Wang Chuan Investment Series
How to Stay Calm After Your Neighbor Made 30x on Memory Stocks
From the AI valuation frenzy, cloud capex surge, and memory chip mania — deconstructing a cycle game with sharply asymmetric returns and risks
Source: chuan.us · Data marked with Source comes from the source text
I · Chapter One
From AI Valuations to the Semiconductor Rally
Timeline2021-2025
The Snowballing of LLM Funding
Anthropic raised USD 72.3 billion cumulatively over five years; OpenAI cumulatively raised USD 190 billion — yet revenue still fell far short of losses.Source
Valuation First
2025-2026
Monetization Without Converging Losses
Tools like Claude Code drove rapid monthly revenue growth, but competition is fierce, switching costs are low, and price-war pressure is immense.
The More You Sell, the More You Lose
2026-05
Memory Stocks Rocket
The Philadelphia Semiconductor Index rose 150% in one year; SNDK closed at USD 1,562, 38 times its price a year earlier.Source
Capital Torrent
Outcome: The valuation inflation of AI companies triggered a capex arms race among cloud giants, ultimately transmitting funding pressure to the memory chip bottleneck.
* * *
How Capital Transmits
II · Chapter Two
The Transmission Path of the Capex Torrent
Resource FlowGlobal AI Capex in 2026
7650BSource
USD
A collective bet by cloud vendors, AI labs, and the supply chain
Big Four Cloud AI Capex
Amazon, Google, Microsoft, and Meta account for ~85% of the global total
6390BSource
Memory Chip Increment
Roughly 30% of AI capex, about USD 225 billion per quarter
900BSource
Capex as Share of OCF
Capex / OCF rose from 62.9% to 86.2%
86.2%Source
Free Cash Flow Turns Negative
Even the deep-pocketed are forced to issue debt or sign long-term leases
NegativeSource
Amazon: Q1 capex USD 44.2 billion, free cash flow negative USD 18.2 billion, and issued USD 53 billion in long-term debt to cover the gap.Source
Key judgment: Cloud vendors are pouring an ever-larger share of operating cash flow — and borrowed money — into AI capex. This cannot continue indefinitely.
* * *
Three Hard Constraints
III · Chapter Three
The Triple Bind Facing Cloud Giants
Deadlocks3.1
Financial Bind: Paper Wealth and Hidden Debt
Step 1
Microsoft, Amazon, and Google contributed cloud compute credits to AI labs at valuations implying theoretical returns of 10x or more
Microsoft, Amazon, and Google contributed cloud compute credits to AI labs at valuations implying theoretical returns of 10x or more
↓
Step 2
But unrealized paper gains require an IPO exit, which would expose the real financials
But unrealized paper gains require an IPO exit, which would expose the real financials
↓
Step 3
More insidious are off-balance-sheet lease commitments: the top five cloud giants total USD 822.3 billion, up 220% in fifteen monthsSource
More insidious are off-balance-sheet lease commitments: the top five cloud giants total USD 822.3 billion, up 220% in fifteen monthsSource
Result: Credit ratings are under pressure, room to issue new debt shrinks, and Oracle is already just two notches above junk.
3.2
Physical Bind: Data Centers Cannot Catch Up with GPU Deliveries
Step 1
Data center build-outs require power supply, grid connection, and transformer delivery — bottlenecks lasting three to five years
Data center build-outs require power supply, grid connection, and transformer delivery — bottlenecks lasting three to five years
↓
Step 2
GPU delivery takes only six to twelve months, leaving large numbers of chips idle in warehouses
GPU delivery takes only six to twelve months, leaving large numbers of chips idle in warehouses
↓
Step 3
Google, Meta, and Amazon added USD 107 billion in assets not yet in service; if 30% are GPUs, that equals one million Blackwell GPUs collecting dustSource
Google, Meta, and Amazon added USD 107 billion in assets not yet in service; if 30% are GPUs, that equals one million Blackwell GPUs collecting dustSource
Result: Capital has been front-loaded, but physical capacity cannot be released in sync, creating a mountain of stranded assets.
3.3
Commodity Bind: High Margins Destroy Themselves
Step 1
Memory is one of the most commoditized tech products: no brand premium, and prices swing with global supply
Memory is one of the most commoditized tech products: no brand premium, and prices swing with global supply
↓
Step 2
After September 2025, HBM demand crossed a tipping point; memory makers chose price hikes over capacity expansion, and gross margins surged
After September 2025, HBM demand crossed a tipping point; memory makers chose price hikes over capacity expansion, and gross margins surged
↓
Step 3
Sandisk's Q1 gross margin hit 78.3% versus 22.5% a year earlier; Micron reached 74.4% versus 36.8%Source
Sandisk's Q1 gross margin hit 78.3% versus 22.5% a year earlier; Micron reached 74.4% versus 36.8%Source
Result: High prices stimulate new capacity and substitution technologies; long-term agreements are fragile, and once supply catches up prices and profits must fall.
1994 – 2008
Historical Echo: Iomega’s 160x Gain and 97% Drawdown
Iomega launched its 100 MB Zip drive in late 1994. Sales soared from USD 140 million to USD 1.21 billion by 1996, and the stock rose from about USD 2 to an equivalent of USD 330 by May 1996 — a return above 160x. Once CD-R and other substitutes took over, gross margin fell from 31% to 25% and the company slipped into loss. By 2008 EMC acquired Iomega for USD 210 million, a 97% drop from its peak market cap.
Source
Where the Three Constraints Meet
Cloud vendors' money, data center physics, and the commodity nature of memory together mean the feast cannot last forever.
Lose a hundred million dollars, and it is your problem. Lose a trillion dollars, and it becomes the world's problem.
Wang Chuan
* * *
Positive Feedback Loop
IV · Chapter Four
Reflexivity: Prices and Beliefs Feed Each Other
LoopReal-economy reflexive demand builds like a tsunami, pulling financial-market reflexive speculators into the game; once it hits the hard constraints of supply and liquidity, the reversal cascades like an avalanche.
* * *
Core Mismatch
V · Chapter Five
Bullwhip Effect: Long-Cycle Supply Meets Elastic Demand
MismatchAn Asymmetry of Time Scales
The memory industry is squeezed between long-cycle supply and short-cycle demand: demand can collapse within months because of prices, substitution, or recession, but supply, once started, cannot brake.
Supply cycleData center / new capacity online
Demand cycleOrders / market sentiment reversal
012345 years
!
The counter-intuitive point of the bullwhip effect:
When orders accelerate, shortages worsen and tempt you to place even more orders. Once supply and demand balance, even if you cancel orders immediately, goods already in production keep flowing into the market, and oversupply keeps getting worse. DRAM chip prices have fallen more than 80% in a short period at least six times in history.
* * *
Conclusions and Takeaways
VI · Chapter Six
A Transferable Decision Framework
TakeawaysThe most dangerous feature of a cycle top is not a compelling story but the simultaneous appearance of low P/E, high margins, and huge gains — which tricks speculators into believing they are doing value investing. When every player is making big money, it takes only one new entrant willing to absorb long-term losses, or one technology breakthrough, to shift the supply-demand balance.
1
Distinguish Revenue from Profit
ARR is just last month's revenue multiplied by twelve, not annual total profit; a highly valued company that keeps losing money and depends on external funding is not a sustainable business model.
2
Beware Paper Gains and Hidden Debt
Uncommenced data-center lease commitments, SPV residual-value guarantees, and compute-credit equity injections all hide real financial risk off the balance sheet.
3
Commoditized Goods Have No Moat
Memory chip prices have fallen more than 80% many times; memory stocks have fallen 95% or gone bankrupt. High margins invite capacity expansion and substitution; long-term agreements crumble in the face of spot prices.
4
Recognize Reflexivity and Mental Tattoos
Treating short-term strength as eternal strength and treating quick money as normal are the two most common mental tattoos at bubble tops. Reversals often complete in days or weeks.
* * *
Editorial Verdict
A Game with Sharply Asymmetric Payoffs
Upside
Staying in the game may still deliver another 2x or more
Downside
Once the trend reverses, the risk is an 80%+ drawdown and a 25-year wait to break evenSource
Advice
Do not mistake short-term windfalls for skill, and never combine leverage with a cycle top
* * *
Editor's Note
How to Read This Article
Why it matters
Wang Chuan strings AI valuations, cloud vendor financials, data-center physical constraints, memory commodity dynamics, and behavioral finance into a causal chain, supporting the argument with numerous verifiable numbers (capex, free cash flow, gross margins, lease commitments). Compared with commentary that is purely bullish or purely alarmist, this article offers a transferable analytical framework: how to identify reflexive demand, how to read off-balance-sheet debt, and how to judge the top of a commodity cycle.
Caveats
Some key figures (such as Anthropic's monthly loss of USD 1.1-1.7 billion) are estimates the author derived by querying AI based on limited public information, not official company disclosures. Some forecasts (Anthropic's year-end IPO probability, OpenAI's listing timing) are based on market rumors. The seventh installment's portrait of the neighbor's outcome is a behavioral-finance extrapolation, not a deterministic conclusion. Readers should treat these as assumptions rather than facts.
Transferable insight
In any market where capital floods in, asset prices rise, narratives strengthen, and more capital floods in, the real top is determined not by the story itself but by whether supply faces hard constraints that money cannot instantly solve, and whether demand is driven by exploration, panic, and liquidity. When both conditions hold, the reversal threshold is usually closer and steeper than most people expect.
Source excerpt