Is This 1999? What the AI Bubble Question Gets Wrong

In March 2000, Cisco Systems briefly held the largest market cap in American history — $555 billion on under $19 billion of revenue, roughly 29x sales. Profitable, dominant, uncontested. The bear thesis was simple: the price had detached from any cash flow Cisco could plausibly ever produce. A quarter-century later, Nvidia trades near $5.2 trillion on roughly $253 billion of revenue — about 20x sales. Broadcom, the other candidate for the analog, trades near $2 trillion on $68 billion of revenue: 29x sales. Identical to peak Cisco.

Same picks-and-shovels position in a transformative IT infrastructure cycle. Same profitable, dominant, uncontested incumbents. Same uncomfortable valuations — with Broadcom literally at peak Cisco's exact multiple. The question is whether they are uncomfortably peaky in the same way Cisco was.

Short answer: public-market multiples rhyme with 2000 — matching it on some measures, exceeding it on others. The fundamentals now are meaningfully stronger. And the speculative excess of every bubble has migrated into private markets, where it's harder to see and harder to short. Less obviously dangerous than 1999, but not necessarily safer.

The Public Scorecard

Do today's broad valuation measures match March 2000? On most measures, this is closer to 2000 than any moment since.

Shiller's cyclically adjusted P/E peaked near 44 in December 1999. Today it sits near 42 — 95% of the dot-com peak and more than double the post-1900 average of 17. The S&P 500's forward P/E is near 21 against a 2000 peak of 25.

Buffett's "single best measure" — has crossed 234% versus ~140% at the dot-com peak. The top ten S&P 500 names now capture roughly 36% of the index versus 25% in 2000. The Buffett indicator is sixty-seven percent above its dot-com peak. Concentration is forty-five percent above it.

Concentration matters. A cap-weighted index is only as diversified as its constituents. When ten names are a third of the index, you're not holding a bet on American business — you're holding a thematic bet on mega-cap technology. The sector and small-cap rotations that protected diversified holders in 2000-2002 are mechanically harder to access today.

Rates cut the other way. The ten-year was 6.45% in December 1999 with an implied ERP near zero. Today it's 4.56% and the implied ERP is positive — thin, but positive.

Net: at or near the 2000 peak on multiples, dramatically above it on concentration and market-cap-to-GDP, less hostile rates. Reasonable people can call this either "bubble" or "fully valued, not bubble" from these numbers alone. The conclusion shifts when you look at what's not on the scorecard.

Rhymes with 1999

Most bubbles are easy to identify in hindsight and impossible to short in real time. The 2026 market shares more than elevated multiples with 1999. It shares the structure.

Concentration tied to a single narrative. In 2000 it was the internet; in 2026 it is AI. Owning the index is a bet on the narrative continuing — whether the 1999 investor saying "I'm diversified, I own the S&P" knew it or the 2026 investor saying the same thing does.

A capex super-cycle to build the new infrastructure. Then: fiber optic, switching gear, the long-distance backbones of WorldCom and Qwest. Now: data centers, advanced packaging, the power grid behind them. Whether the spend earns its cost of capital is the entire question.

"This changes everything" rhetoric. Multiples justified by TAM rather than current cash flow. Sell-side language that was used about the internet in 2000 with results we all remember.

Retail speculation in adjacent assets. Then: day-traded ".com" IPOs. Now: cryptocurrency, zero-day options, AI-themed meme stocks, second-wave SPACs. Same script just different vehicles.

There is also very circular financing between vendors and their customers — important enough that it gets its own section.

Why This Cycle Is Not 1999

If the structure rhymes, the contents do not. The most important difference is the quality of the businesses at the top of the cycle.

In 2000 the high-multiple darlings — JDS Uniphase, Sun, EMC, Lucent — produced little FCF against their market caps. When the recession arrived, capex collapsed and revenue evaporated. The 2026 top five — Apple, Microsoft, Nvidia, Alphabet, Meta — produce free cash flow that dwarfs anything from the 2000 era. These firms have huge moats that were completely absent before.

Roughly $400 billion in annual FCF flows from those five against an estimated $35 billion from the 2000 top-five tech equivalents. Adjusted for inflation the ratio is still six to one in real terms.

Revenue quality also differs. The 2000 darlings sold hardware on long lead times to capex-cyclical customers; the 2026 darlings sell cloud and software on recurring subscriptions with 110%+ net retention. A downturn compresses growth, but it doesn't vaporize the base. Balance sheets and ROE differ too: Lucent in 2000 had $7B of customer financing and operating debt against today's mega-caps holding hundreds of billions in net cash, and top-ten ROE has gone from high teens to above 25%.

A bull pointing at these is right on the facts. The question is whether they're enough.

The warning is the Nifty Fifty from an earlier era. The highest-quality businesses in America — Polaroid, Xerox, IBM, Avon — bought at 40-90x earnings on the thesis that quality justified any price. Investors who held them through 1973-1974 lost most of their real return over the next decade. The businesses survived. The prices did not. Quality is a reason to pay up, but not at any price.

The Private Frontier

The public scorecard tells you what institutions pay for established businesses, not what venture investors pay for two private companies that may define the cycle. In 2000 the speculative core — Cisco, Lucent, Nortel, the headline dot-coms — traded on public exchanges. The 2026 analogs, OpenAI and Anthropic, are private (for now); their numbers have to be pieced together from press releases and leaked round documents. Any honest comparison has to put them on the chart anyway.

The Duopoly Question

ChatGPT dominates the consumer side by an order of magnitude. Enterprise API is closer to a duopoly: OpenAI and Anthropic capture the bulk of paid frontier-grade spend, with Google, Meta, and xAI fighting for the rest. Beneath the duopoly the frontier itself moves every twelve months — whether the incumbents widen the gap or watch it close is the entire bull-bear debate.

A durable duopoly with monopoly rents justifies the valuations. An oligopoly racing toward commoditization does not.

Multiples That Echo Cisco

Cisco at its 2000 peak traded at 29x sales on a $555B market cap — the canonical "too high" multiple for a real business with real customers. OpenAI's March 2026 round valued the company at $852B against a ~$25B run-rate — about 34x sales, above peak Cisco. Anthropic's funding talks imply roughly $1 trillion against a $43B run-rate, about 23x sales — below peak Cisco, in line with peak-cycle SaaS comps, on revenue that compounded sevenfold in twelve months.

Growth Records, Subsidized

The bull case for those multiples is growth. The fastest public SaaS companies — Snowflake, Datadog, ServiceNow — took seven to ten years from founding to cross $1B ARR. OpenAI hit $1B within nine months of launching ChatGPT and scaled to $25B by Q1 2026. Anthropic hit $1B twenty months after commercial API GA, then went from $9B to $43B in the four months ending April 2026 — the steepest revenue ramp in software history.

The issue though is that a meaningful share of this “revenue” is subsidized.

The Circular Financing Loop

Lucent in 1999 carried $7B+ of customer financing on its balance sheet. The carriers buying its switches were largely paying with money Lucent had lent them. When carrier demand evaporated in 2001 the financing was uncollectable, and Lucent — synonymous with "blue chip" at the peak — was effectively worth zero within eighteen months.

The 2026 version: hyperscalers invest in frontier labs predominantly via cloud credits, which the labs spend back on those hyperscalers' clouds. Microsoft has committed $13B+ to OpenAI on Azure terms; Amazon $8B to Anthropic on AWS and Trainium terms; Google ~$3B to Anthropic on GCP terms. Nvidia holds stakes in neoclouds (CoreWeave, Lambda) that then buy Nvidia GPUs.

None of this is illegal and the compute spend is real. But a meaningful fraction of hyperscaler AI revenue growth in 2024-2026 has been funded by the hyperscalers themselves, and the system depends on labs continuing to raise at higher valuations to refresh the cycle. If lab fundraising slows, hyperscaler AI revenue slows immediately. Reflexive systems run fast and can implode fast.

Open Source: The Tail Risk

The risk is less that demand for intelligence collapses, but that the supply of tokens becomes effectively free.

DeepSeek's V3 crystallizes the bear thesis: a Chinese lab under U.S. export controls shipped models within striking distance of frontier capability at training costs an order of magnitude below what closed labs report. Llama, Qwen, and Mistral keep the open frontier six to twelve months behind closed.

The real bull case — enterprises pay a premium for closed models on safety, liability, agent integration, and brand — is plausibly true today. The 1999 parallel: open standards (TCP/IP, HTTP, HTML) commoditized the network layer; the value migrated up to Amazon and Google, founded years after the bubble peak and looking nothing like the names dominating it.

Does AI-stack value sit at the commoditizing model layer or the application and agent layer above? OpenAI and Anthropic prices answer one way. Open-source disruption history answers the other and remains a decent threat to frontier labs. For $6k you can buy a state-of-the-art Mac, install Qwen open-source weights, code up a harness with OpenAI Codex and have a home LLM that is 80-85% as good as Claude or Gemini in your home where your only future expense if your utility bill. Over time, the quality of DIY home LLMs will also improve.

The Excess Has Moved

The bubble-caller's error in 2026 is looking at the S&P 500 — or even the Nasdaq 100 — and concluding valuations aren't extreme enough to warrant the name bubble. On the index level that's partially correct. The S&P is materially expensive but anchored in high-quality, cash-generative businesses with effective moats. It isn't the 1999's Nasdaq.

Some speculative excess appears to have migrated into areas that are less visible in public market indices.”

Private markets first. Late-stage venture marks have expanded sharply — most visibly in the frontier-lab step-ups described above, with crossover and sovereign capital underwriting round-after-round mark-ups at terms that imply premiums to where the same companies would price as public listings.

Crypto plumbing second. Stablecoin float has crossed levels that imply systemic relevance without systemic regulation. Daily 0DTE option volume has reached levels at which the tail is, by some measures, wagging the dog.

Speculative biotech, second-wave SPACs, and small-cap "AI-adjacent" names absorb retail risk capital at valuations with no fundamental anchor.

No single market here is large enough alone to trigger a 2000-style index drawdown. Together they describe a system where excess has been redistributed rather than reduced — and where most investors' AI exposure is hiding in places they don't realize they own.

What Actually Happens in the Unwind

If the parallel to 1999 holds even partially, what comes next? "The bubble pops" did not mean "everything fell by the same amount" in 2000-2002.

The Nasdaq 100 lost ~78%. S&P 500 lost ~49%. Russell 2000 lost ~45%. Russell 1000 Value, left for dead during the late 1990s, lost ~28%. An equal-weight value portfolio finished down roughly two-thirds the loss of the cap-weight index.

The index call and the concentration call are different. ~35% of an S&P 500 position in ten mega-cap technology platforms is a much bigger bet on AI than holders realize. The post 2000 experience suggests that investors may benefit from periodically reviewing concentration risk and considering whether their allocations remain aligned with their objectives, risk tolerance, and diversification goals. At the moment a market cap weighted index does NOT offer much diversification.

Three things would change the bearish view: a sustained pullback in private funding paired with continued hyperscaler capex; open-source frontier models reaching parity on agentic benchmarks; or an enterprise-concentration disclosure from OpenAI or Anthropic showing a meaningful slowdown in net new revenue. Until those move, this cycle and 1999 share more than a chart pattern.

Conclusion

Public markets in 2026 look expensive on multiples, more concentrated than at any prior peak, supported by less hostile rates than 1999, and anchored by higher-quality businesses than their 2000 analogs. None of that prevented the Nifty Fifty from losing a decade for investors who paid quality prices at the wrong moment.

The most consequential speculative bets aren't on the public scorecard. They're in private markets, financed in part by the mega-caps themselves, growing at rates without historical precedent — partly because the growth itself is partly subsidized by the financing.

A diversified, balance-sheet-aware investor doesn't have to call the top. They have to know how concentrated their portfolio actually is, and consider the consequences accordingly.

---

Disclosure

This material is provided for informational and educational purposes only and should not be construed as investment advice, a recommendation to buy or sell any security, or a recommendation to adopt any specific investment strategy.

References to specific companies, sectors, valuation metrics, or historical market periods are included solely for illustrative and discussion purposes. Certain private company valuation, revenue, financing, and market data referenced herein are based on publicly available reports, third party sources, and Modelist Inc. estimates, and have not been independently verified. Forward looking statements, estimates, and historical comparisons are inherently uncertain and may not reflect actual future results. All investing involves risk, including the possible loss of principal. Past performance is not indicative of future results.

Modelist Inc. is a registered investment adviser. Registration does not imply a certain level of skill or training. Additional information about Modelist, including its Form ADV Part 2A, is available upon request or at www.adviserinfo.sec.gov.