AI Capital Cycle ① — Money in the Mirror

September 22, 2025. NVIDIA announces it will invest up to $100 billion in OpenAI. The structure: capital is gated to each gigawatt of data center deployment. About 4-5 million GPUs. If one GW equals roughly $35 billion in NVIDIA chips, then 10 GW = $350 billion in NVIDIA revenue — partly funded by NVIDIA's own equity check.

The same autumn. Microsoft and OpenAI close a restructured deal. OpenAI commits to buy $250 billion of Azure compute over six years (2025-2030). Separately, OpenAI signs a $300 billion, five-year contract with Oracle.

Microsoft is NVIDIA's single largest customer. NVIDIA's FY26 10-K reveals: one customer = 22% of revenue, another = 14%. Together, two customers account for 36%. The market reads these as Microsoft and Meta.

The money came around in a circle. NVIDIA puts capital into OpenAI; OpenAI commits compute to Microsoft and Oracle; Microsoft and Oracle buy GPUs from NVIDIA; NVIDIA recognizes revenue. The circuit is complete.

*When the buyer and the seller are the same person, what does the market price actually mean?*

In April 1996, AT&T spun off its telecom equipment business and named it Lucent Technologies. The years that followed were a golden age for telecom equipment. Fiber was being laid. Internet backbones were being built. New telecom operators (CLECs, ISPs, long-haul carriers) emerged at extraordinary speed.

These new entrants — WinStar, NorthPoint, ICG, Global Crossing — had no capital. To buy equipment, someone had to lend them money. Lucent was that someone.

The structure was simple. Lucent extended credit to the new carriers; the new carriers used that credit to buy Lucent equipment; Lucent booked the revenue. When a carrier went bankrupt, Lucent took a provision. But while the cycle was running, everyone won.

In its fiscal year 2000 SEC filings, Lucent disclosed approximately $7 billion in customer financing commitments. $1.6 billion was actually drawn. Lucent publicly defended this as "normal industry practice," and an executive told one earnings call that the company would become "more like a bank" — meaning it would expand vendor financing aggressively.

The circuit ran flawlessly through 1999. Lucent's stock hit $82.31 on December 20, 1999. Market cap: $258 billion. It was a top-10 company in the world by market value.

But the end of the circuit started at its weakest node.

In April 2001, WinStar Communications filed for bankruptcy. Lucent's exposure was $700 million drawn from a $2 billion commitment. Almost the entire $700 million was written down. (Years later, a court would order Lucent to pay roughly $300 million more to WinStar's bankruptcy trustee.)

Other customers fell in 2001 — NorthPoint, ICG, Global Crossing, PSINet. WorldCom filed in July 2002. Lucent's FY2001 provisions: $2.2 billion. FY2002: $1.3 billion. Combined: $3.5 billion.

The circuit stopped. Lucent's stock hit $0.58 on October 11, 2002. From peak: -99.3% (-95% on a split-adjusted basis). Headcount fell from 165,000 to 30,500. CEO Richard McGinn was fired in October 2000. CFO Deborah Hopkins was replaced in May 2001. In 2004, the SEC settled revenue-recognition fraud charges for $25 million.

Lucent did not survive. It merged with Alcatel in 2006. The merged entity was eventually absorbed by Nokia.

Lucent's story is extreme. Accounting fraud, CEO firing, 99% market cap evaporation, eventual dissolution. In one sense, it's too clean a cautionary tale — *because the business was managed badly.*

But the real lesson of that cycle is not Lucent. It's Cisco.

On March 27, 2000, Cisco Systems passed Microsoft to become the most valuable company in the world. Market cap: $569 billion. Stock price: $80.06 (split-adjusted). Forward P/E: about 150x. Everyone knew Cisco routers were building the internet. Everyone thought they should own Cisco stock.

On October 8, 2002, Cisco hit $8.60. From peak: -89%.

Up to here, the story sounds like Lucent's. But look at revenue:

- FY2000 (ending July 2000): $18.9B
- FY2001: $22.3B (+18% growth)
- FY2002: $18.9B (-15%)
- FY2003: $18.9B (flat)

*The business didn't fail.* FY01 revenue was actually a record high. After modest declines in FY02-03, Cisco kept growing — FY24 revenue was $53.8 billion, roughly 3x the FY00 peak.

But the stock fell -89%. And then, 25 years later, on December 10, 2025, the stock finally reclaimed $80.25 — the 2000 peak. 25 years, 8 months, 13 days.

This is the real lesson of the cycle.

Cisco didn't fall -89% because the business failed. It fell because the circuit stopped. More precisely: *because the multiple that priced in the circuit's permanence compressed.* A 150x P/E embedded an assumption that the circuit would run forever; when it stopped, that multiple compressed to single digits.

Business success and stock success are different things. At a $569 billion peak market cap, investors who bought Cisco at the top took *25 years to recover their capital to the peak price.* Over that time, Cisco's revenue tripled and the company generated $10+ billion in operating cash flow every year. But the entry price was too expensive.

Whether NVIDIA becomes Cisco or Lucent, we don't yet know. But Cisco's 25 years teaches one thing clearly:

*A multiple that prices in the permanence of the circuit collapses when the circuit stops — even if the business never fails.*

Once you've seen the 1999 circuit, the 2025 circuit looks different. Drawn as a diagram, it's simple.

Beside this circuit sits a second one with almost identical shape.

Amazon has invested roughly $16 billion cumulatively in Anthropic since September 2023. In late 2025, it committed up to $20 billion more on commercial milestones. Anthropic uses Project Rainier — a roughly $11 billion AWS-built cluster in Indiana with about 500,000 Trainium2 chips — to train Claude. AWS recognizes the compute revenue.

Exactly the same pattern as Microsoft-OpenAI. Just different protagonists.

And there's more around the edges. NVIDIA didn't only invest in OpenAI. It owns 6-7% of CoreWeave ($2B PIPE plus IPO anchor order), participated in xAI rounds, and has stakes in Mistral, Lambda, Inflection, Perplexity, Reka, Cohere, Wayve, Figure. In October 2025, AMD signed a 6 GW commitment with OpenAI — in exchange, AMD gave OpenAI warrants on roughly 10% of AMD stock (160 million shares at $0.01 per share). The model company that buys GPUs gets equity in the GPU company. The circuit twists one more time.

Here's the full picture.

Beyond the circuit, look at the funding source. In 1999, telecom carriers funded capex with corporate bonds and bank loans. In 2025, hyperscalers fund it from operating cash flow — at least on the surface. But the 2026 guidance shows that self-funding pattern breaking for the first time.

Placing the two circuits side by side, an honest comparison has to do both — *what's the same* and *what's different*.

The differences come first. In 1999, Lucent's vendor financing was, in accounting terms, *receivables and loans*. When customers couldn't pay, provisions had to be taken, and they hit the P&L. In 2025, NVIDIA's investment in OpenAI is, in accounting terms, *equity*. Marks-up and marks-down can happen, but they're separated from revenue recognition. The funding source is also different. In 1999, telecom carriers issued roughly $1.6 trillion in corporate bonds to fund capex. In 2025, hyperscalers convert ~70% of operating cash flow into capex — the gap is filled by bonds ($108B newly issued by Big Tech in 2025), but the ratio is incomparable to 1999.

So NVIDIA's official position is — in accounting terms — true.

This accounting truth deserves recognition. But economically, how does it look? Jim Chanos's framing is the cleanest:

> *"[What NVIDIA does is] channel money into unprofitable companies that then use those funds to buy chips."*

Equity in → capital becomes compute commitment → compute commitment becomes GPU purchase → GPU purchase becomes NVIDIA revenue. Even with five layers of accounting separation, economically *the same company is ultimately making its own revenue with its own money.* Chanos calls this a "pattern that far exceeds the approximately $100 billion in 1990s vendor financing."

So the similarities are simple:

1. Buyer and seller are the same person (or the same capital pool)
2. New entrants (OpenAI, Anthropic, xAI) receive capital to create infrastructure demand — same structure as Lucent giving WinStar capital to buy Lucent equipment
3. While the circuit runs, every participant wins — *without any validation that the business actually works*

Marks has a framing he likes — *what we know / what we can't know / what we can't know but can estimate*. Applied to the AI circuit:

What we can know is measurable fact. The exact money flows in the circuit. The fact that 36% of NVIDIA revenue comes from two hyperscalers (FY26 10-K). Big 5 combined 2026 capex ~$700B. The fact that OpenAI's revenue is far below its compute cost. These don't get debated.

What we can't know is the cycle's timing. Exactly when the circuit will stop. What portion of AI revenue is genuine end-user demand versus recycled own-capital. Whether the peak is 2026, 2028, or 2030. This is fundamentally unknowable — anyone who claims to know is lying.

What we can estimate is *where the first crack is most likely to appear*. The weakest node in the circuit is the first model company that fails to raise its next round. OpenAI and Anthropic are strong. But beyond those two — xAI, Mistral, Inflection, Cohere, Stability — if any one of them takes a down round or fails to fund, one node in the circuit breaks. That signal is measurable.

The second estimable signal is the credit spread on Big Tech corporate bonds. If the $108B issued in 2025 grows to $200B+ in 2026, credit markets will start pricing the credit differently. Spread widening is the circuit's second crack.

The third is the moment when NVIDIA's forward guidance shows sequential deceleration in data center revenue. The quarter NVDA itself says "next quarter's growth will be weak" — that's where the *third crack* shows.

The fourth — and most important — signal, we'll cover in the final memo of this series. The quarterly data from Anthropic's Economic Index on whether AI is actually displacing labor.

This memo doesn't answer whether AI is a bubble or an industrial revolution. That question is fundamentally unanswerable. And every memo — bullish or bearish — that claims to know the answer is selling its own conviction.

What this memo does answer is something different. *What is the structure of the capital flow we are watching?* It is a circuit. A closed loop in which the buyer puts capital into the seller, the seller uses that capital to buy from the buyer, and the buyer recognizes revenue. In accounting terms, it isn't vendor financing. In economic terms, it's the same.

In 1999, the same circuit ended in a -89% drawdown. But *not because the business failed.* Cisco's revenue tripled over the 25 years that followed, and the company generated over $10 billion in operating cash flow every year. What failed was the multiple that priced in the permanence of the circuit. That made the -89%.

The same possibility is open to NVIDIA, Microsoft, OpenAI, SK Hynix — all of them. While the circuit runs, every participant gets rich. When the circuit stops, even if the business never fails, the multiple collapses. And that difference decides everything.

*When the buyer and the seller are the same person, the market price can — for a while — be anything. As long as the circuit runs.*

As long as the circuit runs.