Is There an AI Bubble? The Case That the Numbers Add Up

Few claims are repeated as confidently in tech circles as the idea that the AI bubble is about to pop. A recent essay on a personal blog takes the opposite side, arguing that the spending is huge but not necessarily unsustainable. It is a one-sided piece, built largely on a single company's financial filing, but it raises questions that anyone building on top of AI infrastructure should think through.

The spending is real, and so is the opacity

The author starts by conceding the obvious. AI companies have committed to enormous sums, with figures cited of roughly 3 trillion dollars in commitments and around 1.9 trillion already spent, for a combined total near 5 trillion. These numbers come from estimates, not audited totals, and the author admits that complex financing vehicles make the real picture hard to read. That opacity is arguably the strongest argument for the bubble thesis, and it deserves more weight than the essay gives it.

The counterargument is about the type of spending. Most of it goes into building data centres, which is capital expenditure rather than a recurring cost. Once the build-out slows, the logic goes, spending should fall sharply. The question is whether it will slow, since each new generation of models tends to create demand for more capacity.

Margins, debt and the back-of-envelope math

The essay leans on Anthropic's S-1 filing, which it says shows the company earning about two dollars for every dollar spent on compute. If that holds, running models is profitable and the problem is servicing the cost of building the infrastructure. The author also notes that the filing suggests revenue now exceeds compute and infrastructure outlays combined, though that is a reading of one company's numbers, not proof for the industry.

The debt math is deliberately rough. Paying off 5 trillion dollars over 20 years works out to about 250 billion a year, which the author compares to Alphabet's revenue a few years ago. That comparison is a useful sanity check but not a forecast: it ignores interest, uneven maturity schedules, and the fact that not every borrower is Anthropic. Some debt is also held as convertible notes that can be settled in shares, and the filing reportedly records a large paper loss tied to rising valuation, which cuts both ways if valuations fall.

The revenue trajectory cited is steep: from 41 million dollars in 2023 to 386 million in 2024, 4.585 billion in 2025 and 16.3 billion in the first half of 2026. Growth like that cannot continue indefinitely, and the author says as much. The open question is when it levels off and where.

Demand, moats and the open-source angle

The essay's bet is that unmet demand is large: coding, maths, paperwork-heavy professions and self-driving all have room to absorb more AI. It also points out an inconsistency in some skepticism, where AI is both feared as transformative and dismissed as worthless slop. Both cannot be fully true.

The part most relevant to developers is the moat discussion. The author recalls the leaked 2023 Google memo arguing that neither Google nor OpenAI had a durable advantage, because open-source work was keeping pace. If frontier models are easy to match, no lab wins permanently, and margins get competed down. The essay compares this to a fast-food sandwich war, which is flippant but captures the point: huge demand does not guarantee fat profits when competitors can copy the product.

What to watch

The essay does not prove there is no bubble, and the author is careful to say nobody can see the future. But it usefully separates three risks that often get lumped together:

• Financing risk: opaque debt structures and commitments that are hard to audit.
• Margin risk: commoditisation driven by open-source models and fast followers.
• Demand risk: whether adoption grows fast enough to justify the build-out.

For developers, the practical takeaway is that falling prices from competition and open-weight models may be good news for users even if it is bad news for investors. Watch for audited filings from other labs, how quickly data centre spending tapers, and whether open-source models keep closing the gap. Those three signals will say more about a bubble than any confident online consensus.