In November, the results of the U.S. midterm elections and a series of slowing growth figures from technology companies in the third quarter brought concerns to Wall Street and the artificial intelligence industry regarding expected market growth and resistance to industry policy.
However, a recent analysis by Bernstein of North American data center projects shows that the potential risks are not only in the future but also in data center projects that have already been initiated. The U.S. artificial intelligence boom has lasted for more than three years, and data centers are undoubtedly one of the most prominent and largest industries in this boom. To date, trillions of dollars have been invested in data center construction across the United States. It is undeniable that many of these data center construction plans have become reality over the past several years and have become an important foundation for the continued growth of the U.S. economy, Wall Street stocks, and a series of technology companies.
However, recent reports show that the potential risks contained in these plans may be as great as the growth they bring. According to reports by the Financial Times and the technology media outlet Business Insider, in a data center credibility model study recently released by Bernstein analyst Madison Rezaei's team, a screening of all 8,858 data center projects in North America showed that out of 586GW of announced, under-construction, and land-bank capacity, only 199GW (34%) is truly credible. This phenomenon means that current forecasts of AI infrastructure demand in the U.S. market are built on a very large scale of "fake orders," and that upstream producers on the demand side may be producing chips or equipment for data center projects that can never actually be realized.
"Bragawatts"
In Bernstein's research, it conducted monthly tracking and credit model analysis of 8,858 data center projects in North America, first publishing on October 11 and further updating after the midterm elections. From a proportional perspective, out of 586GW of announced, under-construction, and land-bank capacity, only 199GW (34%) is truly credible; the remaining 387GW, about two-thirds, mostly belongs to "Bragawatts," meaning data center projects that cannot ultimately be realized.
Factors reducing the credibility of data center projects
From a longitudinal perspective, Bernstein found that the credibility ratio of U.S. data centers remained basically unchanged during the three months of measurement: 33% in the first estimate in July, 35% in August, and 34% in October. Even though the scale of announced data center projects expanded from 411GW to 586GW during this period, the proportion of credible projects still remained at one-third. This means that "fake" projects have become a long-term phenomenon or convention in the U.S. data center industry and will continue to affect the artificial intelligence sector, including data centers, in the future.
In Bernstein's research, data center projects are mainly evaluated from four dimensions to determine the credibility of the project itself, including development stage (excluding 198GW), tenant quality (excluding 113GW), developer track record (excluding 48GW), and state-level regulatory risk (excluding 28GW). It is worth noting that the political factor that has drawn the most attention from Wall Street and data center investors over the past several weeks actually poses the smallest risk to data center projects in the model analysis.
The report stated that projects currently at the "announcement only" stage amount to 497GW, accounting for 85% of the total, and Bernstein estimates that about one-quarter of them can be realized. At present, 75GW of "credible capacity" is under construction, which would more than double existing capacity. If all 199GW of credible capacity is realized, it would be 3.4 times the current scale, but this would be a construction cycle lasting five to ten years.
Among the largest influencing factors, namely development stage and tenant quality, these depend to a large extent on the development capabilities and actual operating capabilities of the development entities. From this angle, Bernstein's research results are shocking.
Hyperscalers' self-built projects and data center colocation providers account for only 37% of announced construction capacity but 76% of credible construction capacity; while other categories of developers announced 308GW of construction, but only 33GW, about 11%, was considered likely to be completed in the study.
Bernstein scored each developer from 0 to 5 across four dimensions: power supply (35% weight), capital (30%), leases (25%), and execution (10%). Among them, hyperscalers, Apple, and blue-chip colocation providers were regarded as 5-point developers; 68% of data center projects from developers scoring 4–5 are expected to be completed, while developers scoring 0–1 together initiated 142GW, about one-quarter of the national total, with only 7GW considered likely to be completed, a survival rate of 5%.
Credibility of data center projects by developer tier
This phenomenon in Bernstein's research is caused by multiple factors, but it is visible that this means a very large proportion of the massive data center construction plans currently supporting U.S. economic growth stems from the driving effect of the past artificial intelligence boom and investor optimism, enabling many developers that did not have, or did not fully have, data center construction capabilities to push projects into initiation and financing.
A long-ignored warning
In fact, warnings about the mismatch between the expected scale of U.S. data center projects and actual construction capacity have been common. As early as the months before the midterm elections and the AI stock selloff, the Financial Times repeatedly reported on the potential risks behind the data center boom.
On September 26, the Financial Times reported a mismatch between general forecasts of chip sales and the number of completed U.S. data centers. Morgan Stanley measured this gap by estimating the available power shortfall and concluded: "More than half of the GPU servers sold between 2026 and 2028 may have nowhere to plug in."
The enormous cost the United States has invested in data center construction
Citing data from SynMax Intelligence, the Financial Times said that satellite imagery tracking of U.S. data center construction sites shows that total capacity that can be put into use this year is between 16 and 18GW. In addition, about 2GW of data center projects are planned for completion in 2026 but have not yet begun land clearing, so they are "impossible" to deliver on time.
On October 16, the Financial Times reported again that investors are losing confidence in the expected revenue brought by the AI boom, believing it is insufficient to support the huge debt financing data centers, which will lead to trouble for bonds supporting some of the largest infrastructure projects in the United States. Bond prices issued to finance well-known data centers such as "Hyperion" and "Excalibur," which host companies including Meta, Nvidia, and Oracle, have already fallen far below face value, and debt related to many artificial intelligence infrastructure groups, including emerging cloud service provider CoreWeave, has been heavily sold off.
On October 18, based on Bernstein's related research, the Financial Times said that more than 1,000 communities in the United States this year have sought to impose moratoriums or other restrictive measures on artificial intelligence infrastructure construction. AI laboratories and large technology groups have pledged to invest trillions of dollars in computing power, or "compute," in the coming years. But these huge returns can only be realized if a group of recently established "neocloud" data center providers first manage to raise funds for this unprecedented construction effort. In addition to these financial challenges, any large infrastructure project will inevitably face various obstacles, and public dissatisfaction with data center construction is also growing. According to data center executives, it is becoming increasingly difficult for the entire industry to obtain building permits, and utility regulators have also been refusing grid connections for projects that cannot be self-sufficient. One executive said: "People no longer donate to data centers as easily as before. Nowadays, financing data centers is much more difficult."
Delayed or canceled data center projects
This series of reports by the Financial Times shows that, due to factors such as project delays, public opposition, and mismatches between input and output, the U.S. data center industry has in fact long been in an unbalanced state. However, during the past several months when the artificial intelligence boom has not yet faded, market optimism enabled data centers to obtain financing far exceeding their construction capacity, and a considerable proportion of developers were unable to convert financing into actual data center capacity. This overlooked potential risk could be dissolved amid market prosperity through continued low-interest-rate financing, but as investor confidence weakens with the midterm election results and the successive slowdown in technology company growth, this risk is bound to be amplified.
Comments