Recently, the well-known Wall Street research firm Bernstein conducted a study, and the results were quite surprising.

The cumulative planned data center capacity already announced in the U.S. is as high as 488GW. But only about 170GW can be confidently delivered, accounting for just 35%.

In other words, nearly two-thirds of the planned data centers in the U.S. may or may not be built in the future.

What's more interesting is that the reasons these projects cannot be delivered are quite varied.

Some developers treat data centers directly as real estate: before the project even breaks ground, the land is already listed for sale; others finally start construction, but clients keep changing the design, forcibly increasing the construction budget by nearly $800 million.

Previously, these problems were not a big deal. After all, AI is the hottest investment theme right now. As long as you dare to announce a few GW of data center plans, Wall Street is willing to give a high valuation, and banks and bond investors are willing to put up money.

But now, the situation is starting to change.

On one hand, the Federal Reserve raised rates again in September, tightening the financing environment; on the other hand, more and more data center projects are exposing problems such as construction delays and cost overruns, making investors start to re-examine the risks involved.

Obviously, even Wall Street can no longer stomach the data center pie.

/ 01 / Two-thirds of U.S. data centers are in doubt

Why are two-thirds of U.S. data centers in doubt? The problem mainly lies in two aspects.

One is on the planning side: more and more super data centers are falling into the hands of novices with no construction experience.

BloombergNEF statistics in July showed that among the 100 largest data center plans in the U.S., 52 come from developers with no construction experience, totaling 87GW.

Why are these new players so bold?

The reason is simple: in the U.S. data center business, you do not necessarily have to build the machine rooms to make money.

For some land developers, as long as they acquire land in advance, push approvals, and secure power access, they have a chance to package the land as a data center project and sell it at a high price to real operators.

To put it plainly, it is land speculation, except this time what is being speculated is the AI concept.

The La Osa project in Arizona is a typical case.

In July 2025, Vermaland, a land developer with no data center construction track record, directly announced a $33 billion super project, planning 59 buildings and 3GW of power demand.

How exaggerated is this scale?

As of the end of June this year, CoreWeave's 51 operating data centers had total active power capacity of only 1.5GW.

That is to say, Vermaland's first foray into data centers planned power capacity equivalent to twice CoreWeave's existing operating scale.

But what is more interesting is that in May this year, a Vermaland representative admitted at a government hearing that the project land had already been listed for sale and was preparing to be sold to the end user.

The previously announced $33 billion super plan instead became the best advertisement for selling the land. On the LoopNet land sales page, Vermaland even used "poised to become the largest data center in the U.S." as a selling point.

For this type of developer, the larger the plan, the better the chance the land can be sold at a high price. As for whether it can be built in the future, that is something the buyer needs to worry about.

Similar phenomena are becoming more and more common, and the industry has even coined a sarcastic term: Bragawatts, that is, "bragged gigawatts."

But inflated planning is only one aspect. What is more troublesome is that even if construction actually starts, delivery may not necessarily be on time.

Goldman Sachs research in May this year showed that the historical on-time delivery rate of U.S. data centers is about 72%, but the expected on-time delivery ratio for planned capacity one year and two years ahead may drop to 60% and 50%, respectively.

The reason is not hard to understand. Data centers require large amounts of supporting facilities such as power, substations, and cooling systems, and if any link falls behind, it can slow down delivery.

Moreover, even if these problems are solved, developers still have to face swelling construction costs.

The Barber Lake project in Texas is a typical case.

This project is developed by Cipher Digital, which transformed from a Bitcoin mining company. It plans total power supply capacity of 300MW, has already signed a long-term lease with AI computing company Fluidstack, and Google has also provided partial guarantees.

The client is there, the contract is signed, and by all accounts it should be a high-quality project.

But during construction, the client continuously raised new design requirements, causing repeated adjustments to the engineering plan, and the lease commencement time for some machine rooms was also delayed to the first quarter of 2027.

What is more troublesome is that the project's estimated total completion cost has risen from $2.157 billion to $2.941 billion-$2.966 billion.

Before completion, the budget has increased by nearly $800 million, an increase of more than 36%.

Unfortunately, not all of the extra money can be borne by the client.

According to the agreement, the first approximately $359 million exceeding the original budget is borne by Cipher, and for the portion above that threshold, the client only shares 50%, to be gradually compensated through additional rent over the next 20 years.

This is quite awkward. The original idea was to lock in revenue through a 20-year long-term lease, but when it actually came to execution, before the rent was recovered, the company was bogged down by wave after wave of construction payments.

In the final analysis, the problem with U.S. data centers is that the pace of expansion driven by the capital market has far exceeded the real world's construction capacity.

/ 02 / Wall Street is also starting to recalculate

Over the past two years, one important reason why U.S. AI infrastructure investment has been able to expand wildly is that Wall Street was willing to keep providing funds. But now, Wall Street is becoming less willing to pay for these stories.

The most direct manifestation is falling stock prices.

According to Kpex.AI statistics, from October 6 to 20, eight AI infrastructure companies including CoreWeave, Nebius, and IREN saw a combined market value evaporation of about $51 billion, with the median stock price decline reaching 28%.

Among them, CoreWeave fell nearly 45% at one point. According to Reuters, short sellers alone made about $1.7 billion in paper profits on this company.

The once wildly sought-after AI infrastructure stars have now instead become cash machines for short sellers.

This change has already spread to the IPO market. In October this year, Firmus, an Australian AI data center company backed by Nvidia, originally planned to raise about $5 billion in an IPO, but some investors withdrew their subscription intentions before listing. On October 9, Firmus ultimately canceled the IPO and turned to private market financing.

One important reason is that its partner CDC Data Centres confirmed that the 1.6GW data center project the two sides originally planned to develop together will no longer proceed.

SB Energy, backed by SoftBank, also encountered a similar problem. The company originally planned to push forward with an IPO in September, but later postponed it to the end of the year or even next year.

In addition to the stock market, the bond market is also changing. According to Trepp, a commercial real estate data institution, as of October 23, among the 64 data center-related mortgage-backed securities issued by the end of last year, 60 had widened credit spreads.

Among them, the credit spreads on data center bonds with the highest AAA rating widened by about 0.6 percentage points during the year, while other similar commercial real estate bonds widened by only about 0.01 percentage points. Simply put, investors are starting to demand higher risk compensation.

What is more troublesome is that new bond issuance is also starting to cool. Morgan Stanley research in October showed that at the beginning of this year, the issuance yield on data center asset-backed securities with the highest AAA rating was still below 5%, but now it is close to 7%.

The experience of U.S. data center developer EdgeCore is very representative. In October this year, the company prepared to issue $521 million in asset-backed securities, but nearly canceled the transaction because of insufficient investor subscriptions.

To complete the financing, EdgeCore had to raise its pricing, increasing the credit spread on the highest AAA-rated bonds from the initial more than 100 basis points to 200 basis points, ultimately completing the issuance with a coupon of 5.90%.

And just last year, the $235 million bond issued by EdgeCore had a coupon of only 4.45% and was oversubscribed.

In just one year, data center financing has changed from investors rushing to buy to needing to raise interest rates to find buyers.

For the data center industry, which is highly dependent on debt for expansion, once this negative feedback loop forms, it may be more fatal than a short-term slowdown in demand.

Text/Yuanyuan