U.S. Stocks Dip Again: Is Tech Giant Microsoft Doubting AI Demand?

TradingKey - On Wednesday, a report highlighting changes in the computing power market caused a decline in U.S. AI-related stocks. A research report from TD Cowen stated that Microsoft has canceled agreements with at least two private data center operators in the U.S. for hundreds of megawatts of total power capacity.

On the same day, the Philadelphia Semiconductor Index closed down around 3.3%. Nvidia fell by more than 5.7%, leading the declines among the seven major tech companies. AI-related stocks broadly dropped, with AMD declining nearly 8.9%.

TD Cowen’s channel survey revealed that Microsoft’s cancellations and delays in data center leases were more widespread than anticipated, extending beyond the U.S. to the European market as well. Microsoft has both (1) abandoned over 2 gigawatts of capacity undergoing leasing processes in the U.S. and Europe and (2) postponed and canceled existing data center leases in both regions over the past six months. However, Microsoft remains committed to its $80 billion investment plan.

Microsoft’s Shift May Be Related to OpenAI

"When coupled with prior channel checks, this indicates a potential oversupply situation for Microsoft," stated analysts led by TD Cowen Michael Elias. 

They noted that Microsoft's changes may be associated with OpenAI, the creator of ChatGPT, in which Microsoft holds a stake. TD Cowen had previously confirmed that Microsoft’s data center in Wisconsin was intended to support OpenAI’s operations. Predictions for workload capacity in 2023 and 2024 led Microsoft to frequently lease new capacity. 

However, the emergence of DeepSeek earlier this year has prompted the industry to reevaluate the AI industry. This shift may suggest excess data center capacity for the company.

Growing Concerns About AI Demand

Earlier this week, Alibaba Chairman Joe Tsai warned that there might be a bubble in data center construction, suggesting that new projects could outpace the demand for AI services.

Tsai pointed out that a lot of data center investments in the U.S. seem to be unnecessary or similar, leading to wasted resources and intense competition. For instance, multiple companies are raising capital through SPACs (special purpose acquisition companies) to build data centers, but they have yet to secure actual user agreements, indicating a lack of clear market demand.

Meanwhile, a Goldman Sachs analyst team has downgraded projections for training-related computing power and shipments of AI training servers. Shipments for 2025 and 2026 have been revised down from 31,000 units and 66,000 units to 19,000 units and 57,000 units, respectively (based on 144-GPU equivalent calculations). These adjustments are primarily due to product transition effects and uncertainties in supply and demand.

More Players in Data Center Industry

The report also indicated that OpenAI is increasingly seeking to acquire data center capacity directly from third-party providers, including GPU-as-a-Service suppliers and third-party data center operators. Previously, OpenAI was set to sign a nearly $12 billion agreement with CoreWeave to develop data center-related capacity.

Moreover, OpenAI may construct its own data centers in the long term. Surveys suggest that OpenAI plans to create multiple "Stargate" projects, each representing capacity of 800 MW to 1.5 GW, accumulating potential long-term capacity requirements of over 6 GW. OpenAI is recruiting experts in design, construction, and capacity planning from other hyperscale operators.

Google is filling the capacity vacated by Microsoft in international markets, while Meta is acquiring related capacities in the U.S. Both companies are experiencing significant year-on-year growth in data center demand.

Microsoft Advances Liquid Cooling Technology

In October of last year, Microsoft announced plans for data center operators to retrofit existing facilities to support liquid cooling technology. Recent surveys indicate that Microsoft is developing an internal plan aimed at improving the rack density of traditional cloud data centers by implementing air-assisted liquid cooling solutions.

Traditional data centers typically lack facilities for direct chip liquid cooling, requiring standalone liquid-to-air heat exchangers. This air-assisted liquid cooling approach is viewed as more cost-effective than fully retrofitting an entire data center, as it allows for targeted cooling in specific areas to accommodate high-density workloads.

Stocks in Data Centers and Mining Operations Take a Hit

Data center stocks have felt the pinch as well. On Wednesday, U.S. utility companies supplying power to data centers, such as Constellation Energy and Vistra, fell by 4.4% and 5.9%, respectively. Applied Digital, which is heavily reliant on long-term leases from major companies like Microsoft, suffered a 12.2% drop following the news.

This news also led to a drop in the stock prices of Bitcoin mining companies, which might put even more downward pressure on Bitcoin prices. These mining companies were said to be branching out into hosting services for AI data centers to boost their earnings . A report from Coin Metrics in March indicated that miners were expanding into AI data center hosting to create new revenue streams and make better use of their existing infrastructure for high-performance computing. If demand for data center services started to fall, Bitcoin mining companies could face reduced revenues.