From the release of DeepSeek, which is a low-cost competitor to ChatGPT, to Microsoft scaling back its Global AI Datacenter projects, and Alibaba Director Zhang Yong warning about a bubble in USA AI infrastructure investments, multiple warning signals are flashing: the boom in AI Datacenters may soon be facing a cool down.
Data shows that since the DeepSeek incident occurred in January, Goldman Sachs' "Energy + AI" thematic investment portfolio has been in a downward trend.
Financial blog ZeroHedge previously pointed out that with the emergence of more efficient Large Language Models (LLMs), the market is forming a new trend of 'doing more with less.' On Thursday, Goldman Sachs analysts James Schneider, Michael Smith, and others released a report that advanced their previous forecast for the peak time of global Datacenter capacity from the end of 2026.
In a report to clients, Schneider stated that he updated the supply-demand model for global Datacenters, mainly considering the impact of DeepSeek and the new capacity brought by projects like OpenAI's Stargate.
In the new forecast, he moved up the peak utilization time for global Datacenters to 2025 (previously the end of 2026), while predicting that the supply-demand tension will gradually ease between now and 2027. Despite this, the utilization rate of Datacenters will remain above historical average levels.
Schneider pointed out that there are three major uncertainties for AI Datacenters in the future: first, the monetization ability of consumer-facing AI services is weak; second, large AI infrastructure projects may lead to oversupply; third, efficiency improvements brought by 'small' LLMs aimed at enterprises.
From the demand side, according to the report, Goldman Sachs' global Technology team has recently revised down its forecast for AI training Server shipment volumes, updating the demand growth for various types of Datacenters. This adjustment is related to the slowdown in AI training demand, as well as the pace of adoption of AI inference and corresponding Datacenter workloads.
Thus, Goldman Sachs lowered the growth forecast for demand most directly affected by AI for 2025 and 2026. Additionally, Goldman Sachs adjusted the Historical Data to better reflect the actual changes in Datacenter supply. These changes resulted in an upward revision of the Historical Data demand baseline, but new demand over the next 18 months was revised downward, while the demand trend for 2027 and beyond remains largely unchanged.
On the supply side, Goldman Sachs' model also made corresponding adjustments, incorporating actual supply that was put into operation by the end of 2024, while also adding small Datacenter operators that were previously not tracked. This brought about an upward revision of approximately 2GW in Historical Data supply, which includes both corrections to the Historical Data baseline and actual capacity growth. In the long term, Goldman Sachs expects that the new Datacenter supply coming online in 2030 will increase by 8%, primarily due to some validated new construction projects.
Despite the adjustments in forecasts, Analysts still hold a constructive view on Datacenter operators such as Digital Realty (DLR) and Equinix (EQIX), pointing out that after the AI demand expectations have become more rational, the risk/reward profile of these companies has become more balanced.
Regarding future risks, Goldman Sachs also conducted a quick survey of its clients, asking them what the biggest challenge for AI-themed investments will be in 2025. The results showed that a quarter of respondents chose 'Efficiency Gains'.
