AI Capital Cycle ③ — The IPO Wave, and What's Different from 1999

May 22, 2026. OpenAI confidentially filed its S-1 with the SEC (CNBC, TechCrunch broke May 20). September listing target — Sam Altman said publicly "ready by September." Goldman Sachs and Morgan Stanley lead a 21-bank syndicate. Expected single raise: $50-100B+. At $100B that would be 2-4x Saudi Aramco's $25.6B — rewriting the largest IPO record 2-4 times over.

Same month. SpaceX (with xAI bundled) debuts June 12, ticker SPCX. $75B+ raise, $1.25T-1.75T valuation. By itself the largest IPO ever, and *a month later OpenAI rewrites that record again*. Anthropic in talks May 2026 at $30B+ raise at $900B post-money (Bloomberg). Q4 2026 or 2027 IPO likely. Cursor's ARR reached $2B; xAI holds an option to acquire it for $60B.

2026 is the year of *capital markets validation for AI capital*. And the timing — looks like a 1999 déjà vu.

*But one thing is decisively different.*

You can't evaluate 2026 without seeing 1999's IPO wave.

Quantitative data:
- 1999 total IPOs: 547 (of which 289 internet IPOs)
- 1999 combined raise: $69B+ ($24.66B from internet alone)
- 1999 avg first-day pop: +68% (internet IPOs averaged +90%)
- 117 IPOs (23%) doubled on day one
- VA Linux (Dec 9, 1999): first-day +698%, all-time record. $132M raise → $9.5B market cap (12x MSFT's IPO cap). Within a year: $8.49/sh
- Only 28% of 1999 IPOs had positive net income in year 1
- 1999 IPO class 5-year survival ~50% (Ritter academic data)

Symbolic IPO endings:
- Pets.com: IPO Feb 10, 2000 ($11/sh, $82.5M raise). Liquidated Nov 6. 268 days from IPO to dissolution.
- Webvan: 1999 IPO. Bankrupt 2001. Total losses $1.5B+
- eToys: 1999 IPO. Bankrupt 2001
- Drkoop.com: 1999 IPO. Bankrupt 2002
- 1,100+ internet companies filed for bankruptcy during 2000-2002

*The 1999 IPO wave did not make the cycle top — it merely confirmed it*. The March 2000 NASDAQ peak came *after* the IPO wave; the market recognized the peak only *ex-post*. Greenspan's "irrational exuberance" came in December 1996 — three years too early.

Essence of the 1999 IPO wave: it supplied *primary capital* (cash directly to the company) to unprofitable companies. That capital funded faster growth, created bigger capex cycles, and ultimately the capex didn't get recouped — and the cycle ended.

*This is the decisive difference from 2026*.

The most important difference between 1999 internet IPOs and 2026 AI IPOs is — *the capital markets function of the IPO itself is different*.

1999: internet companies went public *to raise capital*. Pets.com's $82.5M raise was for advertising and operations. When private markets stopped funding, or funded at lower prices than IPO would — the company had *no choice but to list*. IPO was survival capital.

2026: OpenAI, Anthropic, Cursor — all are *already astronomically funded by private markets*. OpenAI is $110B+ cumulative pre-IPO. Anthropic is $64B+. Cursor might even be acquired by xAI at $60B before going public. For these companies, IPO is not *the last resort for capital* — it is the opposite: *the final step where insiders cash out*.

Look at what OpenAI did in October 2025. $10.3B secondary tender at $500B valuation. *Before the IPO, insiders already finished one round of cashout*. Parts of SoftBank's $40B round (March-Dec 2025) were also secondary. *Insiders have already sold most of what they wanted to*.

So the 2026 AI IPO wave — looks like 1999 on the surface but is *economically the opposite*.

*1999 was a wave of capital flowing into companies*. *2026 is a wave of capital flowing out of companies*.

The meaning of IPO success/failure also differs. 1999 IPO failure = company can't get cash → bankruptcy. 2026 IPO failure = insiders can't cash out, but the company is already funded and operations continue. But — *if insider cashout failure leads to lower round valuations, those same companies can't raise their next rounds*. Memo 1's circuit breaks at that node.

Michael Burry in a May 2026 X post: the AI IPO wave is *"the scene of the bloody car crash, minutes before it happens."* Inflation-adjusted — SpaceX + OpenAI + Anthropic combined raise = *roughly equal to ~300 1999 IPOs combined*. In one year, in three megacaps.

Jim Chanos in early 2026: AI is "the displacement idea of this cycle" — the same role the internet played in the 1990s. NVDA's capital spending needs are "in the hundreds of billions of dollars against $13B in revenue." A pattern that "far exceeds the approximately $100 billion in 1990s vendor financing."

BofA's Michael Hartnett: AI's share of US equity market cap has surpassed the 48% peak that defined every bubble of the past century (1920s, Nifty Fifty, Japan 1989, dotcom).

But — *all these critics also include those who described 1999 accurately*. Burry predicted GFC but has been early on multiple bear bets since. Chanos caught Enron but lost expensively shorting NVDA in 2024. Hartnett is a strategist who doesn't time.

*Marks's core insight*: you can't know *when* a cycle ends. But *how* it ends has patterns. 1-3 years after the IPO wave was the dotcom pattern. If the 2026 IPO wave peaks in June-December, the cycle peak is most likely — *2027 or 2028*.

How will OpenAI/Anthropic IPO cascade into the Korean market?

Bullish scenario (OpenAI IPO success, $1T+ valuation, first-day +30%):
- MSFT's 27% OpenAI stake → accounting mark-up ~$135B+ unrealized gain. MSFT EPS one-time impact +30-50% (FASB ASC 321)
- AMZN's ~14-19% Anthropic stake → similar mark-up $100-190B if Anthropic IPO succeeds
- MSFT/AMZN capex confidence preserved → HBM/CoWoS demand sustained → SK Hynix/Samsung operating profit stable
- KOSPI semi weight directly benefits

Bearish scenario (OpenAI IPO breaks issue price in first week — *Pets.com moment*):
- MSFT markdown → Big Tech P/E multiple compression
- Big Tech capex guidance revision → HBM price pressure downside
- SK Hynix/Samsung multiple compression + KOSPI cascade
- Korean market becomes *the first external victim of US capital markets validation failure*

How Korean investors can directly bet on OpenAI/Anthropic:
- Private secondary (Forge, Hiive, etc.) — *but* Anthropic warned May 12, 2026: "these platforms are not authorized; transfers without board approval are void." Risk for some Korean HNW investors.
- Post-IPO — normal US listing access (Kiwoom, Mirae Asset, Toss). Consider FX and settlement.
- Indirect: MSFT (OpenAI) + AMZN (Anthropic)

Korean AI IPO candidates: Upstage. Underwriter selected 2025, KOSDAQ filing H2 2026. Own model Solar Pro 2 (31B params) — claims top-10 frontier ranking. Expected valuation ₩2-3T+. Korea's first generative AI IPO.

SK Telecom Anthropic stake: $100M invested Aug 2023. ~0.3% stake. At 2026 $900B valuation, mark-to-IPO ~₩2.6-3T+. Equivalent to 13-15% of SKT's ₩20T market cap. The single Korean stock most directly exposed to OpenAI/Anthropic IPO.

The single proposition of this memo: IPOs don't make the cycle top. They merely confirm it.

The 1999 IPO wave — through Pets.com, Webvan, Drkoop.com — *didn't make* the cycle top; it was *the ex-post signal that announced it*. The NASDAQ peak came *after* the IPO wave's peak. The 2026 AI IPO wave will *also not make* the cycle top. It will merely show how far the cycle has gone.

But a decisive difference. 1999 was a *primary capital* wave; that capital created capex and extended the cycle. 2026 is an *insider liquidity* wave. Insiders already cashed out via secondary. *IPO itself is not capex funding*. So the capital-markets shock of 2026's IPO wave is likely to appear *with a delay* relative to 1999.

What then to watch?

*First*, OpenAI's first-week price action. If it breaks issue price (i.e., falls below issue price within the first day or week) — that's a mega-version of the Pets.com moment. The single biggest catalyst in 2026 capital markets.

*Second*, the first down round. If a smaller company than OpenAI — Cursor, Cognition, Lovable, Replit, or Mistral — accepts a next round at a lower valuation, that is *the first crack* in Memo 1's circuit.

*Third*, MSFT/AMZN mark-up vs markdown. After OpenAI/Anthropic IPO, how does Big Tech's mark-to-market unrealized gain/loss affect EPS and forward guidance?

*In Marks's framing — we can't know. But the dotcom pattern was 1-3 years after the IPO wave. If 2026's IPO wave is the timer, that 1-3 years is 2027-2028.*