Subscribe

Subscribe to our newsletter and be kept up to date on upcoming Aventine projects

The amount of money pouring into the construction of data centers in the US is so extreme that, according to Epoch AI, a nonprofit that tracks artificial intelligence trends, it accounted for about 0.8 percent of US GDP in the first quarter of this year. At the same time, the business model supporting this investment is unproven, which leads to obvious questions about whether data center investment can be sustained, and what happens if it can’t. 

In order to understand the landscape, we asked finance and data center experts about likely scenarios if the financial fire hose currently funding the data center buildout slows or stops. They described a variety of outcomes. Some projects could vanish before getting started. Others could continue operating at discount rates or be repurposed for alternative computing needs. And still others could become assets for new owners who buy them at knockdown prices. 

The pullback 

One factor determining the fate of any particular data center will be how much of it is already built. 

Projects planned but not yet started could simply be scrapped. The developers might owe kill fees to contractors, but canceling is otherwise likely to be relatively straightforward, said Advait Arun, a senior associate for capital markets at the nonprofit Center for Public Enterprise who published a report on data center financing in late 2025.

Data centers still under construction are another story. Will their abandoned remains be left to decay? Chances of that are slim, said Arun, because many of the contracts for data centers currently under construction have been written specifically to prevent exactly that outcome. Take Meta's $50-billion, 5-gigawatt Hyperion campus in Richland Parish, Louisiana. Meta is building it with the asset manager Blue Owl Capital, which owns 80 percent of the project — part of an arrangement that allows Meta to avoid assuming a large amount of debt. Another part of the arrangement is a commitment from Meta to a 16-year residual value guarantee to pay Blue Owl the difference between the facility's market value and a privately agreed-on value if it decides to give up the lease for any reason. Oracle has similar obligations with its 2.45-gigawatt facility in New Mexico. The result is that there are strong disincentives to abandon data centers before completion, though other measures could be taken to minimize cost.

If necessary, data center developers might also stretch construction timelines or shrink the size of the final project, experts told Aventine. Large AI data centers are not monoliths, but a collection of many buildings, each subdivided into so-called data halls that house servers and have their own power supplies and cooling systems. One way to cut back would be to slow the rate at which these halls are built and put online, said Alex Cordovil, research director for data center physical infrastructure at the market research firm Dell'Oro Group. 

And projects could be shrunk in other ways, said Andy Lawrence, executive director of research at Uptime Intelligence Institute, a market intelligence firm that studies the data center industry. For example, a planned gigawatt facility could be scaled back to a few hundred megawatts. The company building these facilities are “not totally stupid,” he said. “They're not just piling this money in.”

Repurposing

Then there are facilities that are built, energized and ready to go, some with chips already installed, others without. This is where technical details about how these data centers actually work become important, as they determine how the infrastructure can be used — or repurposed — in the future.

If the shell and supporting infrastructure are in place but the chips are not, repurposing a data center so that it isn’t purely focused on AI is more plausible than if chips had been installed. The building could easily be used for more generic cloud computing tasks — from serving up video on demand to running the cloud software we all use every day. The power and cooling would be over-specified for such conventional workloads, but over-specified equipment still works, said Cordovil.

In data centers already outfitted with chips (which account for 60 percent of the upfront data center costs, according to Epoch AI), things are more complicated. If the buildout of AI ends up creating an excess supply of the computing capabilities data centers supply, the likely first response would be for data center operators to cut their prices, said Cordovil, ultimately meaning that end users would be charged less per token. Then the companies would hope Jevons paradox kicks in, with cheaper prices increasing demand and stabilizing the market — albeit at a lower margin. 

If for whatever reason that option is off the table, repurposing is complicated. Data centers built to train AI models use graphic processing units, or GPUs, almost exclusively. These are expensive and consume far more power and create far more heat than the central processing units, or CPUs, used in conventional cloud computing data centers. The chips must also be packed tightly together so they can share data quickly, which requires special heavy-duty power supplies and specialized liquid cooling systems. These requirements make facilities already stocked with GPU chips difficult to repurpose, though if they had to be, they’d be best suited to high-performance computing tasks such as complex scientific AI workloads performed by national labs and research universities, or even cryptocurrency mining. In a pinch, data centers used for AI training could also be used to perform some AI inference — the process of running the models for customer tasks. 

Meanwhile, facilities specifically built for inference are more versatile. They use GPUs to run models but lean heavily on CPUs for other jobs, like user authentication, data preprocessing, tokenization and delivering responses to clients. The combination means that they can be used for other types of compute-heavy workloads beyond AI, said Daniel Bizo, research director at Uptime Institute. That might include applications such as scientific research, financial analysis, cloud gaming or other processes that involve intensive data processing. 

Worst-case scenario

The problem in all of these situations is that expensive infrastructure would be making far less money than it was built for. 

The financial arrangements behind the data center buildout are known for being opaque and interdependent. Though we still don’t understand the full complexity of relationships between chipmakers, cloud operators, investors and AI developers, many critics have argued that their interdependence means that if part of the industry fails for whatever reason — a bankruptcy, say — contagion could spread through the market quickly with destructive downstream consequences.

In the face of plummeting stock prices or even bankruptcy, some data centers could be sold off as distressed assets for far less than it cost to build them. Lawrence pointed out that the GPUs within them are typically useful for only five or six years, which means they are viewed as short-term assets. That means data centers might be carved up, with the GPUs auctioned off and buildings sold separately. Ultimately, it may be that the data center’s grid connection — an asset for which there is currently a five-year waitlist in the US — ends up being the most valuable part of the site, said Arun. 

Data centers might also be carved up, with the GPUs auctioned off and buildings sold separately, said Lawrence. Ultimately, it may be that the data center’s grid connection — an asset for which there is currently a five-year waitlist in the US — ends up being the most valuable part of the site, said Arun. 

Broader fallout

In some parts of the US, anything that disrupts the data center buildout could be met, at least initially, with a sigh of relief: A recent Gallup survey found that seven in 10 Americans oppose local AI data center construction. 

Still, many local politicians have made promises about the economic benefits contingent on these facilities, hoping to see the kind of benefits enjoyed by residents of Loudoun County, Virginia. Sometimes called Data Center Alley because of the density of data centers there, Loudon County benefits from lower property taxes and desirable public amenities like well-funded schools and parks. But Loudon is unusual. It has been home to data centers since the dot-com boom and Virginia laws allow the computer equipment inside data centers to be taxed — a highly lucrative provision expected to bring the county $1.3 billion in revenue next year, according to The New York Times.

As a result, it’s not clear if the benefits Loudon has enjoyed will be broadly replicated elsewhere. (A study published this year by researchers from San Diego State University and others found “little evidence” of data centers leading to improvements in local economic outcomes.) In which case, local governments could be on the hook if “they’re considering data centers to be their primary economic development strategy,” said Arun.

Energy might be a larger problem. The AI boom has prompted some utility companies to contemplate building substantial new infrastructure, including gas-fired power plants, to accommodate growth from data center consumption. If demand doesn’t arrive, the cost may be passed on to domestic ratepayers, said Arun. Another wrinkle is that data centers are currently propping up bets on exotic energy technologies like small modular nuclear reactors, advanced geothermal and long-duration energy storage. If the AI market does soften, said Arun, “we'll have to think more seriously about what we actually need to do to de-risk these technologies.”

Whatever happens to the AI market, data centers have already changed the landscape of the US and aren’t going anywhere. The question is whether they remain uniquely tied to AI, or will be adapted to serve broader infrastructure needs — a process that would likely go hand in hand with a reshaping of the nation’s economic landscape. 

Amazon has struck a deal to upgrade the Calvert Cliffs Nuclear Power Plant in Maryland. John Swarey / Wikimedia Commons.

Next up for nuclear: Squeezing more juice from old plants. As nuclear energy’s renaissance continues, the latest push is to upgrade existing plants with the goal of increasing output by 20 percent or more. The New York Times reports on a string of announced initiatives along these lines from energy and tech companies in recent weeks: Constellation Energy will spend more than $4.3 billion to refurbish 11 reactors at six plants in Illinois, Pennsylvania and New Jersey; Vistra will upgrade three nuclear plants in Pennsylvania and Ohio; Amazon has struck a deal to upgrade the Calvert Cliffs Nuclear Power Plant in Maryland; and Google is investing in upgrades to four reactors in Georgia. The capacity of about three-quarters of the US fleet of nuclear reactors could be upgraded, experts told the newspaper, adding as much as 6,000 to 8,000 megawatts of capacity to the US grid, equivalent to six to eight new large reactors. One big sticking point is that major upgrades can require plants to go offline for months, leading to lost energy production and revenue, so minimizing downtime will be a key challenge.

The debate over AI consciousness is going mainstream. The idea that artificial intelligence could develop consciousness has always been controversial, but it’s usually been a back-burner idea discussed among philosophers and a small subset of AI researchers. In recent weeks, though, the conversation has jumped into the mainstream. The inciting event was the publication of a study by Anthropic in July that identified a set of internal activities inside the company’s Claude models that some researchers say meet their first criterion for machine consciousness. Things have gotten weirder since then. The New York Times reported that, starting last fall, Anthropic co-founder Chris Olah has been meeting with religious scholars to raise the possibility that Claude could be conscious. (Pope Leo XIV has formally rejected the idea.) And Anthropic announced that, from November 12 of this year, sustained and needless cruelty toward its models will violate its usage policy. Meanwhile, Geoffrey Hinton, one of the so-called godfathers of AI, has repeatedly said that he believes today’s models are already conscious. Academics, too, are wrestling with this. In August, aboard a yacht in the Galápagos, philosophers and neuroscientists vigorously debated whether chatbots can be conscious, reported Nautilus. The sessions reached no verdict, according to Wired, with the deepest disagreement being over what kind of evidence could ever settle the debate. But not everyone views this as an open question. In September, Mustafa Suleyman, chief of AI at Microsoft, argued that training Claude on documents that treat its consciousness as an open question creates an “epistemic hall of mirrors” in which an AI becomes convinced its welfare is under attack. The resolution of this question — if indeed it is resolvable — will have major consequences for how society adapts to and adopts the technology. How will we navigate a world divided by those who think AI is capable of consciousness and those who do not?

There’s impressive work in OpenAI’s recent dump of AI math proofs. It was hard to miss the headlines about the 722 AI-generated math papers shared by OpenAI on October 6th. The announcement sent new shock waves through an already panicked community of mathematicians — about how best to determine whether the results were correct, whether humans would ever truly grasp exactly how the models had tackled all these problems and what it means for the future of mathematics more broadly. Have some sympathy, for instance, for Constantin Kogler, a researcher at the Institute for Advanced Study. He and a colleague cracked the so-called exact overlaps conjecture, a major problem in fractal geometry, just 10 days before OpenAI released its AI-derived proof of the same problem. But beyond panic and frustration, there is another, more pragmatic, concern: It’s going to take a long time to work through all of the OpenAI papers, where more breakthroughs likely await. So far, the most consequential result appears to be a significant leap forward in solving the Riemann hypothesis, often considered the most important unsolved problem in mathematics. Science reports that James Maynard — a professor at the University of Oxford, a winner of the Nobel equivalent Fields Medal and a Riemann hypothesis specialist — called the advance “groundbreaking,” adding that in analytic number theory it is “the best headline result for the past 100 years.” There is also a proof of the so-called 3D Kakeya maximal conjecture, which mathematicians view as a stronger version of the result that helped Hong Wang, a mathematician at NYU, win this year’s Fields Medal. Of the papers that have already been reviewed, there are at least nine more that mathematicians consider impressive. There is still work to be done in reviewing and fully vetting all of the proofs, but if the breakthroughs hold up, they provide further evidence that AI has moved beyond humans as far as math is concerned.

Are AI models easier to understand than we realized? The AI world is grappling with a new idea: It might be easier to interpret how large language models work than previously thought. In a 26,000-word essay, Nick Cammarata, a researcher at a company called Goodfire that is focused on AI interpretability, explained that he thinks the theories that came to define why LLMs are so difficult to understand may be flawed. Fundamental to that perceived complexity is the idea of superposition: that neural networks, such as large language models, cram unrelated concepts into single artificial neurons, making it difficult to discern how the larger network works. Cammarata, it’s worth saying, has some standing here: He helped develop the very idea he’s questioning, and also started a project at OpenAI that used AI to try to understand how LLMs work. Cammarata’s new take is that superposition isn’t happening to a significant degree in large language models. His contention is that when humans observe a neuron responding to seemingly unrelated things, what’s actually happening is that people are failing to spot a similarity recognized by the neuron. An example from image recognition would be a neuron that responds to both cats and cars, which has long been cited as evidence that neural networks are uninterpretable. But Cammarata argues, for example, that such a neuron is detecting a distinctive V shape in some images. Large language models are harder to study than image recognition models, but his preliminary analysis suggests that the behaviors of most of the neurons in LLMS make sense too. He’s not claiming to have cracked interpretability, but to have found evidence to suggest it may not be as hard as we thought. As MIT Technology Review reports, it’s unclear if he’s correct. Cammarata himself calls it a working hypothesis, and while many researchers are excited by the work, others are deeply skeptical. Some argue it isn’t even a new idea: Oren Etzioni, a renowned computer scientist who spoke with MIT Technology Review, said it echoes an account of neurons put forward by Geoffrey Hinton and colleagues back in1986.

China’s Growth Machine, from Works in Progress

China’s economic success in recent decades is often attributed to a top-down approach, with the central government calling the shots. This story challenges that idea, arguing that the central government is more of a rule setter, and that in reality, it is local governments across the nation that are responsible for the country’s explosive growth. Subnational governments account for around 85 percent of public expenditure in China, and local officials — rewarded with promotion if they demonstrate sustained economic growth — are more akin to company directors than to the civil servants of the West. As a result, they compete fiercely for companies to set up shop in their backyards, through tax relief, subsidies, cheap land and loans, accelerated permitting and so on. It is this model of regional championing that has led to some of China’s biggest success stories: telecoms company Huawei, car maker BYD, drone maker DJI and Vivo, the biggest supplier of smartphones to India, were all nurtured by local governments. The essay points out that this way of thinking was commonplace in the US in the 1950s and 1960s, when citizens benefited from growth through city services, and argues that a return to local pro-growth policies are what’s necessary for America’s growth to keep pace with China’s. But it fails to wrestle with an important reality: US voters today often reject proposed development, seeing it as a threat to property values, which would make this sort of growth far harder to achieve.

Bending the Curve of Discovery: AI in Science Today and Tomorrow, from DeepMind Institute

Google DeepMind recently launched the DeepMind Institute, a platform for its researchers to explore ideas about a future world in which artificial general intelligence (AGI) is a reality. This is one of the early essays, drawing on the findings of research into how scientists are making use of AI from the following sources: analysis of 15 million interactions with Gemini, Google’s AI assistant; data on more than 2,600 specialized scientific AI models; and a survey of more than 600 scientists. Some high-level findings: Scientists are overrepresented among Gemini users relative to their share of the workforce. They tend to use LLMs for generic tasks like writing and coding but use specialized AI models for domain-specific analysis and modeling — and there’s little overlap between the two. On the plus side, scientists claim to save almost seven hours a week thanks to AI, most of which they plough back into research. In the less helpful column: Around 40 percent of scientists say that their backlog of untested hypotheses is growing as their ability to generate ideas outpaces their capacity to do practical work. Almost half say AI directs them toward more incremental questions rather than novel, potentially more groundbreaking, research. The opportunity, the piece argues, is in figuring out how to capitalize on the gains AI can deliver and minimize the negatives. As for how to achieve this, many of the proposals suggested in the article have less to do with fixing the technology itself and more to do with reforming science as an institution, which is perhaps unsurprising coming from Google DeepMind.

They War-Gamed Dimming the Sun. It Went Even Worse Than You Think, from Wired

You might hope that if you asked highly credentialed experts from across the world to game out how governments should approach the controversial act of solar geoengineering the sun, wisdom would prevail and the various actors would figure out how to make decisions in humanity’s best interests. You would be mistaken, at least according to this story. The idea of solar geoengineering — a controversial plan to mitigate global warming by reflecting sunlight away from Earth — has leaped to the fore recently, with one startup developing technology to release aerosols in the atmosphere at scale to mimic the effects of sunlight-blocking volcanic clouds, and Elon Musk chiming in with support for the approach. But during the war-gaming of how global leaders might respond to the use of such technology, political manoeuvring and sovereignty concerns won out over scientific evidence. Those representing the US were so deeply mistrustful of their China-representing counterparts that they torpedoed any prospect of an international deal on policing the technology, convinced that their rivals were secretly testing such systems. (They weren’t.) Obviously these sorts of exercises don’t predict what will happen in real life, but they do give us a sense of how difficult it might be to agree to rules for such technology before a race to test and deploy it takes hold.