Not just "power shortages," "delays" will become a "key theme" for U.S. data centers in 2026

As the construction of artificial intelligence infrastructure enters a deep-water zone, "delays" are gradually replacing mere "power shortages" as the focus of industry attention. Several executives and investors pointed out in recent discussions that project delays will become a core theme in 2026, and this differentiation trend will distinguish world-class operators from ordinary participants.

The current "delays" encompass a broad concept with complex and diverse causes. From utility companies failing to fulfill power commitments and delays in the arrival of key equipment to construction progress delays caused by a shortage of on-site labor, various situations are emerging.

For builders aiming to construct large-scale AI infrastructure, the complexity of delivery speed and scale is becoming apparent: waiting times for generators, transformers, and liquid cooling components are continuously extending, and since NVIDIA's cutting-edge hardware must use liquid cooling, this still experimental technology has many challenges to overcome.

This widespread industry lag is triggering a chain reaction in the capital markets. Recent cases show that not all delays have the same impact; what truly worries the market are issues involving political resistance, financing failures, and the collapse of binding contracts. These factors not only directly impact the stock prices of related companies but also force some capital parties to reassess the economic feasibility of projects.

Meanwhile, industry giants are trying to address this bottleneck at its source. NVIDIA CEO Jensen Huang secretly convened a summit at the company's California headquarters last week, aiming to promote a "behind-the-meter" power generation model, where data centers build their own power generation facilities to bypass the lengthy wait for grid interconnection.

For the industry, the ability to quickly resolve power supply and engineering delay issues will determine who can survive in this AI race.

Political Resistance and Financing Challenges

Some project delays are not due to technical issues but have encountered political resistance from local governments, especially when the scale of data centers jumps from hundreds of megawatts to gigawatt (GW) levels.

Taking a project in Saline Township, Michigan, as an example, the area voted last September against reclassifying farmland for a large AI data center. The developer, Related Digital, subsequently sued the township, and local officials changed their stance a few weeks ago. The project is planned to be built by Related Digital and Oracle for OpenAI, with a total capacity of 1.4 gigawatts. Michigan's energy regulatory agency also approved the expedited power request for the site this week.

However, local opposition has already had substantial consequences. Before construction officially begins, Oracle and Related Digital need to arrange over $10 billion in project financing. According to insiders, investment firm Blue Owl decided last week not to participate in the investment. Blue Owl is concerned that the leasing and debt terms of the project are less attractive than other Oracle-supported projects, and the persistent local resistance may lead to further delays.

Although Oracle stated that Blue Owl's decision did not affect the timeline and that there may be other financing channels, the news still caused Oracle's stock price to drop by 5%

The Failure of Non-Exclusive Agreements

Another common form of delay occurs when developers and clients have only signed a letter of intent or are in an exclusive negotiation period, but the client ultimately decides not to sign a formal lease.

Fermi, a data center developer co-founded by former Texas Governor Rick Perry, recently encountered such a setback. The company disclosed in regulatory filings that its AI server park in West Texas did not sign a lease with tenants before the deadline.

According to Business Insider, the tenant that ultimately withdrew was Amazon Web Services (AWS). Following this news, Fermi's stock price plummeted by over 40%.

Such situations are not uncommon in the industry. Earlier this year, Microsoft decided not to sign any more data center agreements with CoreWeave in Texas; several tech giants, including AWS and Microsoft, have also abandoned plans to lease facilities developed by Applied Digital in North Dakota. Since lenders typically only begin disbursing funds after developers hold binding agreements with top clients, a mere letter of intent cannot secure billions of dollars in funding.

Contract Breaches and Economic Damage

Compared to the above situation, more catastrophic delays involve those projects that have already secured funding and are further along in construction.

According to insiders speaking to The Information, a few AI data center projects have faced frequent issues, leading clients to withdraw from binding agreements before completion.

This situation directly threatens the survival of the projects. In some extreme cases, due to severe delays, developers have had to propose halving the rental fees for NVIDIA servers as a concession. Industry veterans point out that data center contracts typically favor tenants; if the facilities fail to meet deadlines, clients often do not have to pay full price and have the right to withdraw directly. When clients receive significant price discounts due to delays, the economic viability of the project is severely impacted.

For investors, the core issue is no longer whether the project will be delayed, but whether the team behind it has the capability to solve problems and deliver.

NVIDIA's "Secret" Power Summit

In the face of infrastructure bottlenecks, NVIDIA is taking more proactive measures. Last week, Jensen Huang convened about 100 executives at the Santa Clara headquarters for a closed-door meeting focused on the core constraints of the AI boom: power.

According to insiders speaking to The Information, attendees included influential companies in the global energy and electrical infrastructure sectors, such as Schneider, General Electric, Hitachi, and Siemens Energy, as well as some cloud service providers. Huang reiterated that AI has transformed data centers into "factories"—input power, output intelligence. The message from NVIDIA is clear: whoever can quickly increase power supply will seize the next wave of AI demand.

The summit particularly emphasized the necessity of "behind-the-meter" power. Companies are no longer waiting months or even years for public grid interconnections but are turning to on-site generation. For example, OpenAI's facility in Abilene, Texas, employs this model In addition, Jensen Huang also called on the power industry to improve the energy efficiency of data centers to increase the number of AI tokens generated per watt of electricity.

Other Dynamics in the Infrastructure Sector

In the data center and power infrastructure sector, industry consolidation and new actions are frequent:

• Google Acquires Intersect: Google's parent company Alphabet announced it will spend $4.75 billion to acquire Intersect. The company has been assisting Google in finding land adjacent to high-quality wind, solar, and battery power sources, known as "energized land."

• Regulatory Calls: Vermont Senator Bernie Sanders has called for a pause on the construction of AI data centers.

• New Project Collaboration: Texas Pacific Land Corporation announced a strategic agreement with Bolt Data & Energy to develop large-scale data center parks. Bolt was co-founded by former Google CEO Eric Schmidt