Written by: Xu Chao

The contradiction between supply and demand for artificial intelligence infrastructure is intensifying among the world's top technology companies. According to informed sources, Google informed Meta around March of this year that it could not meet all of its Gemini computing power needs and imposed a usage limit on this social media giant—even the world's largest AI service provider is struggling to cope with the overwhelming demand for computing power.

According to the Financial Times, these restrictions have yet to be lifted and have led to disruptions and delays in several internal AI projects at Meta. As a result, Meta has asked its employees to improve the efficiency of AI computing power usage and is implementing meticulous budgeting for AI tokens internally. Both Google and Meta declined to comment on this matter.

This situation has forced Google to accelerate its expansion efforts. Earlier this month, Google signed a $920 million monthly computing power leasing agreement with SpaceX, owned by Elon Musk. Google CEO Sundar Pichai admitted during the first quarter earnings call this year: "We are indeed facing constraints in computing power recently, and if we could meet the demand, cloud business revenue would be higher."

Meta is not alone in this predicament. Several informed sources pointed out that other Google enterprise clients are also facing varying degrees of restrictions, with Meta being the most affected due to its unusually large demand. This turmoil reflects the explosive growth of AI reasoning workloads, which has become one of the biggest challenges facing the entire industry.

The computing power bottleneck continues to be under pressure, with large clients being the first to feel the impact

Despite major tech companies having invested hundreds of billions of dollars in chips, data centers, and power supply, the supply of AI computing power is still struggling to keep pace with the growth in demand.

Google's cloud business revenue broke the $20 billion mark for the first time in the first quarter, and the backlog of signed but undelivered cloud contracts nearly doubled compared to the previous quarter, exceeding $460 billion. Pichai clearly stated that computing power constraints will continue in the near future.

In this context, the impact on Meta is particularly pronounced. Informed sources indicated that it is the high-intensity demand from large enterprise clients like Meta that is directly driving Google to accelerate its search for external computing power sources. As enterprises scale up the deployment of chatbots, programming assistants, and AI agents, reasoning workloads—i.e., the computing power consumed when executing tasks in practical applications after model training—are becoming the core bottleneck in the industry.

Meta's internal projects are hindered, accelerating the shift to in-house models

Meta extensively uses Gemini internally, covering platform security reviews (including identifying fraudulent content and clearing harmful information), customer service and advertising support chatbots, as well as some internal workflows and code development, while also using other models such as Anthropic's Claude.

According to informed sources, Meta initially chose Gemini because it outperformed the company's in-house developed Llama open-source model. However, due to tightening computing power restrictions, Meta is accelerating its shift to in-house models. Several informed sources stated that Meta has recently begun prioritizing its newly launched Muse Spark model, which is believed to perform comparably to Gemini and will help reduce dependence on external models.

Meta CEO Mark Zuckerberg has previously increased investment in AI talent and infrastructure, committed to creating what he calls "personal superintelligence." Unlike Google, Meta does not have a cloud business and is accelerating the construction of its own data center system, promising to invest a total of $600 billion in the U.S. by 2028.

Google expands through SpaceX, the industry seeks breakthroughs

In the face of pressure on computing power, Google signed a $920 million monthly computing power leasing agreement with SpaceX this month to fill infrastructure gaps. The AI lab Anthropic also reached a similar agreement with SpaceX last month.

Google's imposition of restrictions on Meta provides a rare window for outside observers to glimpse the real pressures faced by the world's leading AI service providers in computing power allocation. Currently, the infrastructure bottlenecks in the entire AI industry are spreading from the training side to the reasoning side, and resolving the supply-demand contradiction still relies on the implementation of a new round of large-scale capital investments.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。