A practical guide to the looming(?) AI bubble

And why a bust wouldn't be good news for skeptics of AI

By now you’ve heard about the AI bubble. There was a bunch of noise about one back in 2025. My first post on Substack was a warning about a potential AI plateau. I imagined what might happen if AI did not progress at the same pace as the prior two years and failed to live up to its hype. In fact, the opposite took place. Claude 3.7 Sonnet and Claude Code unleashed a breakneck push to expand AI capacity, and for the first few months of 2026, to me it seemedlike AI was now underhyped by those who had not yet seen what coding agents were capable of first-hand.

I still feel like most people are underestimating AI. But that certainly can’t be said for holders of capital: the hyperscalers have budgeted roughly $700 billion of capital expenditure this year against less than $100 billion of total AI revenue, and rumors of an impending AI bubble are everywhere again. Detractors of AI might take solace in those rumors.

I want to unpack the math behind the “AI bubble” so that you can form an opinion of your own. And I also want to talk about why AI detractors should pause before running a victory lap.

The math behind the AI bubble

It’s pretty straightforward. When these companies cite a run rate, they anualize their most recent month or quarter, so $2.5 billion of revenue in a single month reads as a $30 billion run rate. It is a projection, not audited annual revenue. With that caveat: Anthropic’s run rate was about $9 billion at the end of 2025. By March 2026 it had roughly doubled to $20 billion, and it reached $30 billion by April. By May, Anthropic itself was claiming a run rate near $47 billion.

Numbers like this are pretty much unheard of. It’s helpful to compare to some companies you are familiar with. OpenAI is valued at roughly $850 billion — compared that to Ford ($45B), Delta ($40B), Target ($55B), Marriott ($70B), Kroger ($50B). Capital markets have thus become obsessed—addicted, even—to the AI trade. Hyperscaler capex alone is running at roughly 2% of US GDP. AI investment accounted for the majority of the country’s GDP growth in early 2026; strip it out, one Pantheon analyst noted, and US business investment would actually be shrinking.

These valuations are a bet on future revenue, not today’s. Investors are pricing in years of compounding growth from here, on the assumption that today’s losses convert to profits once the models are embedded everywhere. As the valuations got bigger and bigger, though, expectations might be getting ahead of reality. Investors need to achieve a return within roughly 5 years. The question is: will organizations and companies be able to figure out the change management fast enough to drive the level of demand required?

There are two sources of AI demand:

1. productivity increases that compel incremental new spending on AI to grow market share, feature sets, new revenue, etc.

2. Efficiency that enables companies to reduce operating costs. Reducing $1.5 of labor with $1 of AI spend would yield gargantuan savings.

It’s hard to predict how large the first bucket will be. But to justify the spending on labor savings alone, AI spend would need to replace between 15-33% of all US labor income, and do it inside the roughly five-year horizon that investors are pricing in. Goldman Sachs’s forecast is around 15 million workers displaced over a decade. The valuations assume something several times bigger, and faster. I walk through that math below.

Every transformational technology comes with a bubble

It isn’t terribly prescient to predict a bubble. Virtually every technology in history that transformed society came with a financial bubble. Canals (1790s), railroads (1840s), oil (1860s), automobiles (1900s-1920s), radio (1920s), electricity (1920s), aviation (1920s), personal computers (1980s), the internet (1990s), solar (2000s). . Britain passed 263 railway acts in 1846 alone before shares collapsed in the Panic of 1847. RCA rose roughly 200-fold over the 1920s, then fell 98% — while radio remade mass culture anyway. The NASDAQ peaked on March 10, 2000 and fell 78%, and the internet went right on eating the world. (The full catalog, with receipts, is in the P.S.) The bubble is coming, sooner or later. It doesn’t make me smart to say so; it just makes me a history buff.

That said, there’s reason to believe the bubble is coming sooner than later. OpenAI’s audited financials leaked in June: a $20.9 billion operating loss on $13.07 billion of revenue, which works out to roughly $1.60 spent for every dollar earned. By Epoch AI’s tracking, the Big Four hyperscalers’ combined free cash flow crosses roughly zero this quarter (Microsoft holds out until 2028), which means that from here, the buildout runs on borrowed money. And the lenders have started hedging. Banks stalled for weeks on a $6 billion loan against SoftBank’s OpenAI stake, with talks resuming only after SoftBank offered guarantees backed by everything else it owns. Blue Owl, one of the biggest private-credit lenders behind the datacenter buildout, gated its tech-lending fund after investors asked for roughly 40% of their money back two quarters in a row. Rental prices for Nvidia’s newest chips fell 31% in three weeks in June, which is what overcapacity looks like before anyone admits to it. Even the Bank for International Settlements — the central bank for central banks, not an institution given to drama — wrote last week that this boom resembles manias that “ended with an eventual reversal in investment, inducing economy-wide recessions.”

All this in mind, the AI bulls have their own case to make. There has been no credit event and June was a record month for investment-grade bond issuance. Alphabet and Amazon each sit on more than $100 billion in cash, and the hyperscalers together threw off close to $450 billion in operating cash flow last year. And unlike the dark fiber of the dot-com bust, which sat dark for years, today’s compute is getting bought as fast as it is built. Microsoft and Google keep saying they cannot build fast enough, and OpenAI and Anthropic are rationing access. For now, every byte of inference finds a buyer. Prediction markets still have the odds of a burst by New Year’s at between 15 to 24%.

Bubbles are easy to predict, but hard to time. Our crystal ball gets less cloudy over the summer and fall: Nvidia reports on August 26, Oracle on September 14. SoftBank’s next $10 billion check to OpenAI is due October 1. And Anthropic is expected to go public. Each one will be a rigorous test of whether AI demand is as durable as the spending assumes—and whether capital is willing to continue leaning into the excitement.

Don’t take solace in the AI bubble

For those who have always been skeptical about AI and concerned about its impact on society, it might be tempting to wait gleefully for the opportunity to shout “I told you so” from the rooftops.

In fact I think the AI bubble bursting is a worst-case scenario for those worried about the negative impacts from AI.

First, it is a false hope for a crash that puts the genie back in the bottle. AI isn’t going anywhere. Google, Microsoft, and Meta have been building formidable AI labs and these companies would survive a bubble and no doubt would be prepared to step in to carry the torch if the opportunity to acquire a distressed frontier model company presented itself.

Even if Google, Microsoft, Meta and other big tech behemoths are somehow incapacitated, Open source models are already widely available. A year ago, the best downloadable models were a clear step behind the labs. That gap has mostly closed. Open weight models from DeepSeek, Qwen, and Kimi now trade within a few points of the closed frontier on coding and reasoning, and the gap keeps narrowing. They are free to download and run on your own hardware. So far, companies have demonstrated that they are willing to pay for the flagship model, but there isn’t any technical reason why they couldn’t switch to an open source option. This potential pricing competition is actually one of the bear scenarios for frontier AI investments.

And if AI isn’t going away, we need to reckon with what will happen to jobs.

Start with what today’s valuations are actually pricing. Bain figures the buildout needs about $2 trillion a year of new revenue by 2030 to pencil, and nobody pays that for software unless it returns a multiple of what it costs — which points at several trillion dollars of value, most of it coming out of work people are currently paid to do, in an economy with a $15 trillion annual compensation bill. Discount that math however you like; whatever remains is still a bet on labor-market disruption several times larger than Goldman’s forecast of 15 million displaced US workers over a decade. So either the labor-market hurricane I’ve been warning about makes landfall on the fastest timeline in economic history, or the financing breaks and we get a crash, most likely with a recession attached. There is no third branch where everyone calms down and nothing changes.

And a recession is where the labor damage concentrates. Nir Jaimovich and Henry Siu found that 88% of the losses in routine jobs since the mid-1980s happened within a twelve-month window of a recession — those jobs vanish in downturns and never come back. Brad Hershbein and Lisa Kahn read millions of job postings and found that employers in the metros hit hardest by the Great Recession raised skill requirements and invested in technology at the same time, and the higher bar stuck after the recovery. And the recovery itself came back as a barbell: mid-wage occupations absorbed 60% of the recession’s losses and got 22% of the recovery’s growth. In 2009, at least, that kind of restructuring required capital — machines, multi-year integration projects. Generative AI is a subscription.

If anything, a crash makes the tools cheaper. That was the lesson of the last big one: 85% of the fiber laid in the late-’90s telecom bubble was still dark years later, and that stranded glass became the dirt-cheap substrate YouTube and Netflix were built on.

AI is here to stay. The question is simply how fast, that is to say, how acute the disruption will be.

What the bubble means for education

Artificial intelligence is rapidly shifting the goal posts for what it means to be prepared for the future. Yet the vast majority of classrooms are not evolving in step to ensure students are ready. In 2025, less than half of teachers report receiving any formal training on AI, and only 31% of public schools have a written policy on AI use in schools. Despite the limited guidance, 61% of U.S. teachers report using AI tools in their work, with one-third saying they use AI weekly — primarily for lesson planning, instructional materials, and adapting content for students. The U.S. education system is flying blind.

Meanwhile, the importance of AI readiness is becoming a primary focus for the economy. Employers are moving the same way. Two-thirds of companies are hiring for AI-specific roles, and more than half of hiring managers say they won’t hire someone who can’t use AI.. As AI becomes embedded in both the economy and everyday learning, supporting teachers to understand and guide its use is essential to ensuring students flourish in the age of AI by developing durable human skills (critical thinking, communication, creativity, and collaboration) — and the agency to apply them.

Now run those two paragraphs through a bust. Every district leader who spent the last three years in wait-and-see mode will feel vindicated, because the headlines will say the skeptics were right all along. State education budgets will crater the way they did after 2008 — ask anyone who got a pink slip in the spring of 2009 — and AI readiness will become the easiest line item to cut. The students who graduate into that economy inherit the worst of both: fewer entry-level on-ramps, and employers who treat AI fluency as a floor. Danny Yagan found the employment scars of the 2008 recession still visible in 2015. Wait-and-see fails in both branches of the fork. If the hype is real, students face the fastest workforce transformation in history. If the bubble bursts, they face it anyway, plus a recession.

The coalition is already here

AI may be one of the most polarizing forces in American life, but AI readiness has emerged as that rare education priority everyone can get behind. Teachers aren’t asking whether AI belongs in their work; they’re asking for guidance, for training that helps them use AI to strengthen their practice rather than leaving them to navigate it alone. Teachers’ unions have reached the same conclusion: when the American Federation of Teachers launched its $23 million National Academy for AI Instruction in 2025, it chose New York City as its home. It’s a clear signal that organized labor understands its members need training to shape this transition rather than be left behind by it.

Parents complete the coalition. Many are deeply concerned about AI’s risks to their children, from academic shortcuts to AI companions, but that concern doesn’t translate into a desire to keep AI out of schools. It translates into an expectation that the adults in the building be prepared: teachers equipped to help students navigate those risks and ensure they’re ready to thrive in the fourth industrial revolution. It’s rare to see teachers, unions, and parents pulling in the same direction, and it won’t last forever -- least of all through a market crash and the budget cuts that would follow. The opportunity for philanthropy is to act while consensus is ahead of capacity.

It is indeed the time for advocacy and activism to push for AI to be human-centered. But that must not come at the expense of pushing forward, harder than ever, to transform every classroom in the country: active learning, real projects, student agency. Learning experiences that build knowledge alongside durable skills will be critical no matter what timeline we find ourselves in.

Bubble or not, the one investment where the ROI is certain is an investment in education that prepares students for the future.

P.S. The full catalog, for anyone keeping score: British canal mania (authorizations went from 1 in 1790 to 20 by 1793, then collapse); Railway Mania (263 acts in 1846; the American echo was the overbuilding behind the Panic of 1873); the Pennsylvania oil rush (oil crashed from $10 a barrel to 10 cents; Pithole grew from four farmhouses to a city of fifty-plus hotels in five months in 1865, then emptied); the bicycle craze of 1896; the automobile shakeout (253 US carmakers in 1908, 44 by 1929); radio (RCA up 200-fold, down 98%); Samuel Insull’s electric-utility pyramid ($500 million controlled on $27 million of equity; its collapse wiped out the savings of 600,000 shareholders); the Lindbergh aviation boom; the 1983 home-computer shakeout (TI exited after $400 million in losses, and the video game market shrank 97%); dot-com; the telecom crash ($500 billion-plus invested, 85% of the fiber dark for years); and cleantech ($25 billion of venture money, more than half of it lost). Every one of those technologies outlived its bubble. The investors, mostly, did not.