AI Hype Cycle

1. The Five Core Contradictions of the AI Hype Cycle

1.1 Revenue vs CAPEX order-of-magnitude mismatch

Mechanism: hyperscalers bet on 10x LLM inference demand over the next 2-3 years. Even if true, sustainable CAPEX only supports $130-160B/yr (see Section 3, 40% compute margin x 5yr asset life). Note: although ARR doubled-tripled vs prior estimates, CAPEX also doubled, so the net gap magnitude is essentially unchanged.

1.2 Token deflation -- pricing power evaporating

• GPT-4 2023-03 to Claude Sonnet 4.6 2026-05: per-token API price has dropped 85-92%
• Open-source threat: DeepSeek R1 / Qwen / Llama 4 squeeze frontier performance into the $0.2-0.5/M token range
• Result: even with 10x token volume, revenue can only grow 1-2x
• Core: LLM is not a utility, it is a commodity. Commodities have no pricing power.

1.3 GPU depreciation mismatch -- long-lived asset vs short-lived technology

• Accounting depreciation: 5-6 years (MSFT extended from 4 to 6 years in 2024, mainly to improve GAAP earnings)
• Technical depreciation: H100 vs B200 vs B300 vs Rubin -- generation cycle 18-24 months, each generation 2-3x perf/$
• 3 years out, H100's economic value vs current GPU = ~20-30% residual
• Risk: once demand turns, the installed GPU fleet write-down is on the order of $100-200B, hitting equity directly

1.4 Circular financing -- accelerating in 2026

Customers and suppliers are the same parties:

Analog: 1999 Lucent vendor financing, blew up in 2000-2001. The 2026 structure is more concentrated than 1999 -- 4 hyperscalers + NVDA channeling over $100B into 2 frontier labs (OpenAI + Anthropic). Even if 30% of this turns problematic = $30B asset write-down cascade.

1.5 Use case vs valuation TAM inversion

• Real AI revenue (2026 Q1 public disclosures): coding assistants (Anthropic Claude Code alone $2.5B ARR 2026-02, industry-wide ~$8-15B), customer service AI (~$3-5B estimate), enterprise copilot (~$10-15B, but mostly to MSFT/CRM middlemen, not LLM providers) -- note: segment-level numbers are T3-T4 estimates, not GAAP
• NVDA market cap (~$5.36T) implies LLM TAM of $500B-1T by 2030
• Bottom-up actual 2030 SOM (sigmoid extrapolation from current growth): $300-500B -- current ARR run-rate is already $65-80B (3-6x to $500B target)
• Gap: ~1.5-3x overvaluation in implied TAM, but still overshoots

2. Bottom-up LLM Revenue Algorithm (not copying sell-side TAM decks)

2.1 Revenue decomposition formula

2.2 2026 actual run-rate (public disclosures 2026-04, T1/T3)

OpenAI + Anthropic combined public ARR is $54B. Google/xAI/Mistral category T4 estimate ranges are wide (no standalone Gemini/Grok revenue disclosure), but even at ceiling, total stays under $80B run-rate.

2.3 Growth curve: still hyperscale today, sigmoid inflection likely 2027-2028

2024 -> 2025 -> 2026 trajectory:

• 2024 industry LLM ARR: ~$10-15B (Anthropic ~$1B late-2024, OpenAI ~$5-7B)
• 2025 EOY: ~$30-45B (Anthropic Feb-2026 $14B run-rate implies $5-10B EOY 2025; OpenAI ~$10-15B EOY)
• 2026-04 run-rate: ~$65-80B (OpenAI $24B + Anthropic $30B + others)
• 2026 EOY (E): ~$100-130B even if Anthropic growth rate halves to 40x annualized
• 2027E: ~$150-220B (assuming sigmoid inflection begins, growth rate decays)
• 2030E SOM: $300-500B

Anthropic alone grew 80x in 12 months = still in the hyperscale phase. Sigmoid inflection likely 2027-2028. $500B 2030 TAM is achievable in the base case, no super-bull required. However CAPEX also doubled ($700-725B), so gap magnitude is unchanged; only timing is pushed back.

2.4 Stress test (bull case)

For LLM revenue to reach $500B by 2030 (i.e., NVDA market cap's implied TAM):

• Current ARR run-rate $65-80B (2026-04), only 6-8x to $500B
• Sigmoid: 2026-04 -> 2030 = 4 years, requires only 60% compound annual growth (vs Anthropic 80x in 12 months, OpenAI ~2-3x in 12 months)
• Enterprise penetration still needs to reach 30%+ (currently ~5-10%)
• ARPU must not crash further -- token deflation remains a headwind, but enterprise ARPU is anchored on seat licenses rather than token price, mitigating
• No AGI capability leap required -- current frontier (Claude Sonnet 4.6 / GPT-5 tier) sufficient to support enterprise penetration ramp

$500B 2030 TAM probability ~40-50%. However NVDA $5.36T market cap still implies $1T+ 2030 LLM TAM to be justified (factoring in NVDA's 30-40% compute margin take), so this portion of the thesis still holds.

3. Sustainable CAPEX Algorithm

3.1 Sustainable CAPEX formula

3.2 Three-scenario substitution (revenue base anchored to 2026-04 $65-80B run-rate)

Conclusion: even in Super Bull (AGI arrives + LLM captures 25% of total software TAM), current CAPEX run-rate still exceeds sustainable by 17%. Base case (highest probability) exceeds sustainable by 4.5x. The magnitude of the core contradiction is irreconcilable.

3.3 Prerequisite for AI civilizational progress: 40% margin floor for tax-collecting enterprises

AI is not narrative; it is the infrastructure layer. The infrastructure is carried by tax-collecting enterprises -- NVDA (compute tax) + MU (memory tax) + TSM (foundry tax) -- each AI workload that passes through is taxed. The continued progress of AI civilization requires these tax-collecting enterprises to be able to fund the next generation of R&D + CAPEX (Rubin / B400 / HBM4 / 18A foundry nodes).

Margin floor mechanism:

• ~40% net margin is the mandatory threshold for a "tax-collecting enterprise". 40% is the median compute_margin in the sustainable CAPEX formula (see 3.2 Base case 40%) -- not arbitrary, but the sustainability math derived lower bound
• Below 40% = R&D + next-gen fab + EUV equipment investment cannot be self-funded = must rely on external financing (government / sovereign / debt) = infrastructure progress stalls OR fully commoditizes
• NVDA: current gross margin ~75% / operating margin ~65% (FY2026 guidance) -- far above floor, easily funds three generations of Rubin / Vera Rubin / Feynman roadmap
• MU: HBM cycle peak gross margin ~40-45% (vs historical commodity DRAM 5-15%) -- just at floor; once HBM prices drop or cycle turns -> margin returns to 25-30% = below floor = HBM4 / 1c node investment constrained
• SNDK: pure NAND commodity, gross margin 25-30% peak -- always below floor; this is why SNDK has no NAND next-gen ramp capability; industry NAND progress relies on Samsung/SK Hynix integrated model + government subsidies
• TSM: gross margin ~55-60% peak -- far above floor; this is why 2nm/A16 nodes can continue to advance

Thesis implication: for any AI-concept stock, check its margin position in the value chain. >40% = real tax collector = LONG candidate (within cycle); <40% = commodity being taxed = SHORT candidate (especially at cycle top). NVDA's high margin is the prerequisite for continued AI civilizational progress, not a bubble signal. The real bubble signal is when margin floor is breached without valuation adjustment -- MU is on this edge.

Corollary: if NVDA margin is forcibly compressed (e.g., antitrust / regulation / DOJ split) = AI progress directly decelerates = this is the true tail risk of "AI nationalization", far exceeding valuation correction risk.

3.3.1 Antithesis: Tax Collection or Extortion? -- The Normalcy Test (2026-06-09)

The normalcy test (epistemology): to judge whether a structure is normal, do not look at how much money it makes today -- measure how far it deviates from the historical baseline. The 50-year normal state of semiconductors: the bulk of value-chain profit goes downstream to platform/device makers (who own the customer and the cash flow), while upstream component makers earn cyclical commodity margins -- DRAM historical gross margin 5-15%, foundry 30-40%. Today: TSM gross margin 55-60%, SK Hynix HBM margins 50%+, NVDA 75% -- upstream extracting super-normal rent from downstream. This is the first systemic inversion of the food chain in 50 years -- that is, the abnormal.

The mechanism of the inverted food chain: in a normal food chain, whoever owns the end customer and the cash flow owns the pricing power -- the four landlords (MSFT/GOOG/AMZN/META) spend $700B/yr of CAPEX, the only real cash source in the entire industry, and by normal logic they should be squeezing prices. Yet pricing power sits with the supply side (Taiwan foundry + Korea HBM + NVDA design). The inversion holds under exactly one condition: buyer panic (FOMO). Only a buyer who dares not refuse will accept arbitrary price increases. An extortion economy derives its pricing power from the buyer's fear, not from the seller's irreplaceability -- this is the essential difference from genuine tax collection (Section 3.3).

The hidden premise of the chokepoint thesis: "buy TSM / SK Hynix / NVDA because they are chokepoints" implicitly assumes the extortion can continue indefinitely. But an extortion economy has a fatal flaw: 100% of its revenue comes from the victims' budgets, and the victims hold $400B+ in cash + in-house silicon channels (TPU / Trainium / MTIA) + collective monopsony power. The landlords tolerate the extortion only because inside the FOMO window "not buying means losing". Once FOMO recedes, the four do not exit one by one at the margin -- they awaken simultaneously = a phase transition: every chokepoint loses its extortion target at the same moment, with no sequential buffer.

Triggers for FUCK OFF: the answer to "why not" is not "never" -- it is "not yet". Any one of three conditions opens the door:

• AI segment realized ROIC < cost of capital is forced into the open (2027 disclosure season, see 9.4 loop 4)
• In-house ASIC perf/$ reaches 70%+ of NVDA (Google TPU already self-sufficient, AMZN Trainium 2 ramping -- the exit channel is under construction)
• Any one landlord cuts CAPEX guidance first, breaking the "not buying means losing" prisoner's dilemma (one says FUCK OFF, the other three follow immediately)

Investment implication (the antithesis of 3.3): the high margins of chokepoint stocks are not a tax collected on a moat -- they are ransom collected inside a FOMO window. The 40% margin floor argument in 3.3 holds only while "the taxed remain willing to be taxed" -- when the window closes, ransom collapses faster than a commodity cycle, because it was never supported by cost structure or irreplaceability in the first place. The normal reverts; the abnormal cannot sustain -- this is the epistemological foundation the normalcy test gives the RESET thesis.

3.3.2 The Victims' Counterattack -- Sequential Game + 4-26 Week Timeline (2026-06-09)

Game structure: not a one-shot prisoner's dilemma, but a sequential game. Statically, the four landlords are trapped in a "whoever stops buying first loses the AI race" prisoner's dilemma, and the extortion looks indefinitely renewable. But the real game is dynamic: each player runs two tracks simultaneously -- on stage, keep placing orders (do not break the narrative, avoid being labeled by the market as "exiting the AI race"); off stage, build the exit channel at full speed (in-house silicon + open-source models + multi-sourcing). The real contest in an extortion economy is not in announcements -- it is in the accumulation rate of BATNA (best alternative to a negotiated agreement). Fight while negotiating; the chips on the table change every quarter.

The counterattack toolbox (escalation ladder, ascending intensity):

Time asymmetry -- why the endgame is certain and only the timing is open: the extorter's margin is a window function -- once FOMO recedes it goes to zero, irreversibly (a broken narrative cannot be rebuilt). The victims' BATNA is a monotonically increasing function -- in-house ASIC perf/$ only rises quarter over quarter, and every tape-out permanently raises the negotiation floor. The two curves must intersect. Even without any exogenous shock, the endgame (extortion premium going to zero) is structurally determined; the only open variable is when the curves cross. The sustainability of an extortion economy = a race between FOMO decay and BATNA compounding -- and the latter compounds.

First-mover profile (who says FUCK OFF first): the two with the highest defection payoff + lowest narrative cost -- META (no cloud resale business; GPUs are a pure cost center, the extortion cost cannot be passed on to cloud customers) and AMZN (deepest in-house silicon via Trainium + a retail cash-flow discipline tradition). GOOG has effectively half-exited already (TPU self-sufficiency; NVDA purchases are gestural) but never needs to announce it -- it is the invisible first mover. MSFT is bound deepest to OpenAI and moves last. This supplies the game-theoretic basis for Section 4 Scenario B's "META or AMZN first".

The counter-extortion path -- the landlords' optimum is not public exit, it is inverting the extortion: L3 public defection carries an overlooked cost: the four landlords' own valuations also carry an AI premium, and publicly cutting CAPEX = personally admitting the AI thesis is cooling = their own stocks take the hit too. So the dominant strategy is not to publicly slow CAPEX -- it is counter-extortion: keep total CAPEX guidance unchanged (narrative intact, own stock protected, and "who gets the giant budget" becomes a negotiating weapon), while the money silently changes destination:

• Mix redirection: the same $700B, but more flows to in-house ASICs (margin internalized) + power/shells/networking (non-extorter vendors) + AMD (price lever) -- NVDA's wallet share shrinks without the landlords changing a single guidance number
• Margin suppression: negotiate net-price discounts off BATNA while leaving list price untouched (face preserved for the extorter), with realized ASP drifting down
• Interface commoditization: UALink vs NVLink + open inference stacks -- announce nothing, standardize away the lock-in layer, and the technical foundation of the extortion is quietly dismantled

Then how does the bubble pop? Through the supplier's financials, not the landlords' announcements. Historical precedent = Apple M1 vs Intel (2020): Apple never announced "cutting compute spend" -- it kept spending, just on itself. What Intel lost was wallet share, and the stock structurally derated without a single customer announcement. The puncture sequence on the counter-extortion path:

1. NVDA's financials show it first (late August / late November): revenue growth downshifts + gross margin guidance slips from 75% + backlog quality (large-customer mix / payment terms) deteriorates -- the landlords say nothing, the market reprices the chokepoint premium on its own
2. The memory chain cracks first (MU/SNDK): no narrative cushion, HBM pricing power decays in lockstep with NVDA order quality, and melt-up positioning is the most crowded
3. Second-tier credit events (CRWV / neo-clouds / ORCL OCI): the financing chain is built on the chokepoint narrative; narrative decay closes the funding window

The silent path and the loud path (public guidance cut, Sections 4/5) are not mutually exclusive -- they are two puncture forms of the same endgame. Counter-extortion maximizes the landlords' interest, so the silent path is the more probable one -- which moves the primary RESET watch-point from landlord guidance to NVDA's late-August earnings (margin + mix). On the silent path the landlords' stocks are relatively protected (narrative kept + margin recaptured); the damage concentrates entirely on the extorters and their financing chain.

The 4-26 week game timeline (counting from 2026-06-09)

Relation to 5.1: 5.1 forecasts the market-price layer (the effect); this section reasons through the game-behavior layer (the cause). 5.1's "most likely trigger in the October-November earnings season" has two behavioral counterparts: loud path = the Week 16-22 public defection window; silent path (counter-extortion, more probable) = NVDA's late-August earnings exposing mix/margin in Week 8-16, with the market completing the reprice by Week 16-22. The rollover scenario (Week 22-26, nobody moves and NVDA's numbers stay clean) does not falsify the thesis -- the covert counterattack is monotonically increasing, time is on the victims' side, and each quarter of rollover revises the next window's trigger probability up.

3.4 Historical comparables (overshoot multiples)

AI overshoot multiple (~10x) is still 3x larger than 1999 Telecom (3x, T2 confirmed). Even using Telecom's -90% drawdown as a conservative anchor, historical precedent lies on one side of magnitude. Not claiming a guaranteed -90% crash, but the mean-reversion pull on magnitude is strong and historical precedents are lacking. 1880s / 1920s numbers are analytical anchors; use cautiously for precise prediction.

4. END GAME -- Three Scenarios + Triggers

5. Timeline Prediction (base case = Scenario B 50% branch)

5.1 Near-term Forecast: Next 4 Weeks + 6 Months (as of 2026-06-09)

Next 4 weeks (through 2026-07-07)

4-week action: no new longs. Hold the OTM put base position. If MU spikes +10%+ post-earnings, that is a better put entry (IV rises but strikes can be pushed further out -- net convexity gets cheaper).

Next 6 months (through 2026-12-09)

6. Positioning (action mapping)

7. Core Takeaways

8. Memory Chain END GAME -- MU + SNDK (price anchor 2026-05-12)

8.1 MU at $800 -- a 1999 Q4 magnitude melt-up top

MU END GAME timeline (anchor $800)

Core judgment: MU $798.49 is a cyclical commodity priced as a secular grower -- a valuation absurdity. Any long entry has expected return of -50 to -88%. Not a long candidate; it is a short / OTM put candidate.

8.2 SNDK at $1500 -- an even more extreme pure-play melt-up than MU

SNDK END GAME -- three-scenario recompute (anchor $1,537.97)

Core judgment: SNDK $1,537.97 is a more extreme melt-up than MU. M&A optionality already priced; downside has no cushion. All four scenario expected returns are negative -- this is a short candidate, not a long candidate. Asymmetric optionality has been exhausted; current price has consumed all upside.

8.3 MU vs SNDK positioning recommendation

9. Forced Subsidy Thesis + Crisis RESET Necessity

9.1 The six channels of forced subsidy

9.2 This is not diffusion; it is forced subsidy

Historical contrast:

• Electricity diffusion 1900-1940: kWh cost steadily fell (from ~$0.50-equiv to ~$0.05-equiv), penetration 5% -> 95%. Cost decline drove diffusion.
• Internet diffusion 1995-2010: dial-up $20/mo -> broadband $50/mo with 10-100x speed; unit bandwidth cost dropped 99%. Cost decline drove diffusion.
• Mobile internet 2007-2020: 3G -> 4G -> 5G, unit data cost dropped 95%; smartphone penetration 5% -> 80%. Cost decline drove diffusion.
• AI 2023-2026: API per-token cost did drop 85-92%, BUT upper-layer toll fees (subscriptions / cloud / power / subsidy) rose 30-50% over the same period. Surface diffusion, actually hyperscaler skim + society pays. Cost increase disguised as diffusion.

Key distinction: token cost decline does not equal user-facing AI cost decline. Hyperscalers capture the token price decline (open-source competition pressure) and skim it back via wrapper subscription / cloud markup / enterprise lock-in. Net effect = users pay more, hyperscalers extract more, true user surplus does not increase.

9.3 Historical comparable -- strong subsidy + strong taxation path must crash

9.4 Necessity of Crisis RESET (mechanism layer)

RESET is not a wish; it is mechanism. The current path has 4 self-defeating loops; any one triggering initiates RESET:

1. Token deflation vs subscription inflation reverse fissure: users will eventually realize self-hosting DeepSeek/Llama 4 + an H100 fleet is cheaper than ChatGPT Pro $200/mo. Enterprise IT departments begin in-sourcing 2026-2027. Subscription ARR growth curve breaks.
2. Grid physical ceiling: VA/TX grids reach limits in 2026-2027; new datacenter permitting stalls. CAPEX wants to spend but cannot. Forced CAPEX cut.
3. Fiscal pressure + CHIPS Act political backlash: 2026-2028 US fiscal deficit exceeds 7% of GDP; CHIPS Act renewal blocked. MU/Intel/TSMC AZ lose subsidies; fab utilization insufficient.
4. Shareholder ROI reality check: 2027 hyperscalers first disclose actual AI segment ROIC (vs internal transfer-pricing optimism); if <8% (below cost of capital), buybacks fully restored + AI CAPEX cut 30-50%.

2 of 4 loops triggering initiates cascade. Probability 65-75% that at least 2 fire between 2026 H2 and 2027 H2.

9.5 True diffusion only post-RESET (healthy path)

9.6 Key differences vs prior cycles

RESET will not be slower; it will be faster, because:

• Open-source threat is historically unprecedented: railroad/electricity/telecom had no "free substitutes". AI has DeepSeek/Llama 4 hitting in the face 24/7. Deflation passthrough 3-5x faster than dot-com.
• Hyperscaler balance sheets stronger: MSFT/GOOG/AMZN/META have $400B+ cash cushion; they will not go bankrupt like WorldCom. RESET form = write-downs + buyback cuts + multiple compression, not bankruptcy.
• But second-tier players more fragile: CoreWeave / Oracle OCI / xAI / AI startup ecosystem leverage worse than 1999 telecom. These will actually fail.
• Political backlash arrives earlier: residential power bills + ChatGPT subscription hikes + unemployment anxiety; populist backlash starts in the 2026 midterm cycle.

10. Integrated Conclusion