Small models "steal" demand for cloud computing power, trillion-dollar data center capital expenditure faces a test

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

A Jefferies report points out that as small language models improve in performance and drop in cost, enterprise AI deployments may shift from centralized cloud setups to localized environments, putting pressure on both cloud compute demand and the huge capital expenditures on data centers.

An important basis for this judgment comes from recent observations of actual AI server usage. In September 2026, Swerve Research Technologies constructed a "congestion index" by tracking wait times in Anthropic server request queues and found that server utilization had dropped 26% from its January to February 2026 peak. This indicator doesn't directly correspond to company revenues, but if utilization continues to decline, it may suggest some computing resources are starting to become idle.

Meanwhile, AI infrastructure investment is still expanding rapidly. The market currently expects that combined capital expenditures by Meta, Alphabet, Amazon, and Microsoft could reach $990 billion by 2027. According to the report, if the growth in AI compute demand ultimately fails to keep pace with such massive investments, some data centers could risk insufficient returns or even become "stranded assets."

Server Utilization Drops, AI Compute Supply Faces a Test

After Meta's personal AI product Muse launched on September 8, Meta's stock price jumped 21% in just a few weeks, highlighting Muse as a recent bright spot for consumer AI applications. However, the report argues that the success of an individual application is not enough to prove the sustainability of the entire AI infrastructure investment cycle.

More noteworthy is that AI companies themselves have issued signals that differ from the generally optimistic market expectations. Anthropic’s CEO, OpenAI CEO Sam Altman, and Elon Musk have all recently spoken publicly about the need to slow AI development; Anthropic’s originally planned IPO was also postponed to November.

At the same time, Anthropic previously signed a $45 billion, three-year compute contract with SpaceX and entered into compute deals worth billions with Amazon and Google Cloud. If AI revenue growth slows while these large-scale compute contracts still need to be fulfilled, enterprise cost pressures could rise even further.

According to reports, Meta was also in talks with Anthropic in July, planning to lease up to $10 billion worth of data center compute resources to them over two years. The report suggests that major tech firms are actively seeking external clients to absorb compute capacity, reflecting that as infrastructure supply grows rapidly, improving asset utilization is becoming a new challenge.

Small Model Gains, Local Deployment May Offset Cloud Demand

Compared to short-term server utilization changes, the report is more focused on evolving AI model architectures.

The "Intelligence per Watt" study, jointly released by Stanford University and Together AI, shows that the performance of small language models is rapidly catching up to that of large models, while energy and compute costs are 50% to 85% lower. If this trend continues, the compute scale and cost structure required for enterprise AI deployment could both see significant changes.

In the past, enterprises relied more heavily on centralized cloud-based large models for AI tasks; as small models are enhanced, some scenarios may shift to local deployment. For industries like banking with high data security requirements, there is already strong demand for local deployment; as small models further reduce deployment costs, the applicability of this pattern may expand even further.

This means that the growth of AI applications doesn't necessarily translate to proportional increases in cloud compute demand. Enterprises might simultaneously expand AI applications while reducing reliance on super-large model APIs and centralized cloud compute.

This is also a key reason why the report questions the nearly trillion-dollar capital spending: If more future AI applications are powered by low-cost, small-scale, localized models, the current investments centered around massive data centers may become misaligned with actual compute requirements.

In this scenario, what truly needs to be re-evaluated is not whether AI demand exists, but how much centralized compute is actually needed, and whether these data centers will achieve returns proportional to the scale of their investment.