Small-scale models are "stealing" cloud computing power demand, putting trillions of dollars in data center capital expenditures to the test.

AI applications continue to expand, but the massive data center investments that underpin this boom are facing new demand-side challenges.

A Jefferies report points out that as the performance of small language models improves and costs decrease, enterprise AI deployments may shift from centralized cloud to localized deployments, thereby putting pressure on cloud computing power demands and huge data center capital expenditures.

A key basis for this judgment is recent observations of the actual usage of AI servers. In September 2026, Swerve Research Technologies constructed a "congestion index" by tracking the waiting time of Anthropic server request queues, and found that the utilization rate of the relevant servers had decreased by 26% compared to the peak in January and February 2026. This indicator does not directly correlate with company revenue, but if the utilization rate continues to decline, it may mean that some computing power is beginning to be idle.

Meanwhile, investment in AI infrastructure continues to expand rapidly. The market currently projects that the combined capital expenditures of the four major US tech giants—Meta, Alphabet, Amazon, and Microsoft—could reach $990 billion in 2027. The report argues that if the growth rate of AI computing power demand ultimately fails to keep pace with such massive investment, some data centers may face the risk of insufficient returns or even becoming "idle assets."

Declining server utilization poses a challenge to AI computing power supply.

Following the launch of Meta's personal AI product Muse on September 8th, Meta's stock price surged 21% within weeks, becoming a recent highlight in AI consumer applications. However, the report argues that the success of a single application is not sufficient to prove the sustainability of the entire AI infrastructure investment cycle.

More noteworthy is that AI companies themselves are also sending signals that differ from the market's optimistic expectations. Anthropic CEO Sam Altman, OpenAI CEO Sam Altman, and Elon Musk have all recently publicly discussed the necessity of slowing down AI development; Anthropic's planned IPO has been postponed to November.

Meanwhile, Anthropic had previously signed a three-year, $45 billion computing power contract with SpaceX and multi-billion dollar computing power partnerships with Amazon and Google Cloud. If AI revenue growth slows while these previously signed large-scale computing power contracts still need to be fulfilled, the company's cost pressures could further increase.

According to reports, Meta also negotiated with Anthropic in July to lease up to $10 billion worth of data center computing power over two years. The report suggests that large tech companies' active search for external customers to utilize their computing power reflects the growing challenge of improving asset utilization as infrastructure supply rapidly expands.

To catch up in performance with smaller models, local deployment or offloading of cloud demand is needed.

The report focuses more on the changes taking place in AI model architecture than on short-term changes in server utilization.

A study titled "Intelligence per Watt," jointly published by Stanford University and Together AI, shows that the performance of small language models is rapidly catching up with that of large language models, while their energy consumption and computing costs are 50% to 85% lower. If this trend continues, the scale of computing power and cost structure required for enterprises to deploy AI may change accordingly.

In the past, enterprises relied more on centralized cloud-based large models to complete AI tasks; as the capabilities of smaller models improve, some scenarios may shift to on-premises deployment. For industries with high data security requirements, such as banks, on-premises deployment already has a strong demand; as the deployment cost of smaller models further decreases, the applicability of this model may continue to expand.

This means that the growth in AI applications does not necessarily correspond to a proportional increase in demand for cloud computing power. Enterprises may expand their AI applications while reducing their reliance on massive model APIs and centralized cloud computing power.

This is also a key reason why the report questions the nearly one trillion dollars in capital expenditure: if more AI applications in the future are carried by low-cost, small-scale, localized models, then the current investment scale surrounding hyperscale data centers may be mismatched with the actual computing power demand.

In this scenario, what really needs to be reassessed is not whether the demand for AI exists, but rather how much centralized computing power AI actually requires, and whether these data centers can achieve a return commensurate with their investment.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.