Someone else dug the hole

Is AI a bubble? Yeah, probably, but it’s a bubble that already exists. The US economy seems greatly over-indexed on AI, and the bubble popping could be very bad. Unless we can turn the corner and have AI really bolster GDP, things look like they could get dark. Even though this is a bubble we created, I don’t think it’s one that has to pop.

The lemonade stand

Imagine you opened a lemonade stand with $50 from your parents. You bought everything for it and started marketing. Eventually it either goes well and you make a profit, or it doesn’t. Nothing new here.

But the difference in this scenario is that the $50 represents half of your parents’ wealth. So if your lemonade stand fails, it’s not great. You might be forced out of your home, cars repossessed, etc. If it does well, then everything is alright, and now you have a house and a lemonade stand.

AI is our lemonade stand, and the US economy is the house.

How much of the house is riding on the lemonade stand

Look, the over-investment in AI basically leads us to this point where we’ve put all our resources into it, and taking them out is near impossible, or at least difficult. The five hyperscalers, the big cloud companies, are expected to spend more than $690 billion on capex across their 2026 fiscal years. They’re borrowing to do it: new debt went from 9% of their capex in fiscal 2024 to 32% by mid-2026. Most of that money becomes data centers, and in a lot of ways the topic of data centers is very divisive. I’ll come back to that.

Here’s how much of the house is on the line.

In the first half of 2025, investment in information-processing equipment and software was about 4% of US GDP, and it accounted for 92% of GDP growth, according to Harvard economist Jason Furman. Without it, growth would have been 0.1% annualized.

The Magnificent Seven were 34.3% of the S&P 500 in December 2025, up from 12.3% in 2015.

A third of the stock market and nearly all of the growth. That’s the house.

Why we can’t just close the lemonade stand

Closing the lemonade stand, or even slowing AI spending down, would take a massive shift in economic investment elsewhere, which doesn’t seem likely. The economy is greatly buoyed by the AI companies, and the rest is suffering.

The St. Louis Fed found that AI-related investment made up 39% of US GDP growth through the third quarter of 2025, against 28% in 2000, at the height of the dot-com boom. Meanwhile, the annual revision to the jobs data cut 2025 down to 181,000 jobs for the whole year, the weakest year for employment growth since 2003 outside a recession.

The AI stocks prop up spending, too. JPMorgan estimated that US households gained more than $5 trillion in wealth in a year from 30 AI-linked stocks, which raised their annual spending by about $180 billion.

2026 looks a bit better. Business investment outside AI picked up in the second quarter, and manufacturing has expanded for eight straight months. It’s not enough to change the picture.

If the lemonade stand fails

If the lemonade stand fails, it’s catastrophic for everyone, it seems. We don’t just lose the stand, we lose the house. The 34% of the stock market is the big issue. You’ll see a lot of money withdrawn very quickly if the pop happens, and not just in AI. Just like the financial crisis, other things will be greatly affected.

The US is the global reserve currency and a large superpower. If its economy felt like it had been cut in half, that would be bad globally. Gita Gopinath, the IMF’s former chief economist, estimated that a correction the size of the dot-com crash could wipe out more than $20 trillion of US household wealth and more than $15 trillion for foreign investors. Foreign investors held $19.9 trillion in US stocks in mid-2025.

Why the lemonade stand doesn’t have to fail

Crypto is the contrast. Crypto was a lemonade stand with nothing to sell: hype without real use. AI has real use, which is why it can be saved.

For the lemonade stand to turn a profit, it’s a combination of some things: GDP has to move, revenue at AI companies needs to move toward profitability, and research breakthroughs have to happen (though this last one I think is where the AI companies catch up).

Revenue is moving. Anthropic’s annualized revenue run rate passed $65 billion at the end of July 2026, up from $9 billion at the end of 2025. OpenAI’s passed $40 billion, roughly double where it ended 2025. Neither report says anything about profit.

GDP moves through enhanced productivity in jobs, manufacturing, shipping, etc., as well as new products and industries. On top of that, huge innovations in medicine, science, etc., driven by AI-enabled or AI-enhanced research.

The 2024 Nobel Prize in Chemistry went partly to Demis Hassabis and John Jumper for protein structure prediction: an AI model that solved a 50-year-old problem. And rentosertib, a lung-fibrosis drug whose target and molecule were both found with generative AI, entered a Phase III trial in July 2026 after a positive Phase 2a in Nature Medicine.

But really, GDP needs to uptick by a measurable amount due to AI, and then we can dig ourselves out of the bubble.

The lemonade stand works, on small teams

If AI is paying off, it should show up in productivity. The Federal Reserve doesn’t see that in the aggregate numbers yet. Its July 2026 note says experiments at the level of individual tasks keep finding productivity gains from AI tools, but productivity across the economy hasn’t accelerated.

That doesn’t match what I see. Engineering teams full of senior+ engineers can ship at an incredible rate.

We had an enterprise SaaS product that was doing “fine”, but it wasn’t really delivering like we wanted. We had the crazy idea to add some new features. It started as a side project and turned into a new product that eclipsed and replaced the old one, mostly because we built with our AI tooling and dev process at the forefront. We started with skills, rules and good AI dev design, and we all used the same things and understood what they did.

It took 1 month to replace the old product, and we immediately began receiving sales at 5x the rate of the prior product.

Larger teams likely don’t see these gains because they’re more encumbered by process, bureaucracy and lack of access to tooling. When a small team of senior folks who know how to build and deploy are together, that’s where the magic happens. If large enterprises want to see these gains, they need to give more autonomy to engineers, hire product-minded engineers, and get other organizations (such as product) out of the way.

The skills I preference now are system design and product thinking: really thinking about users’ problems, what will solve them, and being innovative and not afraid to take risks. Junior engineers are in scary territory. I’d probably tell them to put their skills to use in personal projects, or to consider a new industry, but I’ll save expounding for another post.

What we sell, and where we build

For the lemonade stand to work, it matters what we sell and where we build it.

My long-term fear is that we’re focusing AI on the consumer, on what it can do that’s cool and day to day, and on chat, when the real advancements and large winners will be in research and new product creation. Wearable AI chat devices were, and are, dumb. Chat assistants like Meta’s Muse are cool, but I’m not sure they’re having a measurable GDP effect.

Then there’s where we build the lemonade stand. I think the current shotgun approach of trying to shove data centers into different municipalities is bad. Have a national plan for this. Stop just letting corporations try to do it, angering citizens and residents and not having a sustainable plan. Put them where power and water are abundant, plan nuclear around them, etc.

You should care about the data centers, even if you hate the idea of one down the road. Without the compute, we might not be able to dig ourselves out of a very real hole. It doesn’t much matter if you like the hole or not, someone else dug it.