The question of whether there's an AI bubble and what, if anything, to do about it, keeps recurring. This time, it's a thorough report by Advait Arun, for the good people at the Center for Public Enterprise.

His short answer: yes, it's a bubble, and it's going to hurt when it pops.

Arun’s starting point is straightforward: AI data centers are not magic; they are a hybrid of commercial real estate and infrastructure. On the real estate side, developers borrow to build “powered shells” and then rely on long-term leases to repay that debt. On the infrastructure side, those shells are stuffed with GPUs, megawatts of power connections, cooling gear, and fiber. Both sides of the balance sheet have to line up over decades.

The report walks through that balance sheet with a series of T-charts and tables, showing how assets and liabilities sit with developers, tenants (“hyperscalers” like Microsoft, Amazon, Meta, and OpenAI), GPU manufacturers (NVIDIA, AMD), lenders, and end-users. The picture that emerges is not one of steady, utility-like cash flows but of a chain of mismatches: Long-lived data-center term loans (20–30 years) paired with much shorter tenant leases; GPUs that economically depreciate in 3–4 years, now being replaced on a yearly product cycle; Massive capital expenditure whose only ultimate support is uncertain consumer and enterprise demand for AI services.

Arun then overlays a Minsky-style framework: as long as credit is easy and equity valuations are soaring, everyone can roll their debts and pretend these mismatches don’t matter. Once growth expectations break, refinancing risk cascades down the system—from tenants, to developers, to utilities, and finally to households and state budgets.

The report offers several concrete reasons to think we are closer to "bubble" than "sustainable boom." Namely:

Revenue is nowhere near capital expenditures. Using public statements from Big Tech, Arun estimates roughly 560 billion dollars of AI and data-center investment in 2024–2025 against only about 35 billion dollars of AI revenue from the main hyperscalers. The gap is being financed by stock-market enthusiasm, not customer demand.

Inference is commoditized and margin-squeezed. The actual money in AI is supposed to come from "inference," that is to say: serving model outputs to users. But inference providers mostly offer similar services, often using the same underlying models and GPUs. They are forced to compete on price, while their costs are rising as prompts get longer, models do multiple passes to avoid errors, and traffic increases. So even if usage soars, profits don’t necessarily follow.

GPUs are a shaky form of collateral. The keystone asset of the boom is the high-end GPU. Hyperscalers use these GPUs as collateral to finance their acquisitions. Yet NVIDIA and AMD are moving to annual product cycles, which instantly mark down older chips. Neocloud firms are borrowing against their GPUs at high rates, then seeing the collateral value erode faster than the loan term. The report compares this to paying off one credit card with another: it works until it doesn’t. One figure shows firms lengthening official depreciation schedules to 5–6 years in their accounting even as the real economic life may be half that.

Circular financing, or "roundabouting". The report’s most striking section details the web of cross-investments and purchase commitments tying NVIDIA, OpenAI, Microsoft, Oracle, CoreWeave, and others together. NVIDIA takes equity in CoreWeave; CoreWeave buys GPUs from NVIDIA and sells compute to Microsoft and OpenAI; NVIDIA promises to buy back unused CoreWeave capacity; OpenAI signs multi-hundred-billion-dollar contracts to rent that capacity; Oracle borrows to build data centers to host it, and so on. The sector is effectively using its own future revenues as collateral for itself.

Debt is migrating into opaque channels. While the big public tech names look conservatively levered on paper, the report shows that a growing share of risk is being pushed into off-balance-sheet special-purpose vehicles, junk bonds, and private-credit vehicles. These are in turn backed by GPU collateral and tenant leases that all depend on the same fragile demand story.

These are all worrisome signs. The point about GPUs, and about "roundabouting," particularly feel worrisome.

However, the idea that it's a bubble is by no means certain. Our best analogy for the impact of AI comes from Peter Thiel, who has likened it to a new internet: a technology which will create trillions of dollars of economic value and affect every part of the economy and of our lives, albeit one which will not take us to some sort of post-scarcity Star Trek utopia. We think that's correct.

The real question is whether future demand for AI will be large enough to justify all this investment. And it's very likely that it could.

Every serious forecast from the hyperscalers assumes not just today’s chatbots, but the progressive colonization of the entire software stack by AI: copilots embedded into every workflow, AI-native applications that do not exist yet, and, eventually, models that are not just stochastic parrots but tools tightly coupled with databases, robots, and physical processes. If that world materializes even halfway, then today’s numbers are not evidence of irrationality but of a J-curve. Railroads, electrification, mobile broadband all went through a phase where capital spending vastly overshot contemporaneous cash flow. In hindsight, the mistake was usually not that investors built too much infrastructure, but that regulators and incumbents underestimated the downstream productivity shock.

There is a perfectly coherent story in which what looks like a bubble is simply the ugly, front-loaded investment phase of a genuine platform shift. Every serious forecast from the hyperscalers assumes not just today’s chatbots, but the progressive colonization of the entire software stack by AI: copilots embedded into every workflow, AI-native applications that do not exist yet, and, eventually, models that are not just stochastic parrots but tools tightly coupled with databases, robots, and physical processes. If that world materializes even halfway, then today’s numbers are not evidence of irrationality but of a J-curve. Railroads, electrification, mobile broadband all went through a phase where capital spending vastly overshot contemporaneous cash flow. In hindsight, the mistake was usually not that investors built too much infrastructure, but that regulators and incumbents underestimated the downstream productivity shock.

The other point that cuts against the bubble narrative is that these data centers are not tulip bulbs, they are multi-use industrial infrastructure. A megawatt of firm power and fiber to a site does not care whether the racks inside are running frontier models, basic cloud workloads, scientific computing, or industrial simulations; GPUs that are no longer cutting-edge for training the latest foundation model can still be perfectly economical for serving slightly older models, recommendation engines, or enterprise inference. In that sense, the system is much more resilient than the stylized charts suggest: capacity can be repurposed, workloads can be degraded, prices can adjust. Even if we do see a painful repricing of AI equities or some spectacular failures among over-levered neoclouds, that is closer to a rotation within the tech sector than some sort of wholesale crash. The underlying assets will find uses because computation, unlike subprime mortgage CDOs, corresponds to a real and steadily growing demand curve.

Discussions of bubbles can take on a metaphysical bent. In some sense, the story of Western civilization, or of human life itself, is a "bubble." All civilization is like a bike, it falls over if you stop pedaling, its course is always presuming some future functioning to keep it going.

Looked at from the vantage point of 2025, the dotcom bubble and crash, and the accompanying telecom bubble and crash, look like a relatively modest pullback at the beginning of a very long boom.

It seems indisputable to us that AI will be at the center of all our lives and all business applications a few years from now, and that this represents a very large business opportunity indeed.

With all that being said, of course, while one should hope for the best, one should also prepare for the worst, so we should still pay attention to Arun's report.

He warns that state and local governments are already heavily exposed. At least ten states are forgoing more than 100 million dollars per year in tax revenue via broad data-center exemptions, often for decades, even as they become fiscally dependent on tech-sector capital gains.

On top of that, the federal “AI Action Plan” and the recent push for expanding CHIPS-style tax credits to AI data centers and even GPU purchases would, in effect, he argues, invite Washington to backstop the bubble. A federal guarantee on GPU-backed loans, as OpenAI has openly floated, would not be a neutral industrial policy; it would hand NVIDIA and AMD a blank check to accelerate product cycles and, he thinks, dump the depreciation risk onto taxpayers.

Regardless of whether that's true, Arun clearly expects some kind of crash, and his main recommendation is interesting: the government should be ready to acquire distressed data-center and energy assets.

If it is a bubble, and the bubble pops, we will likely see data centers without tenants and power projects without anchor customers, but with perfectly good physical infrastructure and interconnection rights. Rather than letting that capital stock rot, the federal government and states should be prepared to buy or refinance these assets at fire-sale prices and repurpose them, he argues, for applications such as secure cloud for critical infrastructure, industrial computing, national-security workloads, or other power-intensive industries. Arun explicitly points to TARP as a precedent where public purchases of distressed assets ultimately made money for taxpayers.

We actually like that idea. A structured federal program that acquires distressed data-center assets at market prices would finally publicly commit the government to industrial policy. It would give the state a clear ownership stake in the physical backbone of the AI and cloud economy, rather than leaving national-critical infrastructure entirely in private hands whose investment horizons often shrink at the first sign of a downturn. In geopolitical terms, it would amount to recognizing compute capacity as a strategic resource and treating it accordingly.

And as TARP itself demonstrated, such programs can make money. CPUs, GPUs, substations, cooling plants, and interconnection rights all have residual value; compute demand does not evaporate, it shifts. A temporary public balance-sheet intervention could stabilize the sector, allow bankrupt or over-levered players to reorganize, and ensure that the national computing base remains intact rather than being liquidated piecemeal. By the time the cycle turns, the government could sell or lease these assets back into a healthier market, recouping its outlay while having kept the country’s digital and energy infrastructure whole. In that sense, a data-center TARP would be both pro-market and pro-industry: disciplining excesses, preserving productive capacity, and aligning public power with the long-term national interest.

A good fallback plan to have in the back-pocket. In fact, it would be a good idea to publicly talk about it now, so as to pre-empt future accusations of a "bailout."

Policy News You Need To Know

#GroyperScare — We have tried to stay out of the discourse on the supposed takeover of DC by groypers, meaning adherents of Nick Fuentes's antisemitic ideology, but we will point to this excellent and sober reporting by Emily Jashinsky in UnHerd showing that the threat is, at least for now, when it comes to actual positions of influence, highly overstated. It's a really good report.

#Immigration — ICE launches a program to investigate unvetted sponsors of unaccompanied minor illegal aliens. In one of the most horrifying stories we can remember of the past four years, during the Biden admininstration 450,000 unaccompanied minors who were smuggled across the border were placed with unvetted sponsors, some of whom are sex traffickers and felons.

#Immigration — Speaking of immigration-related horrors, a child was just killed in Idaho by an illegal alien driver with an ICE warrant. The victim has been identified as 8-year-old Mora Gerety. This shows, even after a year of mass deportations, the titanic scale of the task faced by the Trump administration to undo the damage of decades of neglect of the immigration system. May little Mora rest in peace.

#FutureOfChristianity — A new survey of 1,000 regular Protestant churchgoers reveals a striking ideological drift: far fewer adults in the pews now identify as firmly conservative (34%), with growing shares calling themselves moderates or progressives, even as nearly one-third openly prefer socialism to capitalism—including 60% among LGBTQ-identifying respondents and notable strength among younger, higher-income, and black churchgoers. Only 46% say biblical teaching guides their political choices, down from 51% just two years ago. For Republicans, this signals a widening gap between institutional conservatism and the worldview of the congregations historically considered part of the coalition’s backbone. Today's Evangelicals are not like yesterday's Evangelicals, and our coalition is changing whether we like it or not.

#Homelessness — HUD is announcing a complete overhaul of the Continuum of Care program, designed to attack homelessness. It's a deliberate break with the Biden-era approach that poured nearly all funding into Housing First pipelines while insulating local providers from accountability, despite rising homelessness and worsening street disorder. The new $3.9 billion NOFO redirects resources toward transitional housing and treatment-oriented services, reversing a system in which roughly 90% of awards supported a model that, in practice, rewarded permanent dependency while neglecting addiction and mental-health drivers of homelessness . Research from cities such as San Francisco, Los Angeles, and Portland has repeatedly shown that expanding unconditional housing subsidies without treatment requirements coincided with sustained increases in unsheltered populations and fentanyl-driven mortality, whereas jurisdictions that emphasize structured transitional housing, sobriety-linked services, and law-enforcement partnership, such as Houston and San Antonio, have documented measurable reductions in chronic street homelessness. Secretary Turner’s mandate that 70% of projects compete for funding, coupled with restored access for high-performing faith-based providers and alignment with HHS treatment-first priorities, represents an very welcome change. Homelessness policy will now prioritize recovery, public order, and measurable exits to self-sufficiency rather than the automatic renewal of what HUD accurately calls a “self-sustaining slush fund” that failed to demonstrate outcomes.

#GreenNewScam — Over at the Institute for Energy Research, Robert L. Bradley, Jr. marks the ten-year anniversary of the Paris Climate Agreement, offering a tough but fair assessment: it's pretty much dead. The deal, Bradley argues, has collapsed under the weight of its own unreality, vindicating early critics like James Hansen who warned it was "a fraud." With COP30 revealing that nearly all major emitters are off-track and that political enthusiasm now centers on symbolic "clean energy" branding rather than enforceable emissions cuts, the piece contends that global climate governance has reverted to the logic of past failures such as Kyoto. US commitments are rated "critically insufficient," not because Congress has failed to regulate enough, but because the UN framework demands economically self-defeating targets while ignoring the reality that fossil fuels remain the world’s cheapest and most reliable energy sources. The article frames the true policy choice not as a futile race toward Net Zero mandates but as a shift toward adaptation, energy abundance, and regulatory realism; in other words, an approach aligned with market-driven innovation rather than multilateral wish-casting.

#MentalHealth #Kids — In a new study published in JAMA Network Open, RAND reports that roughly one in eight Americans aged 12 to 21 now turn to AI chatbots for mental-health advice, with usage highest among 18- to 21-year-olds and more than 93% of users calling the guidance helpful. After a decade of federal spending growth and bureaucratic expansion in the health care system, nearly half of adolescents with major depressive episodes still receive no care. As a result, young people are reaching for cheap, immediate, and private AI tools, even though no clear standards exist for evaluating chatbot accuracy or safety. For better or worse, AI will guide the future of our mental health.

#Yoorop — It's easy for Americans to mock Europe’s faltering growth model, but it's a problem for America too, argues AEI's Jim Pethokoukis, very rightly. The latest evidence from Brussels, namely its move to dilute and delay core provisions of the AI Act, amounts to an overdue confession that the continent’s precaution-first regulatory culture has throttled its own technological base. As Jim explains, Europe’s decade-long fixation on privacy maximalism and anti-scale competition rules, from the GDPR to the Digital Markets Act, has yielded bureaucracy and dependency rather than innovation. The emerging shift in Brussels toward a more permissive approach is welcome but dangerously late in a world defined by continental-scale industrial competition. America needs a Europe capable of building, scaling, and rearming, which means encouraging EU partners to abandon regulatory habits that mistake economic sclerosis for social justice.

#GovernmentWork — TSA officers who demonstrated exemplary service will be receiving a $10,000 shutdown bonus from DHS.

#Trade #Tariffs — In an op-ed at the journal, AEI senior fellow Desmond Lachman argues that the Trump administration's tariffs are in fact increasing the trade deficit.

#StatePolicy #Tax — Georgia is preparing abolish its state income tax, Katie Cook explains at the Independent Women's Forum. The move is motivated by the state’s long-running competition with its low-tax Southern neighbors: despite twelve consecutive years as Area Development’s top state for business, Georgia still suffers from high living costs, rising pressure on middle-class households, and one of the Southeast’s steepest income-tax burdens. The Senate Special Committee on Eliminating Georgia’s Income Tax now argues that a full repeal, financed by unwinding corporate tax breaks and broadening the consumption-tax base, would not only return $2,000 to $3,000 per year to typical families but also unlock the growth dynamics seen in Tennessee and Florida after their own transitions to no-income-tax models. As appealing as that sounds, we hope they also remember the example of Louisiana, where the abolition of the state's income tax, primarily funded by an increase to consumption taxes, proved unpopular. Still, states are the laboratory of democracy. Let a thousand flowers bloom!

#Slavery — Ok, this is not exactly about within our domestic federal policy remit, but it's just too funny to pass up. A bunch of African nations have called for "slavery reparations" from the West. At the Pacific Research Institute, political scientist Wilfred Reilly points out the all-too-obvious problem with that idea.

Friday Essays

Want to have a good laugh? New York Magazine's The Cut reports on the "Moms of the Upper East Side" Facebook group, which, they inform us, has 35,000 members. After the election of Zohran Mamdani, the group apparently went into full "meltdown." Who thought the leopards would eat MY face??

Speaking of hilarious, and speaking of the Leopards Eating People's Faces Party, the New York Times did a long reported piece on the Sierra Club, and how the conservationist organization embraced every woke fad, and as a result destroyed itself from the inside. Delicious.

This week, we want you to have a good time, we want you to laugh, which is why we're bringing you this third piece of hilarity. It's hard to do this essay justice in summary; through the author's experience of going to Burning Man, he skewers the pretensions of the Silicon Valley world.

More serious: Peter Thiel wrote a now-famous email in which he made the perhaps obvious, but also easily-forgotten, point, that it doesn't matter if capitalism is better in some sense than socialism, if young people don't have enough opportunities and perceive, with some fairness, the system to be rigged against them, they will embrace socialism nonetheless. In the wake of the election of Zohran Mamdani, this diagnosis seems even more prescient than ever, which is why Sean Fischer at The Free Press decided to interview him to expand on those thoughts.

We had missed this great essay from Jon Askonas in Compact when it appeared last month, so we are discussing it now. His argument is easy to summarize: conservatives failed to conserve what they wanted because they underestimated the role of technology, and specifically the role of technology in destroying or at least changing underlying material conditions that make such traditions possible.