1999.AI

AI and dot-com déjà vu

Jamie Dimon once defined a financial crisis as “something that happens every five to seven years.” Well, it’s been 18 years since the last crisis. As you age, cycles become more visible: You’ve seen this movie before and begin to recognize the moment as a point on a curved line. Slowly … then suddenly, the line changes direction — for better or worse. Recently, echoes of 1999, i.e., peak dot-com, have been growing louder. I believe we’re witnessing the initial stages of the unraveling of the AI bubble … but unlike in 1999, we could be in for a twist ending.

Tipping Point

If you were raising capital in 1999, the hero wasn’t a profitable business model, but a suffix: dot-com. The defining philosophy of the era was “get big fast.” Entrepreneurs and investors believed the internet represented a once-in-a-generation opportunity to capture margin and market share. By 1999, 39% of all venture capital investments were being deployed into internet companies. My firm, Red Envelope, raised $30m at a valuation of $120m on revenues of $30m (EV/Revenues of 4x) losing $20m. Most profitable specialty retailers were trading between .8x and 1.2x revenues. Spoiler alert, the markets did eventually show up and inform me this made no sense. That same year, 80% of U.S. IPOs were related to internet companies. Pets.com, the poster child of the dot-com bubble, had the correct thesis — consumers would buy pet food and supplies online — but the company was a decade early. (See Chewy, founded in 2011.) Like many B2C internet startups, Pets.com incurred net operating losses, but spent heavily on advertising in the run-up to its IPO. In 1999 the Pets.com sock puppet mascot was so popular it was a balloon in the Macy’s Thanksgiving Day parade. A few months later, Pets.com was one of 17 internet companies to buy Super Bowl ads, up from two in 1998. The following month the company went public, raising $82.5 million. In less than a year, however, Pets.com declared bankruptcy and shuttered operations. Similar fates befell Webvan (an early iteration of online grocery delivery), eToys.com (once considered a brick-and-mortar-toy-store killer), and hundreds of other B2C startups.

Domino.com

The first dominoes to fall were B2C firms, as their business models relied on consumers ready to buy dog food via dial-up modem. The fallout took longer to reach B2Bs, as enterprise companies have longer sales cycles and stickier customers. Sun Microsystems, whose tagline was “we’re the dot in dot-com,” powered B2C startups. At its 2000 peak, Sun was valued at $205 billion — nearly as much as General Electric at the time. But as its internet clients went bankrupt, the business collapsed. Sun reported net income of $1.85 billion in 2000, but that number halved to $927 million in 2001. Sun lost $628 million in 2002 and $2.4 billion the following year. From peak to trough, the company shed 96% of its market cap; it was eventually acquired by Oracle for $7.4 billion in 2009. Along similar lines, DoubleClick was the advertising company of the era, with a $12 billion valuation. But as dot-com startups stopped advertising, its valuation dropped to $800 million, and it was soon taken private. In 2007, Google acquired DoubleClick for $3 billion, demonstrating that (some) technology developed during Web 1.0 was sound, even if the dot-com business models weren’t.

Eventually, the falling dominoes hit the infrastructure layer, causing a separate, but related, telecom crash in 2001. At its peak, Nortel Networks carried 75% of North America’s internet traffic. In the summer of 2000, just as the dot-com bubble was bursting, Nortel was valued at $230 billion. A year later more than 90% of its value had been erased. Along with Global Crossing and Lucent Technologies, Nortel had extended vendor financing to the same dot-coms that were now bankrupt. None of the three survived the crash. In retrospect, their downfalls seem obvious, but at the market peak, just as the falling dominoes were moving from B2Cs to B2Bs, 74% of stocks had buy recommendations, up from 60% four years earlier. Hype cycles aren’t just entrepreneurs, i.e., storytellers, getting out over their skis; they’re business models that incentivize consensual hallucination. In unrelated news, Goldman Sachs and Morgan Stanley (lead underwriters for SpaceX) have buy recommendations on the company with price targets of $205 and $300, respectively.

Echoes

The echoes of the dot-com and telecom implosions are deafening. OpenAI’s leaked financials reveal the company lost $21 billion in 2025. A C-suite exodus, the lawsuit from Apple, and reports that OpenAI is considering delaying its IPO until 2027 all feel very 1999. The company’s financials aren’t sustainable — for every dollar subscribers spend on ChatGPT, OpenAI spends nearly three. Its business model resembles an LLM hallucination. Case in point: OpenAI is projecting $100 billion in advertising revenue by 2030, but the company’s ad business is on pace to fall short of its own forecast by 90%, according to eMarketer. The biggest red flag, however, is Sam Altman’s request for a bailout cosplaying an investment opportunity — offering U.S. taxpayers a 5% stake in OpenAI. As my Markets co-host Ed Elson wrote last week, “The idea is to provide every citizen a share in the profits of AI. But here’s the rub: AI has no profits.” A key component of capitalism vs. socialism is citizens get to make up their own minds re which stocks they buy … or don’t. Even Senator Bernie Sanders is floating a sovereign wealth fund financed by a one-time 50% tax on AI equities. When the far left and far right agree on something, it’s almost always a terrible idea (e.g., anti-vaccine sentiment, isolationism, antisemitism, etc.). Forcing the American taxpayer to invest in a well-connected private firm isn’t socialism, it’s cronyism in the form of an SOS signal. It’s also a testament to the power of marketing, as OpenAI’s advertising spend in 2025 alone would’ve been enough to buy every Super Bowl ad spot for the past seven years.

Domino.ai

Last year I observed that circular financing deals were common toward the end of the dot-com bubble, and in retrospect, a strong signal of fragility, as one falling domino triggers a chain reaction. The circular financing deals connecting B2C and B2B AI companies with companies building AI infrastructure are easy to look past as long as customers, especially enterprise users, continue spending. According to the Economist, corporate spending on AI increased 13x from 2025 to 2026. But recently that narrative has hit a speed bump.

In May, Axios reported that an anonymous company spent $500 million in a single month after failing to put usage limits on Claude licenses for employees. Uber blew through its entire AI budget for 2026 in just four months. DoorDash, Meta, Microsoft, and Salesforce are now pivoting from “tokenmaxxing” to sobriety, i.e., limiting it to proven use cases. (FYI: Tokens are what the LLMs call chunks of data; a token equals about four characters.) To date, the best use case for AI is coding, but the dominant tech trade of 2026 — sell software stocks to buy chips — is showing signs of falling apart, suggesting that investors overestimated the scale and timeline of the AI disruption.

Meanwhile, Palo Alto Networks CEO Nikesh Arora told CNBC that widespread adoption depends on token costs coming down 20% this year and 90% next year. Meta CTO Andrew Bosworth summed up the about-face in an April memo to employees. “Nobody should be using AI tools just for the sake of using them. All motion is not progress, and token usage alone is not a measure of impact of any kind.” Good point. Except that runaway enterprise spending is what’s driving Anthropic’s $47 billion in annual recurring revenue and justifying the company’s $965 billion valuation. The pivot to measuring productivity is a good thing, but in the short term it benefits cheaper open-source models (i.e., China) that deliver 80% of the frontier model’s bang for 20% of the bucks.

Take a pause, and look (again) at the above graph. This is the grim reaper knocking at the door of every VC who has gone all-in (in ‘25 and ‘26) on AI. Zooming out, AI could end up being similar to electricity — a foundational technology that distributes value to end users. In that scenario, “the real jumps in productivity and job displacement come from new companies and processes rather than incumbents grafting new technology onto existing workflows,” John Burn-Murdoch wrote in the Financial Times. “The fact that incumbent software and knowledge work companies are finding only modest productivity gains by incorporating AI into existing workflows and organizational structures, while usage, revenue and productivity explode at Anthropic and OpenAI — companies built around AI, with products written and reviewed by it — is perhaps early evidence of the same dynamic playing out here, only much faster.”

Boom

I’m as bullish on AI as I was on the internet in 1999, but with hindsight and scars, I know not to conflate valuations with value, as transformative technologies take longer to deploy than the carnival barkers claim. The danger with the AI bubble isn’t that the technology is overhyped and still in search of use cases, it’s that the speculation is so concentrated that the 10 most valuable companies in the S&P 500 account for 43% of the index’s total market cap. In other words, when AI sneezes, the U.S. economy’s lungs may begin to fill with fluid.

Every bubble creates extraordinary wealth. The question is who keeps it. I suspect AI will create enormous value — but unlike search, social, or e-commerce, much of that value will leak past shareholders and into the hands of customers (think jet transportation, vaccines, and the PC). The biggest winners won’t be shareholders in AI companies, but the people who use the technology. In sum, we may have passed a wealth tax, just not the one AOC envisioned.

Life is so rich,

P.S.

Speaking of Silicon Valley’s hyperbolic tendencies, our head of research, Mia Silverio, breaks down tech’s obsession with humanoid robots. To read Mia’s analysis in our Extra Credit newsletter, click here.