Meta Platforms, Inc. venturing into AI cloud services does not equate to admitting a comprehensive GPU surplus. It is more likely allocating different generations of AI computing power to different economic uses: the latest GB200/GB300/Rubin resources are prioritized for next-generation model training, while the previous generation H100/H200 are shifted more towards inference and external compute sales. Meanwhile, Alphabet (Google) previously could not fully meet Meta Platforms, Inc.'s demand for accessing Gemini due to capacity constraints, which conversely indicates that cutting-edge models and high-quality inference capacity remain tight. Therefore, this is not the end of AI CapEx transactions but an evolution of the business model from "pure cash-burning infrastructure" to "chargeable platform assets."
Another securities firm also stated that Meta Platforms, Inc. leasing or selling computing power is merely the implementation of "Meta Cloud" and is unrelated to "oversupply." Tracking recent trends in token usage and Annual Recurring Revenue (ARR) shows that AI demand is still in the early stage of steep growth. The backlog for cloud services among North American cloud providers is approximately 14 times their quarterly revenue. "This is merely a reallocation of computing power from Meta Platforms, Inc. to leading large model companies like Anthropic; the industry-wide supply-demand gap will not change."
Analysts emphasize that the market should not simplistically interpret Meta Platforms, Inc.'s external compute leasing/sale as a sign that "AI computing power demand has peaked." "More accurately, this is Meta Platforms, Inc.'s attempt to transform AI infrastructure from a pure cost center into a leasable, chargeable, platform-based asset." In fact, Meta Platforms, Inc.'s intention to sell computing power is not new information; it was hinted at during the annual shareholder meeting in May, stating that if the built computing power exceeded internal needs, it would consider selling the surplus or launching API services.
For the hardware supply chain, analysts believe the key observation is not "whether Meta Platforms, Inc. leases computing power" itself, but whether the actual token usage and ARR growth rates of companies like OpenAI and Anthropic continue to rise. If the ARR and token usage of model companies like OpenAI/Anthropic continue their growth trajectory, and hyperscaler CapEx does not see a substantial downward revision, then the hardware investment thesis remains valid.
"Short-term volatility provides another window for strategic positioning. In late July, as the earnings season reconfirms the strength of the sector's fundamentals and inflation data confirms it has peaked, technology stocks are expected to reach new highs," said Zhang Qiyao.
Comments