Pretty much every finance thread online now takes it for granted that the AI bubble is about to pop. A widely shared essay on Hacker News pushes back on that. Its author argues that the huge sums AI companies are spending look more like a one-time construction bill than a broken business model. The post drew a score of 185 and plenty of debate, and it’s worth a close read. Its argument has real weight in some places and real holes in others.
🏗️ The Core Claim: Capex Isn’t Forever
The essay starts with the scary number. AI companies have committed to at least $3 trillion in spending and have already spent about $1.9 trillion, so the industry’s total bill comes to roughly $5 trillion. The author agrees that nobody can earn a profit if spending stays at that level.
He doesn’t think it has to stay there. Most of the money is going into data centers, which he treats as one-off purchases rather than ongoing operating costs. Once the buildout ends, he expects spending to drop sharply.
His main evidence is Anthropic’s S-1 filing. By his reading, it shows:
- $2.08 in revenue for every $1 of compute, so AI is profitable once you set aside infrastructure costs
- Revenue that now exceeds compute and infrastructure investment combined
- Revenue growth from $41 million in 2023 to $386 million in 2024, $4.585 billion in 2025, and $16.3 billion in the first half of 2026 alone
Those growth numbers are hard to wave away. If other labs are growing anywhere near that fast, the claim that “nobody will pay for this” loses a lot of its force.
🧮 The Back-of-Envelope Math
The author also does some rough debt math. Paying off $5 trillion over 20 years works out to about $250 billion a year. For comparison, Alphabet brought in about $162 billion in revenue in 2019, or around $213 billion in today’s dollars. His conclusion is that the bill is enormous but not unthinkable.
He also points out that some of the debt sits in convertible notes, which can be repaid in shares instead of cash. He cites $72 billion in share-based losses owed to lenders in Anthropic’s first half of 2026. That figure looks so large mainly because the company’s valuation has climbed to an estimated $1 trillion.
🔍 Where the Argument Gets Shaky
This is where I’d push back on the pushback. Some of the essay’s weak spots matter a lot:
- GPUs aren’t buildings. Calling data center spending “one-off” ignores that AI chips wear out or become outdated within a few years. Some of that capex will keep coming back, so it behaves more like a recurring cost.
- Revenue isn’t profit. Comparing a $250 billion yearly debt payment to Alphabet’s revenue skips over the costs that come out of revenue first. To service that debt, the industry would need far more than $250 billion in sales.
- Interest is missing. Spreading $5 trillion over 20 years with no financing costs makes the math look friendlier than it really is.
- Paying in shares only works while the valuation holds. Settling debt with stock is easy at a $1 trillion valuation. In a downturn, that same mechanism becomes painful very quickly.
The author admits the most important weakness himself. The financing vehicles are complex enough that nobody can say exactly how much the industry owes. In bubble debates, that kind of opacity has usually been a warning sign.
On track record, this is an individual essayist, not an analyst with a history of market calls. Read it as a well-built argument, not a forecast.
💡 The Sharpest Point: Skeptics Can’t Have It Both Ways
The essay’s best line of argument is about a contradiction it sees in many AI critics. They fear AI will take over jobs, yet they also insist it only produces “slop” that nobody will pay for. As the author puts it, “money is going to change hands when that happens.”
He bets that demand is still largely untapped. Coding, math, legal and medical paperwork, and self-driving cars are all far from fully adopted. He believes AI consumption could reach “10x what it is now” even if the models stopped improving today.
📌 What Practitioners Should Take From This
- Watch unit economics, not headlines. Compute margins and revenue per dollar of compute tell you more about the industry’s health than total capex does.
- Don’t build on free tiers. The author notes that labs can cut costs “instantly” by restricting free access, so plan for prices going up.
- Hedge on vendors. Even if the industry overall is healthy, individual companies with heavy debt can still fail.
The real question may be timing rather than whether a bubble exists. Can revenue grow fast enough to outrun depreciation and debt? Anthropic’s next few quarterly reports should give us a much clearer answer. The full essay is on Hacker News.