By Juan Garces De Marcilla, Founding Member, Montesa Capital
The AI bubble debate gets confused because two different questions are being asked at once: is AI overvalued eventually, and is it about to crash soon. The first question is genuinely unsettled — reasonable, credentialed people disagree, and the structural risk we wrote about previously (AI’s success slowly eroding the consumer income it depends on) is real but plays out over a decade, not over the next earnings cycle. It tells you nothing about whether 2026 or 2027 brings a crash.
The second question — is a burst likely in the next 18 months — is the one that actually matters for capital allocated today. Here is the strongest verified case that it isn’t.
1.This isn’t 1999.The revenue is real.
The dot-com comparison breaks down on contact with the numbers. Microsoft’s AI business has surpassed a $37 billion annual revenue run rate, up 123% year over year, with commercial backlog standing near $627 billion. Nvidia’s Q4 data center revenue hit $62.3 billion, up 75% year over year. These are not projections or pre-revenue startups burning venture capital on Super Bowl ads — they are large, profitable, audited businesses with accelerating cash earnings. Dan Ives, the Wedbush analyst who has covered tech since the dot-com era, makes the contrast directly: in 1999 the average tech stock traded at 30x revenue on little more than slide decks; today’s AI leaders are generating hundreds of billions in cash against real infrastructure and real customers.
2. Adoption is still early — the demand runway hasn’t been tested yet.
Bubbles burst when the market for a product turns out to be smaller than investors assumed. That’s hard to argue convincingly here: by Wedbush’s count, only about 3% of US companies and less than 1% globally have meaningfully deployed AI. Whatever you think of current capex levels, the claim that AI has already saturated its addressable market and is now running out of buyers doesn’t survive contact with the adoption data. There is a lot of room left to grow into the spending before demand becomes the limiting factor.
3. Balance sheets went into this cycle unusually strong..
This is the most underappreciated point in the debate. Unlike 2000 — and unlike a classic credit bubble — most hyperscalers entered the AI buildout with historically low leverage or outright net cash positions. By late 2025, the aggregate liabilities-to-assets ratio across major hyperscalers had fallen to around 48%, near 2015 lows, compared with roughly 80% for the S&P 500 broadly. That matters because it means even as capex begins to outpace operating cash flow — which independent analysis projects could happen in aggregate by roughly Q3 2026 — these companies are financing the gap from a position of strength, with investment-grade balance sheets, not from a position of fragility.
4. Independent, mechanical frameworks back this up.
JPMorgan ran the AI rally through a formal five-factor bubble-diagnostic test in December 2025 and concluded that current investment is linked to actual enterprise revenue rather than speculation alone — a structured, falsifiable test, not a sentiment-based opinion. Separately, Federal Reserve Chair Jerome Powell has drawn an explicit distinction from the dot-com era on the same grounds: that AI companies are generating real revenue and that data center spending is contributing to measurable GDP growth, not just to stock multiples.
The one thing actually worth watching
Intellectual honesty requires naming the real near-term risk rather than ignoring it: the financing structure of this buildout is shifting from cash flow to debt faster than the bulls like to admit, and that shift is not evenly distributed. Oracle is the clearest stress case — its five-year credit default swap spread roughly quintupled between early 2025 and March 2026, reflecting real concern about its debt-funded, highly customer-concentrated buildout tied largely to OpenAI. That is a legitimate watch item. But it is a concentration risk in specific balance sheets, not evidence of a sector-wide bubble about to pop. The distinction matters: a single overleveraged counterparty stumbling is a credit event, not a verdict on the asset class.
The bottom line
The strongest version of the bear case isn’t “AI demand doesn’t exist” or “the revenue is fake” — both claims are contradicted by the earnings data. The strongest version is a slower, structural one: whether the income AI displaces over the next decade quietly narrows the customer base that all of this revenue ultimately depends on. That’s a real question, and we’ve made the case for it elsewhere. It is not, however, an 18-month catalyst. For the time horizon that determines whether capital deployed today survives the next two years, the data argues for real revenue growing into real spending — with leverage concentrated in a small number of names worth watching individually, not feared collectively.
Montesa Capital—research notes are for informational purposes and do not constitute investment advice.