The AI Hardware Unwind

P1 -- The only criterion is AI ROI data

Whether the AI revolution works has exactly one test: can application-layer unit economics close? Current reality: value-chain profits pool in a single hardware layer (NVDA gross margin 71%; the memory cartel lifted conventional DRAM roughly 700% in four years), model labs burn billions a year, and the application layer scales without profits. In an industry whose two downstream layers are broken, hardware-layer profit is not profit -- it is an advance against a future that cannot pay.

P2 -- Open source has capped the paying ceiling; the revenue base is cut at both ends

Kimi/DeepSeek-class open-source, near-free models cap what enterprises will ever pay: the capability gap closes at roughly 1 point per month (80.6 vs 88.7 today, hitting the 85-88 agent threshold in H1 2027), so closed labs have lost pricing power -- there is always a nearly-free anchor. On the consumer side, only about 5% pay (Adam's anchor; consistent in magnitude with public paid-conversion estimates). Anthropic and OpenAI are therefore squeezed from both ends: hardware taxes them upstream while open source caps their revenue downstream -- the middle layer's business model does not close mathematically, and survives on circular financing (which is precisely the root cause of the vendor-financing evidence chain in Section V).

This chart is P2 in real time: token consumption is usage voting with its feet. On this price-routed marketplace, Chinese/open models take roughly seventy percent of the flow. The honest boundary: OpenRouter is only the price-sensitive edge of the market and excludes Anthropic/OpenAI first-party enterprise API traffic -- but substitution always begins at the price-sensitive edge and spreads toward the core; the marginal buyer's revealed preference is the leading adjudicator of pricing power. "Open source caps the paying ceiling" is not a forecast; it is this month's leaderboard.

P3 -- Hardware margins this high are unhealthy; the rent is killing its own demand

A 71% gross margin is a tax on the entire industry. Until the tax falls, the application layer can never reach scale economics and the AI revolution cannot run -- monopoly profit is destroying its own demand side; the structure is self-negating. Hardware margins must normalize (into the 40s, manufacturing territory) and token costs must approach utility pricing before the AI industry can actually develop. That is not a bear's wish; it is a survival requirement of the industry itself.

P4 -- Therefore de-bubbling is self-healing, not disaster

What the market is doing right now -- repricing credit name by name, punishing capex raises, marking GPUs down to 27% of original price -- is the normal mechanism that beats the tax rate back to a sustainable level. Every short section of this handbook is not a bet against AI; it is a bet on AI after the rent is gone. The clearing is the precondition for Wave 3 real applications. The reason the entry window around 2032 is a generational opportunity is precisely that it is the first moment application-layer unit economics can work.

Correction 1: The capex direction was falsified

Phase one is not "guidance cuts" -- it is raises getting punished. TSM raised to $60-64B and was sold 4%; GOOGL raised to $195-205B and was sold 5%. The real cuts arrive in the October FY27 guidance season, scripted as "component pricing normalization."

Correction 2: Buyers are panic-locking supply, not cutting orders

MU signed 16 multi-year agreements worth $22B; Musk publicly wailed "pretty insane" while locking years of capacity. The $950B Korea-US supply package of 7/25-26 (see Section IV) is the largest specimen of the same behavior: peak-panic supply lock-ins are a top signature. The buyers' revolt belongs to the 2027 contract-renegotiation season, not to now.

Correction 3: The paying cliff is a trajectory, not a current state

The China/open-source band closes on the 85-88 agent threshold at roughly 1 point per month (now 80.6 vs 88.7), arriving H1 2027 -- exactly the quarter when Anthropic's ~$40B/yr and OpenAI's committed contract step-ups take effect. The scissors close on a calendar date.

Correction 4: The -80% belongs to 2028, not May 2027

Layer 1 (H2 2026) = datacenter de-bubbling, -30~40%. Layer 2 (2028) = the digital-plus-physical double kill after the relay narrative is falsified: -70~80%, P/E 15-20x terminal. The character of the interval between them was repriced by the Section V cascade vet: downgraded from "long leg" to "tactical bounce leg."

Correction 5 (new in v2.1): The "77% December hike pricing" was a misread

Live FedWatch (07/28): December hold 78.2% / cut 15.4% / hike just 5.4%. v2.0 misread the hold probability as hike probability. But the June dot plot's YE2026 median is 3.8% (vs 3.4% in March) -- the FOMC's own median path implies one more hike. Market 5.4% vs dots implying a hike: the divergence itself is the position (see Section IV).

P0 / now -> 2026-08-15 / short leg

The meat grinder: mega-cap earnings week

The reaction function has been confirmed twice by TSM and GOOGL. July 29-31 is the same grinder running tests #3-#5; 80% odds at least one name gets sold.

• 07/29-31: MSFT / Meta / AMZN / AAPL earnings. Single criterion: does a capex raise get sold or rewarded?
• 08/12: July CPI. Oil plus AI input costs (electricity +4.0% YoY, memory/component pass-through) hit the print. A hot print re-arms December hike pricing from 5.4% (May precedent: it reached 51% during the inflation shock).

Triggers: the "component pricing" number upgrading (April's $5B -> ?); the quality of MSFT's answers on the OpenAI $250B inside RPO; whether Meta announces an equity raise; FY27 capex adjective counting ("significantly higher").

P1 / 2026-08-15 -> 09-30 / short leg

The final review: NVDA's reaction-function verdict

The numbers will be spectacular (GOOGL's $205B capex is NVDA's order book). The only variable is which rulebook the market prices them under.

• 08/26: NVDA earnings. Sold (55-60%) = beat-then-die #4, the last sanctuary falls; rewarded (40-45%) = the tax regime renews, weak hands push F&G toward Greed.
• Mid-September: hawkish FOMC hold; December dots set (June median already 3.8%).
• September: the Anthropic $900B+ round result = the circular-financing life-support verdict (vendors share one incentive: get it closed).
• Late September: front-running research on the SPCX 12/12 lockup expiry begins.

Triggers: first mark-downs printing inside NVDA's own equity book (CRWV/Nebius/xAI); customer-concentration disclosure; whether guidance stops being explosive.

P2 / 2026-10-01 -> 11-03 / short leg

Main harvest: the earnings-revision print layer

Price front-ran in July; October is when it hits the books: full depreciation, the priciest memory contracts entering COGS, and the FY27 guidance season. The precise shape of "ugly Q3" = margins plus guidance plus book losses all darkening at once -- not a revenue miss.

• Mid-October: Q3 earnings; operating margins compress across the board; the pre-scripted "component pricing normalization" cut language debuts.
• Mid-October: first mark-to-market print. GOOGL's $94.1B SpaceX stake -> an estimated $25-30B book loss, whiplashing Q2's $99B gain.
• Early October: NAND Q3 contract price QoQ. Below +30% = memory reversal confirmed = the final confirmation light for the Jun'27 leg.
• 11/03: midterm elections; peak inflation-scapegoat politics.

Target zone: QQQ 560-600 | SPX 6,300-6,700 | VIX 28-40.

P3 / 2026-11-03 -> 12-12 / gear-change window

Side-switch window: the mechanical bounce

Short-covering plus election resolution plus front-running the 2027 pivot. The bounce has no sustainable logic, but it has mechanical logic. Your discipline is to use it, not deny it. v2.1 note: after the Section V vet, every rally here and in P5 gets bear-rally discipline -- rentable, never holdable.

• At VIX 30-40, sell cash-secured puts at the -40% tier on GOOGL/TSM: get paid to wait for Wave 2 entry prices.
• Wave 2 tranche 1 (1/4 of ammunition only): GOOGL > TSM > GEV > AMZN.
• On an SPCX bounce to $130-150, build post-December expiry puts (the lockup trade; first target the $70-80 flow floor).
• Never-buy list: memory / neoclouds / humanoids / TSLA / SPCX longs. (Conflict disclosure: the thesis tracker's sister-long basket contains a TSLA long stub (2026-05-12), in direct conflict with this list. This handbook is the later document (post 07/26) and governs for now; final call belongs to Adam.)

P4 / 2026-12-12 -> 12-31 / short leg

Risk cluster: lockup expiry + FOMC

A potential replay of December 2018: Warsh hiking into a weak tape = the last hawkish performance, and the ritual by which the bubble's death cures inflation and paves the 2027 turn. v2.1 correction: the hike is not the baseline (market prices 5.4%) -- it is the tail case, armed if inflation re-heats through August-November (triggers in the Section VI Fed card).

• 12/12: SPCX lockup expires; $200B+ of potential supply; the front-run already happened in October-November.
• Mid-December: FOMC. Baseline hawkish hold; if CPI runs hot consecutively, a 25bp hike lands on a weak tape (the Dec 2018 structure).

P5 / 2027 Q1 / gear-change window (downgraded from long leg in v2.1)

Ignition: the pivot and the third-year engines

Original judgment: the third-year rally needs no AI repair (Nasdaq +50% on the 2003 rubble, +29% in 2019 on zero earnings growth). v2.1 downgrade rationale (Section V): the debt-cycle variable severs the comparability of 2003 and 2019 -- 2019-type (no debt clearing, pivot = rally) and 2003-type (rally from -78% after the debt cleared) are different species. This cycle's debt structure looks like 2000-02; durable-rally probability cut from 65-70% to 40-50%.

• Fed pivot: wealth-effect destruction + layoffs = demand destruction = inflation cured. But the base rate: the Fed cut 11 times in 2001 and the market fell two more years. A pivot is necessary, not sufficient.
• Capex cuts -> FCF explosion: the Meta 2023 template -- a regime that punishes capex raises symmetrically rewards cuts. This engine stands.
• Reset launchpad: a covering rally from -20~30%. Downgraded to tactical.
• National will: the 2028 election needs a floor; an OpenAI IPO restart = the house's maintenance machine flips back ON.

P6 / 2027 Q2 -> Q3 / adjudication leg

Ride and scout: the scissors close, elasticity decides

The rally (if any) shares the stage with the physical-AI climax. In the same quarter, two known curves intersect: the band touches the 85-88 agent threshold x Anthropic's ~$40B/yr and OpenAI's committed step-ups take effect.

• The version verdict: real demand elasticity after cheap tokens >1 (Version A) -> fill tranche 3, hold to the end.
• Elasticity <1 (Version B) -> the rally is Layer 2's final mania -> longs stop; open the 2028 short list (pure-play humanoids / TSLA / Figure if it IPOs).
• The SNDK Jun'27 leg harvests with the NAND reversal in H1'27 (the -50~70% path).

Consumer-side indicators: OpenAI ad rollout progress (= the subscription surrender note); free-tier capability vs paid-tier gap; NET pay-per-crawl monetization.

P7 / 2028+ / terminal

Layer 2 collapse: the terminal

Physical AI falsified (humanoid mass production slips + robotaxi economics fail) + no next relay narrative. The original report's terminal verdict lands here -- one relay baton late. v2.1 note: if the Benner rhythm wins, the liquidation extends to 2030-2032; the buy year is not the crash day.

• NVDA -70~80% (digital plus physical double kill), P/E to 15-20x, margins normalize to manufacturing 40s.
• Memory margins from 84.6% to teens or losses (trough custom); compute becomes a utility, token cost approaches water and power.
• Hunting ground: residual Jan'28 positions + 2028 LEAPS. Prey: pure-narrative humanoid names.
• Real applications (agents / embodied / vertical AI) close the ROI loop on de-rented soil -- Version A's Wave 3 starts here.

The mismatch of money

Capital pools exactly where it is least needed: 76% operating margins in memory, 71% gross margins in GPUs -- while the application-layer ROI loop, the only thing that decides whether this revolution works, starves (logic spine P1-P3). The margin gradient in Section IV -- 76 / 71 / 60 / 6.4 -- is not four numbers; it is a map of misallocation. Vendor financing is money-mismatch in its extreme form: the seller lends the buyer money to buy the seller's goods, and the money circles in place while value moves not one step. What the market is doing now -- repricing name by name, punishing capex raises -- is chasing money from the wrong places back to the right ones.

The mismatch of time

A five-year, $950B lock signed twelve months before the renegotiation season; three-year depreciation schedules resting on GPUs worth 27% of cost; 2028 LEAPS dying of 2026 theta. Behind every loss sits a pair of unaligned durations. Hence this handbook's single instrument rule: the instrument's duration must match the thesis's phase -- rolling 2-12 months before 2028, LEAPS only from R5 onward. The reason Benner's 150-year-old card earns its place in Section II is that it performs exactly this calibration: it tells you which beat of the rhythm you are standing on, so you do not live a sell year with a buy year's moves.

The mismatch of emotion

Greed at the top makes buyers panic-lock capacity at peak prices (Correction 2); fear at the bottom will keep the same people out of a generational entry (2032). The buried knife catcher is not unlucky -- he is paying the invoice for an emotion out of phase with the market. The KS-4 F&G discipline is an anti-mismatch machine: never add shorts in Fear, scale at Greed 55+/75+ -- it forces you to act precisely when your emotions and the market's phase point opposite ways. Every line of R1-R6 in Section III -- "load at rally tails, never chase weakness" -- is the same discipline made concrete.

The mismatch of trust

Trust in the wrong place is a negative asset: BDC marks so stale that Saba bids at 20-35% off, ratings running a quarter behind CDS, circular financing booking "revenue" against self-guaranteed debt, the HY index still denying what single names have already set records saying. The T1-T5 legend in Section IV is a price list for trust -- how much a court filing or a market price is worth, how much a "person familiar with the matter" is worth, graded cell by cell. The stratification signal (single-name CDS vs the index) is, at bottom, a mismatch between two layers of trust inside the same market -- and its resolution names whose trust was wrong.