In October 2025, when credit markets looked stable and strong, I warned that a private credit winter was coming. I wasn’t clairvoyant: I just had better data, the answers were buried in contractual terms signaling distress was ahead. Now we’re seeing those signals again: while everyone’s worried about software companies, they should actually be worried about AI startups.
Most people liken the AI boom to the dot com bubble. But the right comparison is the lesser-known portal wars.
Let’s go back to 1998. The internet had just gone mainstream, and a new kind of company was taking over the web: AltaVista, Excite, Lycos and Yahoo were each racing to become your home base online–the “interface” to the internet. They competed on adding vertical workflows: features like news, email, weather and shopping. Venture capital poured in and they grew fast. For a moment, it looked like any one of them, or all of them, could win, each differentiating themselves in domains in which they were marginally better from the other.
Then a pair of Stanford graduate students created a new model, a search algorithm called PageRank, which used the web’s own link structure as a proprietary data signal. No email, no weather, no shopping, no news. PageRank made a bet: the “interface” to the internet would be won by information retrieval, not by vertical workflows. Excite’s CEO famously turned down PageRank at a price of $750,000 and within a few years, PageRank became what we know as Google and every other portal had been rendered irrelevant.
We are watching the same movie with AI startups today.
Thousands of companies are building products on top of the same AI foundation models — OpenAI’s GPT, Google’s Gemini, Anthropic’s Claude – with no added proprietary content or data, only workflows, each aiming to be the “portal” for their vertical. They have different names, different user experiences, different pitches to investors. What most of them share is that the intelligence powering their product is available to every competitor, every established company, and increasingly to ordinary consumers at low or no cost because they’ve added nothing proprietary to enhance their offerings.
But here’s what most people are missing: the portal wars showed that information retrieval power, not workflows, wins interfaces to new technologies. That reality only compounds with enterprise technology: fiduciary‑grade AI requires proprietary data and information to win professional interfaces in the long term.
The recent US Executive Order on AI sends a clear signal that accountability and trust are not optional. As AI moves deeper into regulated work, the defining question is no longer whether a system can generate an answer, but whether professionals can verify it, and confidently stand behind it.
Unless startups can marry foundational models with proprietary data and content, when the companies behind foundational models inevitably add new capabilities, entire categories of startups, and the value investors poured into them will vanish overnight. When that distinction becomes clear to the market, the correction will not stay in Silicon Valley. Corrections are never confined.
The evidence is already in the deal terms.
Across filings, two disclosures are emerging. In investor materials, companies tout foundation model access as a competitive advantage. In risk factors, companies disclaim responsibility if that access changes, gets more expensive, or disappears. Read together: AI makes us a better investment than our competitor, but we have no control over that thing that makes us a good investment. That’s the gap. Gaps like that are where bubbles form.
In M&A, sophisticated buyers are already protecting themselves from this gap. A new class of seller representation is emerging, with sellers assuring buyers that their AI use creates no dependency that would prevent the business from continuing. In some deals, sellers are being asked to represent that their core IP contains no foundation model exposure at all. When these types of terms show up in M&A, the smart money is surreptitiously telling us exactly where the value in these companies actually lives.
This gap between what’s being pitched to investors and what companies are willing to stand behind is the bubble. When the bubble bursts, value will concentrate with two types of AI providers: (1) foundational model providers and (2) the institutions that own the data those models need to be trustworthy. The more investors focus on the Excites and Lycos of this story, no matter how fast they are growing today, the larger the bubble becomes.
This is not a prediction that AI fails. The internet succeeded. But when the bubble burst, the S&P 500 lost nearly half its value and the Nasdaq took 15 years to recover. The people who understood the underlying structure of that market had options. Most Americans didn’t because most Americans didn’t know they were exposed.
The question at this point isn’t whether there is an AI bubble. It’s whether the market expectations of where value accrues match the medium-term reality. They don’t. That’s the real AI bubble that investors should be paying attention to right now.
Written by Dan Wertman, VP of Noetica at Thomson Reuters (Noetica was acquired by Thomson Reuters in 2026)
