Dear Aventine Readers, 

Welcome to our new, more frequent newsletter. As of today, we’ll be sending weekly updates. You can expect the same in-depth look at the way developments in science and technology are affecting us and the world we live in, but we’ll be keeping things a little shorter and — we hope — snappier. We hope you enjoy it.

We’re also introducing some new features, including one we’re sharing today: Views from Substack. In recent years, Substack, the self-publishing platform, has become a major source of independent thinking and analysis, and its writers are increasingly shaping mainstream discourse. But there’s too much out there for most people to keep track of. So once a month we’re going to highlight voices that represent a range of thought on a particular topic with the goal of keeping you up to date on emerging ideas. In this issue we’re looking at the AI bubble. Are we in one or not? And what does it mean either way? Substackers went deep, as you’ll soon see. 

Before we dive in, I’d like to introduce you to Jamie Condliffe, our newly appointed editor at large. Jamie has been the primary writer of this newsletter since we launched over two years ago, and he’s now taking on a bigger role at Aventine, contributing editorial ideas and taking on larger writing assignments. Jamie has a background in science, technology and business journalism, with a PhD in engineering from the University of Oxford. He’s previously worked at The New York Times, MIT Technology Review, Protocol and Sifted.

If you have any feedback about what we’re doing, particularly around the changes we’re making to the newsletter, please reply to this email or write to us at dmattoon@aventineinstitute.org. We also love to hear from you on topics you’d like to see covered. 

Sincerely, 

Danielle Mattoon 
Executive Director, Aventine

Subscribe

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

You know people are starting to worry about darker economic times when you find yourself reading about tulips. So it was on Substack over the past weeks, where certain commentators wheeled out the classic example of the tulip mania that gripped the Netherlands in the 1630s. Back then, money piled into tulip bulbs, the market collapsed, and many wealthy merchants were left holding the bag in what’s often said to be the first ever recorded asset bubble.

This time it’s not tulip bulbs but the infrastructure required to build out the future of AI. And it’s not just wealthy merchants at risk, but potentially the global economy. On Substack, writers whoh focus on technology and economics shared their very many views on how likely such a scenario might be and how it could all play out.

To establish the stakes here, it should be noted that sums being discussed are monumental, with UBS predicting that the big tech firms are expected to spend $375 billion on AI infrastructure this year alone. Putting this in historical context, Paul Kedrosky, who writes a Substack about complexity in the financial markets, crunched some numbers and calculated that AI capital expenditure in the US this year sits at around 1.2 percent of the nation’s GDP, a greater proportion than the 1 percent invested during the peak telecommunications infrastructure buildout around 5G and about a fifth of the 6 percent invested in railroads during the frenzy of that sector’s rapid growth during the 1880s. “These are very large numbers given that before 2022 AI capex was likely less than 0.1 percent of GDP,” Kedrosky wrote. “It has, in three years, grown by at least 10x from there, and perhaps more than that.” 

Putting it in another context, Paul Krugman, winner of the Nobel Prize in Economics and former New York Times columnist, was early-ish in pointing out the warping effect of all this spending, which, he wrote in August, “accounts for about half of US economic growth in the first half of 2025. Without that surge we’d probably be looking at an economy at stall speed, that is, growing so slowly that it could easily slip into recession.” 

Where is all the money to build these data centers coming from? Kedrosky, in the same post, did a useful deep dive and listed a number of sources. One is the huge cash flows inside companies like Alphabet, Amazon, Meta and Microsoft. Another is debt issuance, in which those big tech companies issue bonds to secure cash, and yet another is venture capital and private equity investments, which allow smaller companies to build out AI infrastructure. Over time, the methods of raising capital may become more exotic: After OpenAI CEO Sam Altman said that we could expect his company to “spend trillions of dollars on infrastructure in the not very distant future,” Dave Friedman, who writes about the intersection of AI and finance, postulated that financial instruments based around computing power might be necessary to help finance this huge infrastructure buildout. He was so taken by the idea that he followed up with several posts about what might lead them to fail and how exactly they could work.

Meanwhile, former Bloomberg Opinion columnist turned Substacker Noah Smith has been chewing over all of this to determine how worried we should be. His conclusion: “Some of the basic conditions of a financial crisis are at least starting to fall into place.” Smith’s analysis goes roughly as follows: A slice of the financial sector known as private debt buys up debt such as that from the big tech companies’ bonds, partly using money borrowed from banks. The private debt sector is growing in size, and it is almost certainly exposed to the AI infrastructure build out. That means that there’s potentially a large amount of risk being passed on to banks, and little way of really understanding quite how overexposed they are. “If all the private credit funds are lending to data centers, then their correlations are probably pretty high — if there’s a bust in AI, a lot of them will go bust at once,” he wrote. “It seems like there’s a chance it could hurt the U.S. banking system.” 

One counterargument was shared by a Substack called Shareholdersunite, which turns stock market research into quick-to-read digests. It pointed out that artificial intelligence could, as a general purpose technology, transform the economy so totally that it pays for all of the early investment many times over. If that happens, worry about a financial crisis might feel laughable. But we don’t know if that’s guaranteed, nor do we know whether it will happen within the most helpful time frame. As Smith noted, the railroad infrastructure buildout wasn’t a mistake, but it was premature, and the excessive spending combined with the fact that the system wasn’t used at full capacity until years later helped precipitate a financial crash. 

Nate Jones, meanwhile, who has worked in product leadership roles at tech companies and now writes about AI strategy, argued that concerns about AI returns lagging investment are overblown, driven by a string of recent events and news stories that stoked such fears. Among those events and stories: the coverage of OpenAI’s GPT-5 rollout as botched and underwhelming, an MIT study that claimed 95 percent of all enterprise AI projects are failing to provide return on investment, Meta’s decision to restructure its AI team, and Sam Altman’s “Yes” when asked if we are in an AI bubble. Jones argues that too much has been read into it all and that there may be more reason for optimism than pessimism right now. His reasoning:  Significant progress is being made in agentic AI, but it’s not as obvious as advancements in chatbots, so it’s not widely experienced as progress; AI progress isn’t constrained by software performance but by availability of computing power; and, if you read the MIT study in detail, the lack of ROI has more to do with how managers are deploying AI than the technology itself. (We’re going to dive into this topic in a future newsletter, so watch this space.)

Finally and most recently, the Exponential View, a Substack that covers how emerging technologies are shaping our lives and the economy, published a long and quant-heavy post laying out a “practical framework” to identify whether or not we are in a bubble. The framework, put forward by Azeem Azhar and Nathan Warren, boils the question down to five metrics: the health of the economy, industry strain, revenue growth, the state of valuations and quality of capital. The authors then define thresholds intended to indicate whether any of these measures are flashing red. “If two of the five gauges head into red, you’re in bubble territory,” they warn. For now, three of their gauges are firmly green, a fourth is leaning yellow and the final one is leaning red. So according to their framework, there’s no need to worry yet. 

If there’s a broader lesson to take away from all of this, it’s probably one from Shakeel Hashim over on the Transformer Subtack. He argues that many of the historic boom-bust cycles that everyone has been so furiously writing about share a common theme: Yes, they caused financial turmoil, but they also left behind infrastructure that fundamentally changed the world we live in. In the Netherlands, the tulips still bloomed.

Forecasting AI progress until 2040, by Epoch AI

Predicting the progress of AI is notoriously difficult, but the nonprofit Epoch AI takes one of the most rigorous approaches. It has tracked data from over 3,000 AI models from as far back as the 1950s up to today, using metrics that range from test scores to the energy needs of training runs, in an effort to predict what the future of the technology might look like. In this nearly 100-minute podcast, accessible on its Substack, Jaime Sevilla, Epoch AI’s director, and Yafah Edelman, its senior researcher, discuss how they’re forecasting AI progress until 2040 and what could cause it to move faster or slower. Its results are divisive: “Economists are going to point … and say, ‘These people are insane, they say we can get to 10 percent growth a year by 2035,’” says Sevilla. “And then, the AI people, they’re going to look at us and they’re gonna be like, ‘Oh, these people are insane. They’re only projecting 10 percent growth by 2035.’”

China is quietly saving the world from climate change, by Noahpinion

Noah Smith, mentioned earlier, argues here that the sheer scale of Chinese production of green tech — solar panels, wind turbines, heat pumps — is driving costs so low that adoption will become inevitable, domestically and globally. The wrinkle: China is also burning more coal than ever to meet surging energy demand. But the nation’s overall carbon emissions appear to have plateaued or even declined slightly as renewable capacity has soared. If production of green tech continues at this pace, Smith argues, the future he describes — in which “China’s dogged industrial policy and peerless manufacturing prowess [make] green energy so cheap that simple economics are going to take over” across the world — could materialize.

How to Write the AI Action Plan, by Statecraft

Dean Ball, who from April to August was a senior policy adviser at the White House Office of Science and Technology Policy, was the lead author of “America’s AI Action Plan.” (He’s now a senior fellow at the Foundation for American Innovation.) Speaking to the policy-wonk Substack Statecraft, Ball spills the beans on working on the report. The idea, he said, was to move beyond typical AI policy documents that sometimes come up short and instead “identify between two and six specific policy actions that federal agencies can take now — with existing statutory authorities and budget — to meaningfully advance the ball on that objective.” One interesting detail: He used AI to anticipate areas in the plan on which agencies might push back, helping preempt criticism. Ball also shares some interesting predictions on future AI policy, from how political pushback against AI may manifest (potentially via NIMBYism against data centers) to why it might make some sense to sell more AI chips to China (if you do it carefully).

Topology of "China AI" by Concurrent
And China's Big AI Diffusion Plan is Here. Will it Work? from Matt Sheehan's Newsletter

The phrase “Chinese AI” is often used as a blunt catchall. But here’s a 7,000-word topology of China AI from Concurrent as presented by an AI engineer currently working in the country. Among its takeaways: Americans underestimate the scale of China’s AI workforce; China underestimates the quality of top US talent; and state intervention in China is more complex than many think, with tech companies gaining leverage after DeepSeek’s debut. He also points out that while the US boasts a few standout champions (Google, OpenAI, Anthropic), China’s dense cluster of firms may ultimately drive more competition and innovation. For a little more context, Matt Sheehan, a senior fellow at the Carnegie Endowment for International Peace, has a complementary post exploring China’s new policy to diffuse AI through its economy.

How People Use ChatGPT, by Forked Lightning

Harvard economist David Deming, along with researchers at OpenAI, published a working paper for the National Bureau of Economic Research on how people use the company’s chatbot. He’s also published accessible write-ups of the research on his Substack, Forked Lightning. Some interesting details: Usage of the platform took off for all users late 2024 (suggesting a tipping point at which ChatGPT became actually useful); gaps in use by gender and income status are closing quickly; personal usage is growing faster than work usage. As for what people are using it for? Practical guidance (such as tutoring, self-care advice or creative ideation), writing, and seeking information (akin to web searches), in that order. Together, those three uses account for nearly 80 percent of all ChatGPT traffic.