One-line takeaway: "Bubble" is not a single-choice question
Whenever someone shouts "AI is a bubble," first ask: "Which layer are you talking about?"
Industry bubble? Asset-price bubble? Or earnings bubble? Mixing these three is like lumping onion, garlic, and chives together as "allium plants" — technically correct, but pick the wrong one in the kitchen and the dish is ruined.
---
Layer 1: Industry Bubble (Falsified)
Definition: The capex being poured in has no real users; infrastructure sits idle.
This is the easiest layer to refute.
- Anthropic's inference gross margin is 70%. Not 7%, not 17% — 70%. Keep 70 out of every 100 dollars of revenue after direct costs. That's higher than Salesforce selling software, and far above Starbucks selling coffee.
- ARR (annual recurring revenue) grew 10x in a year. Customers aren't just "trying it out" — they sign long-term contracts, prepay, and queue for token quotas.
- Micron just signed a $22 billion long-term supply agreement — customers willing to lock in multi-year capacity, unheard of in traditional memory.
- The hardware chain rose 2-5x in four months. This isn't idle capital speculating; it's physical supply shortage. An EUV lithography machine costs $200 million, and ASML's order book stretches three years out. This isn't "hype" — it's "can't be built fast enough."
- The memory sector ended last year at just 3-4x P/E — the market pricing in near-bankruptcy. This year it re-rated to 10x P/E: not a crazy surge, but a recovery from "extremely undervalued" to "reasonable but still cheap."
- Analogy: oil-shipping stocks at the peak of the 1970s oil crisis traded at only 3-5x P/E, because the market knew such windfalls weren't sustainable. Memory faces the same logic — the market is already discounting it.
- Leading platforms trade at 7-8x P/E, versus the Nasdaq average of ~25x and Nvidia above 50x. Memory valuations look almost shabby.
- 2024-2027: Extreme supply-demand mismatch; memory makers print money
- 2027-2029: New capacity gradually releases; prices start to loosen
- 2029+: Supply and demand balance; ROIC normalizes (20-30%)
- Howard Marks, *Is It a Bubble?* (Oaktree Memo, Dec 2025)
- Jim Chanos AI Bubble Warning (Jun 2026)
- Micron FY2026 Q3 results: revenue $41.46B (+345.7% YoY), gross margin 84.9%
- Micron $22B long-term supply agreement disclosure
- South Korea market's 5th circuit breaker of 2026 (June 26; SK Hynix -9%)
- CICC: global AI is not in a full-blown bubble, but crowding is high
Conclusion: The industry-bubble layer is empty when peeled. Demand is real, payment is real, capacity bottlenecks are real.
---
Layer 2: Asset-Price Bubble (Valuations Aren't Absurd)
Definition: Stock prices far exceed intrinsic value; P/E ratios defy gravity.
This layer needs careful peeling.
Jim Chanos (the man who shorted Enron and Tesla) recently warned about an AI bubble too, but with a sharp angle: not shorting chip stocks, but shorting cloud vendors' "depreciation time bomb." Chipmakers recognize revenue and profit immediately; cloud vendors capitalize the costs — pretty balance sheets now, ugly income statements once depreciation kicks in.
Conclusion: This layer has some water in it, but not much. The market is already pricing memory with a "cyclical stock" framework.
---
Layer 3: Earnings Bubble (Real, but Physically Capped)
Definition: Supply-demand mismatch has produced extremely high return on invested capital (ROIC), but that return is unsustainable.
This layer is solid.
Memory's ROIC is currently 200-400%. For reference: Apple is ~30%, Moutai ~25%. This isn't a "good business" — it's a money printer.
But the printer has physical limits:
1. A new fab takes 3-5 years from decision to mass production. Today's high prices are the result of 2021-2022 investment decisions; current expansion plans won't release capacity until 2028-2029. 2. EUV equipment is monopolized by ASML, with capacity growing at most ~30% per year. Not unwilling — the lithography machines simply can't be built faster. 3. Customer qualification locks. HBM (high-bandwidth memory) can't just be bought and used; qualification takes 6-12 months. Once Micron, SK Hynix, or Samsung customers qualify a supplier, switching costs are enormous.
So the "earnings bubble" script:
Three years is the physical cap — not a prediction, a hard constraint of fab construction cycles.
---
Howard Marks' Framework: Is It Different This Time?
Oaktree's Howard Marks wrote a memo last year titled *Is It a Bubble?* His core point:
> "Ours is an extraordinary moment in world history. A transformative technology is rising, whose proponents claim it will change the world forever." > > "Early participants reap huge gains, while onlookers feel intense envy and regret, piling in driven by fear of missing out."
That's the classic bubble recipe. But Marks also reminds us: not every bull market is a bubble, and not every bubble bursts.
What's different this time: 1. Cash flow already exists (Anthropic's 70% gross margin, Micron's 84.9%) — not the "eyeball economy" of 2000 2. Healthier capital structures (Big Tech total debt of $385 billion; leverage ~20% below prior heavy-investment cycles) 3. Long-term contracts lock in demand (Micron's $22B; Nvidia near $100B in purchase commitments)
But the risks are real too: 1. The depreciation bomb (cloud vendors' capitalized costs will eventually enter the depreciation cycle) 2. End-demand suppression (memory prices up 165%; smartphones up 7%, laptops up 15% — dampening adoption) 3. Geopolitics (South Korea's market hit its 5th circuit breaker of the year; Middle East conflict affects energy costs and rate paths)
---
Summary Analogy: AI Is Not Dot-Com 2.0
| Dimension | 2000 Dot-Com Bubble | 2026 AI Cycle | |---|---|---| | Revenue | Almost none | Real and growing fast | | Gross margin | Negative | 70-85% | | Valuation | Unlimited (no earnings) | Memory at 10x P/E, already discounted | | Capital structure | High leverage, burning cash | Strong Big Tech balance sheets | | Demand certainty | "If a million people visit" | Multi-year locked contracts | | Capacity constraint | None (infinite bandwidth) | Physical cap (fabs take 3-5 years) |
AI looks more like oil-shipping stocks in the 1970s oil crisis — not "fake," but "cyclical." The windfall comes from supply bottlenecks; it fades when supply releases. The key: can you convert windfall profits into a moat (technology, customer relationships, scale) during the boom?
---
Takeaways for Ordinary People
1. Don't think in a binary bubble/no-bubble frame. Three layers, three different conclusions. 2. Memory's "earnings bubble" has a 2-3 year window left. Determined by physical cycles, not sentiment. 3. Watch the "depreciation time bomb." Cloud vendors (Microsoft, Google, Amazon) haven't yet absorbed AI infrastructure depreciation; the hit should show up in 2027-2028. 4. Profit distribution along the chain is shifting — from "shovel sellers" (chipmakers) to "gold diggers with shovels" (application layer). But that shift takes time.
---