Essay argues the AI bubble may not exist, citing Anthropic's S-1

In a blog post dated 8 November 2026, the author pushes back on what they describe as an assumption in online discussion: that the AI bubble exists and will pop any day now. The author's own position is cautious. Nobody can see the future, but there is a "pretty realistic chance" that the growth of AI companies in recent years is justified and that they have more room to grow. The author also says they are not an expert on financial bubbles.

The first argument is about spending. The author concedes that debt and spending commitments run through complex vehicles that make true expenses hard to pin down, and calls that one of the most salient points in favour of a bubble. Still, they put commitments at no less than 3 trillion dollars, cite an unsourced estimate that 1.9 trillion has already been spent, and arrive at around 5 trillion dollars overall. Their answer is that this money mostly goes into building data centres, which are one-off purchases rather than operating costs, so spending should plummet once the data centres exist. Running AI is, in their reading, cheap: Anthropic's S-1 filing says the company makes $2.08 for every $1 of compute, and the author reads the filing as showing Anthropic now earns more than its compute and infrastructure investments combined. In a pinch, companies could also end free use and pause expensive research, which can be done instantly, unlike leases and spending commitments.

Debt is the one serious long-term cost. The author's back-of-the-envelope math: repaying 5 trillion dollars over 20 years is 250 billion dollars a year. For scale, Alphabet reported almost 162 billion dollars of revenue in 2019, or 213 billion adjusted for inflation. The needed revenue could be well over 250 billion a year, the author says, but paying it off is not unthinkable. Some debt is in convertible notes that can be repaid in shares; Anthropic recorded 72 billion dollars of loss in the first six months of 2026 in the form of shares due to lenders, which the author attributes to a valuation that has risen to an estimated 1 trillion dollars, leaving ample equity to make lenders whole.

The second argument is that profits can grow a lot. The author cites Anthropic's S-1 revenue as 41 million in 2023, 386 million in 2024, 4.585 billion in 2025 and 16.3 billion in the first six months of 2026 (the unit is not stated in the sentence). They assume other major labs such as OpenAI see similar growth, and say the rate must level off eventually but not necessarily soon. In their view there is unused potential in frontier models: most or maybe all programmers and mathematicians are worse than an AI swarm with a good harness, and more expensive. AI cannot replace lawyers and doctors but can automate their paperwork, and self-driving cars are becoming real. Adoption takes time, but usage should skyrocket even if AI gets no better. The author also calls it contradictory for sceptics to fear AI's impact while insisting it produces only slop, and says it is hard to see AI consumption not being 10x what it is now, with no timescale given.

Third, the lack of a moat. The author recalls the leaked 2023 Google memo "We Have No Moat, And Neither Does OpenAI", which argued that open-source work keeps pace with the labs and that no company holds a durable advantage. The author accepts this but compares it to Popeyes in the chicken sandwich war: no restaurant had an enduring edge, yet it was profitable for all. Not having a moat does not mean a company cannot make money. The average person will not run a local model, brand loyalty keeps people with the product they started on (the author says they use ChatGPT mostly because it was their first LLM), and large customers are locked in by integration and contracts to buy set amounts of tokens.

Fourth, capital. The author says it is premature to worry about AI companies failing to raise money when most have not had an IPO. SpaceX, which the author calls partially an AI company, raised more than 80 billion dollars when it went public. Many AI companies have not issued public debt, and those that have, such as Alphabet and Meta, have excellent bond ratings. Profits can also fund growth and debt repayment, and from Anthropic's S-1 the author thinks the industry may be past the stage of risky bets.

Finally, valuations. The author notes that bubbles are usually preceded by obviously inflated valuations, and the dot-com bubble was largely driven by firms that could not turn a profit or had P/E ratios in the hundreds. Alphabet, Meta and NVIDIA have all seen market capitalisation explode, yet the author says all have stable P/E ratios in their 20s, which looks like a healthy trillion-dollar company rather than irrational pricing. Last, the author considers the scenario in which AI is a fad and GPU and data-centre demand is propped up by invested capital. They test it on Alphabet, whose prospectus splits revenue into Google Cloud (rented compute) and Google Services. Their conclusion is that Alphabet's higher earnings are not mostly from data centres: Google Cloud grows faster in percentage terms, but Google Services grows far more in absolute terms, up 37.8 billion dollars or 12%. The text available to us breaks off mid-sentence while giving the Google Cloud figure, so the end of the argument is not covered here.

Key facts

The author puts AI commitments at no less than 3 trillion dollars, cites an unsourced estimate of 1.9 trillion already spent, and totals around 5 trillion; they argue this is mostly one-off data-centre building that will fall away.
Reading Anthropic's S-1, the author says the company makes $2.08 for every $1 of compute and that revenue went from 41 million in 2023 to 4.585 billion in 2025 and 16.3 billion in the first six months of 2026 (unit not stated).
Rough math: repaying 5 trillion dollars over 20 years is 250 billion a year, which the author compares with Alphabet's almost 162 billion dollars of 2019 revenue (213 billion inflation-adjusted).
The author says Alphabet, Meta and NVIDIA have stable P/E ratios in their 20s, unlike dot-com firms with P/E ratios in the hundreds, and that SpaceX raised more than 80 billion dollars in its IPO.
The author admits opaque debt vehicles are a strong point for the bubble case and says they are not an expert on financial bubbles.

Why it matters

The post is a counterweight to the widespread online view that an AI bubble is about to burst. It does not bring new data. It reads figures from Anthropic's S-1 and Alphabet's prospectus and argues from them that spending is front-loaded, revenue is growing fast and valuations are not inflated the way the dot-com era's were. For readers following the AI-bubble debate, it shows how the opposing case is built: one-off capex, strong unit economics, available capital and modest P/E ratios.

Who it affects

Investors and analysts weighing exposure to AI companies and to the GPU and data-centre chain are the obvious audience. It also touches the AI labs and large platforms named in the post (Anthropic, OpenAI, Alphabet, Meta, NVIDIA, SpaceX) and anyone tracking whether heavy AI infrastructure borrowing is sustainable.

How to use it

Treat it as a checklist of questions rather than a verdict. Which of the 5 trillion dollars is one-off build-out and which recurs? What do the S-1's compute margins and revenue figures actually say? How are debt vehicles structured? Where the author cites filings, readers can check Anthropic's S-1 and Alphabet's prospectus directly.

How solid is it

This is one person's opinion piece, not an analyst report, and the author says they are not an expert on financial bubbles. All S-1 figures are the author's reading of the filing; the post does not link or excerpt it. The 1.9 trillion spent and the 1 trillion Anthropic valuation are given as estimates with no source. The 20-year repayment term is a hypothetical, and the post does not say which companies hold the debt or at what interest rate. The 10x consumption view has no timescale. The text available to us is cut off mid-argument in the Alphabet comparison.

Risks and caveats

The author's own concessions point to the weak spots: spending commitments run through complex vehicles that hide true expenses, and the 250 billion a year figure is very rough and could be well above that. The case also depends on assumptions that spending will plummet after build-out, that revenue growth continues, that other labs mirror Anthropic's growth, and that the lack of a moat does not erode margins. Favourable P/E ratios are cited for Alphabet, Meta and NVIDIA, but Anthropic's own roughly 1 trillion dollar valuation is only an estimate.

“Simply saying that a company doesn't have a moat isn't enough to say that the company can't make money.”

— From the blog post "I'm Not Convinced The AI Bubble Exists"