Mental Models for the AI Semi & Server Space — Fable Extraction

Extracted recursively from Nomura's "Asia AI Semi & Server — Is the cycle over?" (Anchor Report, 30 June 2026). Three passes: Pass 1 inventories the heuristics (breadth — 50 models in 7 clusters). Pass 2 finds the machinery underneath — the causal engines that generate the Pass-1 models, their second-order effects, and what breaks them. Pass 3 compresses to the five kernels that regenerate everything, plus the working playbook.

Standing caveat: the logic below is meant to be durable; every number attached to it (who holds which bottleneck, which share, which multiple) is a 30-Jun-2026 calibration and will rot. Keep the model, re-anchor the number.

PASS 1 — The inventory

A · Scarcity & allocation — who eats

• A1 · Allocation is destiny. A GPU/ASIC maker's realizable revenue is capped by its secured share of the binding input — in 2026, TSMC CoWoS — not by its TAM, design merit, or order book. Price the allocation, not the ambition. Grounding: nVidia locked ~60% of 2026F CoWoS by booking ahead; AMD's GPU and AWS's ASIC both "fell short of beginning-of-year expectations" despite genuine product progress — they simply didn't hold the slots.

• A2 · The destiny variable migrates. What gates shipment changes over time; the allocation lens must be re-pointed at whatever currently binds. Grounding: Nomura's own 2027F view — "CoW capacity allocation will matter less than whether vendors can secure substrates and other small components (CCL, capacitors)." The 2026 destiny variable (CoWoS slots) is already being demoted.

• A3 · Book early, book broad. Dominant buyers reserve strategic resources beyond direct need — including adjacent materials — partly to deny rivals. Booking is a competitive weapon, not just procurement. Grounding: nVidia booked CoWoS "plus key materials like T-glass" ahead of peers; the strategy "has worked out."

• A4 · Booked capacity gets consumed, never released. When the flagship product slips, the vendor back-fills its reserved wafers with older-generation product rather than cede the slots. Wafer totals are stickier than product mix — read allocation as commitment and mix as the adjustable. Grounding: with Rubin cut from 2.5mn to ~2mn units, nVidia raises HGX/Blackwell mix (18%→30% HGX) explicitly "to consume promised CoWoS allocation."

• A5 · Elephants and grass. When two dominant buyers compete for a shared scarce input, everyone else's ceiling is the duopoly's leftovers, regardless of execution. Grounding: "when the elephants (nVidia and Google) fight, the grass (other xPU/ASIC vendors) gets trampled"; the two ≈80% of TSMC's AI revenue, up from 70-75%.

• A6 · Scarcity is how moats get breached. Credible substitution of an incumbent comes from capacity/spec unavailability, not from price. When the leader can't cover the spec, the customer qualifies the challenger. Grounding: Google chose Intel EMIB-T for TPU v9 because the >9x-reticle footprint was something "TSMC's CoWoS roadmap could not address at the time Google made its decision."

• A7 · The qualification clock. Sourcing decisions lock 2-3 years ahead, against the incumbent's roadmap as of decision date. A later roadmap fix doesn't unwind the lost socket — it only contests the next one. Grounding: Google decided at end-2025 against TSMC's then-9.5x roadmap; TSMC's April-2026 jump to 14x reticle defends future sockets, not TPU v9.

• A8 · Second-source split ratios are the tell. Customers deliberately dual-track suppliers to de-risk and gain leverage; the drift in the split ratio front-runs share shifts better than any announcement. Grounding: Vera CPU dual-packaged at TSMC (CoWoS-R, ~40%) and Amkor (S-SWIFT); Google TPU split Broadcom:MediaTek ≈ 66:34 with MediaTek's TPU share "more than doubling to 30%+" in 2027F — the ratio move IS the MediaTek thesis.

B · Bottleneck physics — where the constraint lives and moves

• B1 · The constraint migrates to the step the giant doesn't control. When the dominant supplier aggressively expands what it owns, the binding constraint relocates to the nearest step it doesn't own. Grounding: TSMC's aggressive CoW build shifts the bottleneck to WoS/substrates/small components — "TSMC cannot be the only one 'working hard' to meet demand."

• B2 · Capacity is a plan; output is a min(). Realizable output = the minimum across every co-input; always haircut nameplate. Grounding: TSMC targets 2,000kpcs CoWoS capacity 2027F; Nomura models 1,800kpcs output — a ~10% haircut attributed to WoS and component limits.

• B3 · Planning conviction scales with balance sheet — so the mismatch lodges at the periphery. Big, close-to-customer suppliers plan nearest to real demand; small component makers believe latest and build least. The shortage therefore concentrates at the smallest nodes, systematically. Grounding: "many suppliers in 2H25 underestimated AI order upside more severely than TSMC did"; the resulting shortage list is capacitors, PMIC, CCL, substrates — the periphery, not the center.

• B4 · Greenfield arithmetic dates the relief. New capacity ≈ 2 years from build-start (stretching to 2+ as tool lead times lengthen); brownfield buys the first year, then the wall. Build-start date + 2 = relief date. Grounding: the late-2025 build-start ⇒ "supply will stay constrained into 2027F"; TSMC's next front-end jump "not until 2028F"; equipment lead times "prolonging this to two years or more."

• B5 · The constraint stack differs by altitude. Power gates datacenters; wafers gate chips; small components gate racks; they bind and relieve on different clocks — map each layer separately. Grounding: AWS "constrained mainly by power, then chips and server components"; simultaneously the chip layer is CoWoS-bound and the rack layer is memory/CPU/PCB-bound.

• B6 · Shortage exports itself. Suppliers convert consumer/auto/industrial capacity to AI; the AI shortage spills into non-AI end-markets with a lag — a tradeable second-order effect. Grounding: conversion trend "will accelerate in 2H26-27F and squeeze consumer, automotive and industrial products."

• B7 · The leader's outsourced step is the follower's franchise. Whatever the dominant player structurally refuses to own becomes a granted monopoly-adjacent business for its subcontractors. Grounding: TSMC fully outsources WoS → ASE's oS business = 59%/51% of LEAP revenue in 2026/27F.

• B8 · New entrants climb a yield ladder. Entrants to advanced packaging start where failure is cheap — CPUs carrying no HBM — and graduate to HBM-laden accelerators only after yields mature. Cost-of-scrap ordering predicts entrant product mix. Grounding: "the HPC chips ramping on OSATs' CoW processes from 2H26 are mostly CPUs (which don't carry expensive HBM content)" — AMD Venice at ASE, nVidia Vera at Amkor, Cobalt 200 at Amkor.

C · Demand epistemology — real vs booked

• C1 · Triangulate bookings, usage, funding. Demand is real only where all three agree: order books (bookings), tokens/traffic (usage), and FCF/financing (funding). Each leg alone is gameable; the intersection isn't. Grounding: the report's cycle call rests on exactly this triangle — record bookings, token growth as verification, FCF strain as the named 2027 risk.

• C2 · Tokens are the ground truth. Token-consumption growth is the demand primitive that licenses overriding conventional cycle-top signals. Grounding: Google processing 1.3 quadrillion tokens/month (20x+ y-y), 16bn/minute via Gemini app; explicitly among "demand-side factors sustaining."

• C3 · Watch the app to front-run the silicon. End-application share shifts lead custom-silicon orders by quarters. Grounding: Gemini traffic share ~8%→20-27% underwrites the TPU/MediaTek call; Claude's rising share tracked as the tell for Anthropic's $400bn+ infra commitments across AWS/Google/nVidia/SpaceX.

• C4 · Build the tracker nobody else has. The research edge is an upstream measurement, not a better opinion on downstream data everyone shares. Grounding: the DC-build tracker (280 projects) is "a leading indicator ahead of Asia supply chain data points"; the earliest 2025 signal was BMC bookings — visible only because someone was tracking BMC.

• C5 · One demand currency, footed. Convert every signal to a common unit — GW → chips (TDP, ~70% compute load) → wafers (÷16 GB300, ÷9 VR) — so demand foots against capacity and against every company model. If it won't convert, you don't understand it yet. Grounding: Fig. 3's back-of-envelope is the report's spine; the same wafer number recurs in every chapter.

• C6 · Tenant conversion confirms; halts are the canary series. A shell campus naming a hyperscaler tenant converts speculative capacity into real demand; project halts are the inverse signal — count both as time series. Grounding: Cipher reaching 700MW contracted; Meta-Reliance 168MW; vs Crusoe's Project Jade cessation and the halted MSFT/G42 Kenya project, both removed from the demand base.

• C7 · Neoclouds are transition shock-absorbers. CSPs bridge platform gaps through neoclouds rather than expand outgoing-generation capacity; neocloud order surges mark transitions — and smooth over the air pocket that would otherwise be visible. Grounding: "CSPs may be unwilling to expand current-generation capacity but still need tokens to bridge the gap"; neoclouds "playing a bigger role during the [GB300→VR200] transition."

• C8 · Circular financing pulls demand forward and concentrates risk. Vendors investing in customers, backstopping their loans, or renting them compute all accelerate the cycle and create correlated fault lines. Read every such deal in both directions. Grounding: nVidia's $2bn stakes in CoreWeave and Nebius plus equity rights in IREN; Broadcom leveraging its balance sheet for residual-value guarantees on the $35bn AI XPV platform; SpaceX renting Colossus to Anthropic ($1.25bn/mo) and Google ($920mn/mo).

• C9 · Deflate capex by component inflation. In an inflationary component regime, nominal capex growth overstates real compute growth — part of the "demand" is just pricing. Grounding: ~$25bn of Microsoft's ~$190bn FY26 capex is "higher component pricing"; Meta raised guidance partly "on higher component pricing."

• C10 · Backlog quality is counterparty quality. RPO/backlog is demand visibility only to the extent the counterparties can fund it; concentrated RPO is leveraged demand. Grounding: Oracle RPO $638bn (+363% y-y), heavily OpenAI; AWS backlog $244bn; the report separately flags whether LLM makers' commitments are financeable (hence the Broadcom backstop mattering).

• C11 · Overbooking is Schrödinger's strength. Book-to-bill >2x is simultaneously real demand and stockpiling; it resolves only at the next platform transition or supply relief. Model the unwind explicitly rather than picking a side. Grounding: ASPEED's >2x book-to-bill flagged as carrying "underlying risk of a subsequent correction… stockpiling amid a tight-supply, cost-inflationary environment," with a modeled 2028F pullback (37mn→35mn units).

D · Technology forcing functions — what the physics mints

• D1 · Reticle size is the master spec. Package-footprint growth is the upstream variable that forces packaging technology choice, substrate area, thermal solution, and the whole content chain. Track the reticle roadmap as the leading indicator of everything downstream. Grounding: 3.3x (Blackwell) → 5.5x (Rubin) → ~6x (Feynman interposer) → 14x CoWoS roadmap; TPU v9's >9x footprint literally decided the Intel-vs-TSMC packaging question.

• D2 · Bigger, hotter chips mint new BOM lines. Each step in die size and TDP creates content that didn't exist: SiC carriers, microchannel lids, full liquid cooling, 18L→44L→(78-104L?) boards, M7→M9Q/PTFE materials. Content-per-unit is a growth axis orthogonal to unit growth. Grounding: TDP 700W→1,400W→2,300W across generations; Feynman's footprint/TDP drives SiC thermal plates (~6% of GWC 2028F revenue, a product that doesn't exist today).

• D3 · Synchronized migration spikes the node. When multiple large-die programs land on the same process node in the same year, that node's demand goes non-linear. Watch for cadence convergence. Grounding: Rubin + TPU v8t/v8i + Trainium 3 all reaching 3nm in 2026F — "a somewhat synchronized large-die AI chip migration cadence."

• D4 · Paradigm shifts cluster at leap years. Breakthroughs arrive in bundles (~every 3 years), so execution risk concentrates: 2028F needs 336/448G SerDes + CPO + EMIB-T + GPU-on-GPU SoIC + CoPoS + new CCL simultaneously. Date the leap; that's where both the risk and the new-content opportunity sit. Grounding: the 2H24-25 Oberon leap; "2028F will mark the next technological leap," with the full breakthrough list attached.

• D5 · The transitional-tech trap. Bridge technologies (NPO, XPO, CPC) offer near-term profit with a 3-5-year obsolescence tail if the end-state (CPO) leapfrogs them — and the uncertainty itself stalls capital commitment across the chain. Prefer end-state exposure when identifiable. Grounding: "higher medium-term risk to investments in transitional technologies… given CPO could ultimately be the end-state solution."

• D6 · Continuity de-risks; demos outrank slides. Package/architecture continuity across generations predicts smooth ramps; physical hardware shown beats roadmap claims as evidence of design decisions. Grounding: Rubin Ultra's shift to a Rubin-like 2-die package (vs prior 4-die MCM expectation) was "validated by nVidia's Kyber compute blade demo accommodating four GPUs"; the transition expected "smooth… akin to GB200-to-GB300."

• D7 · Agentic AI re-rates the CPU. Agent orchestration is CPU-bound: CPU:xPU ratios move from 1:8 toward ~1:1, standalone CPU racks appear, core counts lift ASPs without unit growth, and management/BMC content rides along. A commodity category became a growth category in about two quarters. Grounding: "Agents don't rent cores"; nVidia's $200bn Vera TAM claim and ~$20bn CPU revenue visibility; AMD's server-CPU TAM raise $60bn→$120bn+; Arm's ">4x today's CPU capacity"; ASPEED's TAM re-segmentation.

• D8 · Testing intensity peaks at immaturity. Test time scales with chip complexity × process immaturity: it doubles at a new generation's ramp, then gets streamlined as yields mature and programs are trimmed. Testers monetize the ramp, not the steady state. Grounding: next-gen XPU "testing time still needs to double versus Blackwell before being trimmed down once mass production stabilizes" — the finding that de-risked KYEC's 2026F.

E · Money mechanics — who captures, who pays

• E1 · The bottleneck owner books the price hike. Pricing power sits at the currently-binding step and moves with it. Find the constraint owner; that's where the earnings surprise is. Grounding: CCL/substrate price actions; TSMC's 5-10% N2/N3/N5 hikes "with potential upside in scale and scope — even mature nodes."

• E2 · Same inflation, opposite signs. One input-cost line (memory above all) is simultaneously the buyer's FCF squeeze and the seller's ASP windfall. Your sign depends on chain position — always ask who pays and who collects. Grounding: surging memory costs drive hyperscaler 2027F FCF pressure and general-server ASP +28% / revenue +67% in the same forecast.

• E3 · Price elasticity scales inversely with size. The smallest qualified player at the bottleneck has the most earnings torque per point of price. Grounding: "given TUC's smaller operating scale versus other AI CCL makers, its earnings growth is very elastic to pricing tailwinds" — hence TUC's 197% EPS growth vs EMC's 175% on the same hike cycle.

• E4 · Capex anchors; FCF governs. Capex guidance anchors next year's demand, but capex growing faster than free cash flow is the sustainability limit — a capex line you can't fund is a future air pocket. Grounding: top-5 capex consensus ~$930bn (2027E) against FCF ~$41bn; named as the thing "investors could be concerned" about.

F · Valuation & market behavior

• F1 · Revisions are the price engine. The durable catalyst is sustained upward consensus EPS revisions — the ratchet — not one-off multiple expansion. The ratchet stalling (revision breadth rolling over) is the real cycle top, truer than price hikes/LTAs/overbooking, which have called past tops but misfire when demand is structural. Grounding: "sustainable upward consensus earnings revisions remain the biggest catalyst"; every company chapter is organized around the revision.

• F2 · The multiple is the trade. Targets are built as (high-end-of-historical-band multiple) × (forward EPS/BVPS); the narrative is what licenses the band position, and it is the thesis's most fragile part. Grounding: TSMC 25x, ASE 25x, MediaTek 25x, EMC 32x, TUC 30x, ZDT 28x, ASPEED 50x — each explicitly "at the high end of the historical range."

• F3 · Basis-rolling is the aggressiveness dial. Analysts roll the TP's earnings basis outward (2027F → avg 2027-28F → 2028F) to keep targets rising as the cycle extends; the further the basis, the more the call leans on cycle longevity. Check which year the TP hangs on. Grounding: ASE moved to avg 2027-28F; ASPEED from 50x 2027F to 50x 2028F; ZDT and GWC to 2028F bases — all in one report.

• F4 · The consensus-gap slope is the actual bet. Where the analyst's numbers diverge from consensus increasingly into out-years is where the differentiated claim lives; a flat gap is just noise. Grounding: MediaTek net profit +2.7% vs consensus 2026F → +19.3% 2027F → +35.5% 2028F: the entire ASIC bet sits in the slope. (Contrast ASPEED: above consensus near-term, -20.6% below in 2028F — a dated disagreement about the correction.)

• F5 · Re-rating front-runs revenue. In a shortage, being a qualified exposed player earns the multiple before shipments arrive; band breakouts precede the P&L. Grounding: TUC re-rated on "a worsening CCL industry shortage should support a broad-based sector re-rating"; ZDT/EMC/TUC price bands broke their historical ceilings months before the modeled revenue inflection.

• F6 · The reflexive macro loop. The boom is self-limiting through rates: AI-driven economic acceleration → earnest Fed hikes → 10y yields far above 5% → complex-wide de-rate regardless of estimates. The cycle's success is its own tail risk. Grounding: Nomura's strategist scenario — insurance hikes turning into "earnest, consecutive rate increases"; yields listed as a top risk beside component shortages.

• F7 · Publish grids where points would lie. When the point is unknowable, honest work publishes the sensitivity surface and names a base case; false precision is the tell of weak research. Grounding: 2029F CoWoS "2,500-3,500kpcs depending on price hikes," with a full price × content-share grid (Fig. 22); VR200 as "15-20%" of 2026F shipments.

G · Ecosystem structure

• G1 · Bet the arms dealer. The ASIC design partner monetizes a hyperscaler's silicon ambitions with less binary risk than any chip vendor — and with multi-customer optionality. Grounding: MediaTek (TPU v8t/v9) TP raised 3,400→5,800 on ASIC sales assumptions of $2.5bn/$14bn/$36bn; Broadcom/Marvell/Alchip/GUC populate every hyperscaler's IC-partner row.

• G2 · The stack has layers; climbing them is the margin story. L0 land/power → L1 shell → L2 ODM → L3/L4 compute buyer. Crypto miners pivot to AI landlords; hybrids that climb layers (own GPUs and the shell) capture the mix shift. Grounding: the report's explicit L0-L4 framing; IREN evolving from Bitcoin miner to GPU-owning cloud with a 5-year nVidia service contract; Cipher/Hut 8/Core Scientific mid-pivot.

• G3 · BOM membership beats product merit — and transitions are the entry windows. The tradable unit for suppliers is qualification into a platform's supply chain; positions are sticky within a generation and contestable only at generation changes. Map the BOM (who supplies what to which platform) and watch the windows. Grounding: the Fig. 140/150 supplier matrices are the report's most operational pages; ZDT's thesis is precisely new-window entry ("producing VR200 Bianca HDI boards… ramp Google's switch boards by late 3Q26F").

• G4 · Segment demand by which supply chain it draws on. Sovereign programs (HUMAIN, Stargate-country deals, China's $295bn plan) are a distinct demand class; include only what draws on the supply chain you're modeling. Grounding: China's data centers excluded from the CoWoS math "since they likely adopt domestic compute chips" — an explicit exclusion discipline most Street models skip.

PASS 2 — The engines underneath

The 50 models are not independent; they are outputs of six causal engines. Each engine is stated as mechanism → what it generates → what breaks it. This is the recursion: models made of models.

Engine I — The Allocation Cascade

Mechanism: upstream demand measurement (C4/C5) → buyers race to book the binding input (A1/A3) → the duopoly crowds out marginal buyers (A5) → the giant expands what it owns, so the constraint migrates outward (B1/B3) → periphery owners inherit pricing power (E1/E3) → their estimates ratchet up (F1) → the ratchet licenses high-end multiples (F2/F5). Generates: the report's entire Buy-list ordering — the durable longs are bottleneck owners one or two layers down (ASE, Unimicron, EMC, TUC, ZDT, KYEC), while the marquee chips are demand drivers rather than capture points. Also generates A2 automatically: as the constraint migrates, the "destiny variable" migrates with it. Breaks when: the revision breadth rolls over (F1's own kill switch), or supply relief arrives on schedule (B4) and the pricing power fades where you're long — the cascade doesn't stop, it moves, and the failure mode is holding last year's constraint owner.

Engine II — The Mismatch Generator (why the bottleneck always moves outward)

Mechanism: planning conviction scales with balance sheet and customer proximity (B3) → in every demand shock, the center (TSMC) plans closest to reality and the periphery under-builds → the constraint lodges at the smallest node → price hikes there (E1) fund capacity → greenfield takes 2+ years (B4), during which demand grows again → the constraint hops to the next least-convinced node. Generates: the observed conveyor: 2026 CoW/memory/CPU → 2027 WoS/substrate/CCL/capacitors/PMIC/optics → 2028 SerDes/CPO/thermal (the report's own sequence). Predicts you can forecast the next bottleneck by asking "who in the chain still doesn't believe the demand?" — conviction lag, not technology, locates it. Breaks when: demand actually stops growing (the conveyor needs fuel), or when a coordinating actor (TSMC co-investment, hyperscaler prepayments, Broadcom-style backstops) transmits conviction down the chain faster than organic ordering — watch prepayment/LTA structures reaching small suppliers as the damping signal.

Engine III — The Demand-Truth Triangle

Mechanism: three legs — bookings (orders, RPO, book-to-bill), usage (tokens, app traffic), funding (FCF, financing) — and a set of named wedges that push bookings above real demand: strategic overbooking (A3/C11), stockpiling under inflation (C11), neocloud bridging (C7), circular financing (C8), capex-price inflation (C9). Generates: the report's title answer. "Is the cycle over?" = "do usage and funding still corroborate bookings?" As of June 2026: usage yes (tokens 20x), funding straining (2027F FCF), bookings inflated but backed. Each wedge carries its own unwind trigger: transition air pockets (C7), financier withdrawal (C8), inventory digestion (C11). Breaks when: treated as static. The triangle's legs update at different frequencies (usage monthly, bookings quarterly, funding annually-ish) — the discipline is noticing when a leg turns before its narrative does. The funding leg turns first this cycle (memory costs → FCF), which is why the report names it the top risk.

Engine IV — The Content Ratchet

Mechanism: physics escalates (reticle, TDP — D1/D2) → each escalation forces new packaging/materials/thermal (D4) → new BOM lines mint new suppliers (D2) → entrants climb the yield ladder into them (B8) → customers dual-source for leverage (A8) → share contests open at each platform transition (G3). Generates: revenue = units × content/unit × price, and this cycle uniquely grows on all three axes at once (units: racks 54.5k→62k; content: 22L→44L boards, SiC, liquid cooling; price: across-the-board hikes). Explains why supplier revenue forecasts (EMC +110%, TUC +100%) can exceed any believable unit growth — most of it is content × price. Also generates D7 as a special case: agentic AI is a content ratchet on the CPU side of the rack. Breaks when: physics offers a cheaper path — a packaging/integration breakthrough that reduces content per unit (the report's own CoWoP and panel-level candidates), or when a platform transition simplifies rather than complicates. Content ratchets reverse rarely but violently; watch for "cost-down generation" language in roadmaps.

Engine V — The Reflexivity Pair (the cycle attacks itself twice)

Mechanism, macro circuit: AI capex boom → measured economic acceleration → Fed forced from insurance hikes to earnest tightening → 10y >5% → the multiple (F2) — the most levered part of every thesis — de-rates complex-wide (F6). Mechanism, micro circuit: shortage → overbooking (C11) → inflated order signals feed capacity planning (B3's data input) → over-conviction builds exactly at the top → glut. Meanwhile the smoothers — neocloud bridging (C7), mix back-fill (A4), vendor financing (C8) — hide the transition seams that would otherwise reveal where real demand ends. Generates: the deep reading that the system's stabilizers are also its opacity: everything that makes the cycle resilient (flexible mix, bridge capacity, financed demand) also makes the eventual turn harder to see in order data. Hence the primacy of usage data (C2) — tokens can't be double-ordered. Breaks when: — it doesn't; it's the boundary condition. The practical model: position for the cascade (Engine I) while renting, not owning, the assumption that the reflexive pair stays dormant; the two circuit breakers (yields, revision breadth) are cheap to monitor daily.

Engine VI — The Milestone Ledger (research as dated falsification)

Mechanism: every load-bearing thesis is attached to a dated, observable event; the portfolio of open questions resolves on a calendar, not on argument. Grounding examples: TPU v9 tape-out end-2026E ("a key reality check for EMIB-T timing"); VR200 = 15-20% of shipments concentrated in 4Q26F; Rubin-Ultra transition 2Q27F; substrate constraints capping ASPEED's 3Q26F; Crusoe/G42 halts as demand-reality events. Generates: the difference between a view and a position. The EMIB-T question (A6/A7) isn't argued to conclusion — it's carried as an explicit coin-flip with both branches pre-priced (MediaTek "benefit of the doubt" on one side, TSMC's SoIC/CoPoS defense on the other). Second-order: whoever maintains the best milestone calendar systematically front-runs resolution repricings. Breaks when: milestones are allowed to slip silently. A slipped date is information (usually negative); re-dating without noting the slip is how theses become zombies.

Second-order observations (models about the model-set)

• The same entity appears on multiple sides. Broadcom is arms dealer (G1), financier (C8), and backstop guarantor simultaneously; nVidia is chip vendor, customer-investor, and compute renter. Exposure through one lens is correlated through the others — sum a name's roles before sizing it.
• ASP inflation contaminates every unit read. E2/C9 imply that in this regime, revenue series systematically overstate volume everywhere: capex, server revenue, supplier sales. Deflate first, or you'll double-count the same price move as both demand and earnings power.
• Allocation data and demand data are different things. A4 means wafer allocations can look rock-solid while end demand rotates underneath (mix shifting to last-gen back-fill is the tell). Watch mix, not totals, for early warning.
• The report's method IS a model. Foot everything to one number (C5), publish grids not points under uncertainty (F7), show every revision (F1's discipline), label inferred vs confirmed (Feynman's "?" cells) — these are heuristics for doing research in this space, as portable as any market model.

PASS 3 — The kernels

Compression test: which minimal set of questions regenerates all 50 models? Five kernels, plus the method rule. If you internalize these, Pass 1 is re-derivable on demand — including for constraints, players, and technologies that don't exist yet.

• K1 · Scarcity decides who eats. Under a binding constraint, secured position beats product merit: allocation caps revenue (A1-A5), unavailability breaches moats (A6-A7), split ratios telegraph the shifts (A8), qualification beats quality (G3), and the arms dealer eats regardless (G1). Ask of any player: what has it secured, and whose leftovers cap it?

• K2 · Constraints migrate outward on a conviction gradient — follow them. The bottleneck moves to what the giant doesn't control (B1), lodges where planning conviction is weakest (B3), relieves on greenfield arithmetic (B4), and differs by altitude (B5). The owner of the current constraint has the pricing power (E1/E3); the owner of the next one is the trade. Ask: which step binds now, who still disbelieves the demand, and what's the build-start date?

• K3 · Demand is real only where bookings, usage, and funding agree. Every wedge between booked and real has a name (overbooking, bridging, circular financing, price inflation) and an unwind trigger (C7-C11); usage is the leg that can't be gamed (C2/C3); the funding leg turns first (E4). Ask: which leg is weakest right now, and what event exposes it?

• K4 · Physics mints content; content decouples revenue from units. Reticle and TDP escalation (D1/D2), migration convergence (D3), and paradigm-leap years (D4) create BOM lines from nothing; entrants climb in where scrap is cheap (B8); revenue = units × content × price and the axes move independently (Engine IV). Ask: what does the next spec step force into existence, and who owns it?

• K5 · Prices ride revisions; multiples ride narrative; yields revoke both. The revision ratchet is the engine (F1), the high-end multiple is the trade (F2/F5), the basis-roll and consensus-gap slope reveal the analyst's real position (F3/F4), and the reflexive loops (F6, Engine V) are the standing tail. Ask: what keeps the ratchet turning, what licenses this multiple, and what single print revokes it?

• K0 · The method rule (meta-kernel): convert everything to one unit and make it foot; publish intervals where points would lie; date every claim's resolution; show every revision; label inference. Research that can't be footed, falsified, or diffed is opinion.

The underwriting playbook — pricing a name in this space

1. Position (K1): what has it secured — allocation, qualification, BOM slots? Whose leftovers cap it?
2. Constraint (K2): does it own the current or the next binding step? If neither, it's a demand story — apply K3 harder.
3. Foot (K0): reconcile its implied volumes to the common unit and the capacity ceiling. If its plan requires more than its allocation, someone's number is wrong — probably its.
4. Truth (K3): how much of its backlog is wedge (double-order, bridge, financed)? What's the unwind event?
5. Content (K4): is its content-per-unit rising with the spec roadmap? Content growth survives unit stalls.
6. Torque (E3): smallest qualified player at the bottleneck = max earnings elasticity; size accordingly.
7. Multiple (K5): where in the historical band, what narrative holds it there, which forward year is the basis, and what's the consensus-gap slope telling you about the real bet?
8. Calendar (Engine VI): list the dated events that resolve the thesis; if you can't, you don't have one.

The standing watchlist (dated, as of 30-Jun-2026 — re-anchor quarterly)

• EPS-revision breadth across the complex — the true top signal (K5).
• 10y yield vs 5% / Fed shifting from insurance to earnest hikes — the multiple's revoker (F6).
• CoWoS output-vs-capacity gap — narrowing = constraint migrating; re-point Engine I holdings (K2).
• TPU v9 tape-out (end-2026E) — the EMIB-T coin flip; resolves TSMC-packaging share risk vs MediaTek/Unimicron upside (A6/A7).
• VR200 mix in 4Q26F / Rubin-Ultra transition 2Q27F — transition air pockets that the smoothers are currently hiding (C7, Engine V).
• Memory price trajectory → hyperscaler FCF — the funding leg's turn (E2/E4).
• Halt/tenant-pull count (post Crusoe-Jade, G42-Kenya) — the demand-reality canary series (C6).
• Substrate/CCL spot ASPs and small-supplier LTA/prepayment structures — pricing-power fade and conviction-transmission damping (Engine II).
• Book-to-bill normalization vs inventory (ASPEED-class signals) — the overbooking resolution (C11).

Source: all grounding drawn from Nomura Global Markets Research, "Asia AI Semi & Server — Is the cycle over?", Anchor Report, 30 June 2026 (Jeng, Lee, Teng, Yang, Chen, Hu). The models are this extractor's synthesis; the report supplies the evidence. Logic durable; calibration expires.

04 — An Autonomous Research System for Anchor-Class Sell-Side Work

Designed from first principles against Nomura's "Asia AI Semi & Server — Is the cycle over?" (30 June 2026). The question this document answers: if you wanted software agents to produce research of this class — not a summary of it, the research itself — continuously and autonomously, what would you build? The plan starts from the research problems, derives the epistemic operations they require, and only then names tools, skills, and workflows. Anything in the inventory that cannot be traced back to a problem should be deleted.

1. What the artifact actually is, and what "replicate" means

The report is not 120 pages of content. It is one decision, warranted: stay long Asia AI semis after an 85% run, because the cycle has not peaked — plus the tradeable decomposition of that call into 25 tickers and 9 target-price raises. Everything else is warrant.

Read closely, the warrant rests on exactly three pillars, and they are worth naming because they are the specification:

1. A proprietary leading indicator, footed to supply. The data-center build tracker (280 projects → GW → chips → CoWoS wafers) lets the analysts see demand before it appears in Asia supply-chain data, in units that reconcile against TSMC's capacity. The edge is not a better opinion; it is an earlier, commensurable measurement.
2. Internal consistency at scale. The same 1,800kpcs CoWoS-output number flows into the exec summary, the allocation table, the TSMC chapter, ASE's LEAP bridge, ASPEED's BMC SAM, and the server forecast — and foots everywhere. A reader who checks any two sections against each other finds they agree. That is what makes 120 pages feel like one mind.
3. Vintage discipline. Nearly every table is old-vs-new. The product is not the level; it is the revision — what changed since December, and why. The narrative is literally a diff with reasons attached.

So "replicate" means: reproduce the warrant, not the pages. A system that emits the same figures without provenance, footing, and diffs has replicated nothing; a system that maintains those three pillars over a living state can emit this report — and next year's report about bottlenecks that don't exist yet — as a rendering step.

The honest ledger: where agents lose, where they win

Be precise about the gap between a Claude-driven system and the Nomura desk, because the architecture must be shaped by it.

Agents lose on: - Field checks. "Our latest supply chain survey suggests…" is human relationship capital. Per-node kpcs figures, OSAT booking tone, testing-time-per-chip — not scrapeable, not licensable. This is the single genuine moat in the report. - Licensed data. Bloomberg consensus (BEst), Similarweb, TrendForce pricing. Deterministic but paid. - Being talked to. Companies pre-brief analysts. No feed replaces that.

Agents win on: - Exhaustiveness. The desk samples; an agent fleet can read everything — every earnings call including the 40 no analyst covers, every 8-K exhibit, every Taiwanese-language board resolution, every ERCOT interconnection filing, every Korean HBM trade story. Breadth substitutes for access more often than people expect. - Language. The supply chain discloses in Chinese, Japanese, and Korean first. MOPS filings, DIGITIMES Chinese edition, Nikkei, Korean IR — native multilingual reading is a structural advantage over any Anglophone desk. - Persistence. The system never forgets a vintage, never loses a source URL, never fails to re-check a number it published. - Consistency. Perfect footing is trivial for software and hard for six analysts with spreadsheets. (The report has visible OCR-era artifacts and at least one duplicated figure.) - Adversarial rigor at scale. Every major claim can get a dedicated refuter agent before publication. Humans do this for the thesis, not for 500 numbers. - Accountability. The system can score every dated forecast it ever made against outcomes and publish the scoreboard. Sell-side does not.

The design principle that falls out: substitute exhaustive public-source triangulation for privileged access, and persistent versioned state for analyst memory — and wherever the field-check moat is genuinely load-bearing, degrade to calibrated intervals rather than fake the point estimate.

2. The five research problems

Strip the report to its irreducible questions and there are five. Every section serves one of them.

P1 · Demand reality — Is the demand real, how large, arriving when?

The hardest problem, because the conventional signals (price hikes, LTAs, book-to-bill >2x, overbooking) all read "cycle top," and the analysts must decide whether to override them. The report's answer: construct an independent measurement upstream of consensus (the DC build tracker), triangulate against usage ground truth (tokens processed, Gen-AI traffic share), and separate real demand from booked demand (strategic overbooking, neocloud bridging, vendor financing, SpaceX-style circular deals all inflate bookings). Failure mode if skipped: you are long a double-ordering mirage, or flat for a supercycle.

P2 · Constraint location — Which step of the chain binds, who owns it, when does it relieve, where does it go next?

The report's signature intellectual move: TSMC's CoW is not the bottleneck precisely because TSMC controls it and is expanding aggressively; the constraint migrates to what the giant does not control (WoS, substrates, CCL, capacitors), and its owners inherit the pricing power. Supply relief is datable: greenfield ≈ 2 years from build-start, so late-2025 starts ⇒ constrained through 2027F. Failure mode: you buy the marquee chip vendor when the money is being made two layers down.

P3 · Allocation under scarcity — Given a binding resource, who gets what?

Zero-sum share assignment over the constrained output: nVidia ~55%, Google ~27%, AMD 8-9%, AWS 5-6% of 1.8mn wafers. This is what converts a macro call into per-company revenue. Note its logical structure: it is a constrained estimation problem — shares must sum to the (interval-valued) total, and each customer's share is evidenced by roadmaps, guided capex, rack shipments, and channel noise. Failure mode: per-company forecasts that don't jointly reconcile to physical supply — which is precisely the error the rest of the Street makes.

P4 · Translation to securities — What does the above do to revenue, EPS, and what's priced in?

Mechanical but merciless: segment bridges (LEAP, SiC thermal plates, BMC SAM) → quarterly P&L → EPS → gap vs consensus → target price = chosen multiple × chosen forward basis. The report's craft details matter: the multiple sits at the high end of the historical band (the re-rating IS the trade), the EPS basis rolls forward to keep TPs rising, and the consensus gap widening into out-years marks where the differentiated bet lives (MediaTek: +2.7% vs consensus 2026F → +35.5% 2028F). Failure mode: right thesis, no tradeable expression — or a TP that silently disagrees with its own model.

P5 · Falsification and timing — What would make this wrong, and when do we find out?

The report carries its own kill switches: EPS-revision breadth rolling over (the true top signal), 10-year yields through 5%, TPU-v9 tape-out on EMIB-T (end-2026E — a dated reality check on the biggest structural threat), project halts (Crusoe Jade, MSFT/G42 Kenya as canaries), memory-cost trajectory versus hyperscaler FCF. Failure mode: a thesis with no resolution criteria is an opinion, and it will be defended long after it dies.

P0, cross-cutting: the pillars from §1 — provenance, footing, vintage discipline — are not a sixth problem but the quality bar every answer must clear.

3. From problems to operations

Each problem decomposes into a small set of epistemic operations — the verbs the system must be able to perform. There are nine, and the entire tool inventory in §5 exists to perform them.

Two of these deserve emphasis because they are where an autonomous system can exceed the desk:

O4/O5 together produce a mechanism the report only gestures at: infeasibility as signal. If the sum of evidenced customer demand exceeds the capacity interval, that is not a modeling error to be smoothed — it is a measurement of overbooking, publishable as such. The constraint solver's residual is a research finding. (The report reaches the same conclusion — "biggest-ever component supply mismatch" — by analyst intuition; the system gets it as an arithmetic byproduct, quarterly, for free.)

O7 at scale is a new product tier. The desk red-teams its thesis over lunch. A workflow can spawn a refuter per material claim — 200 adversarial passes per publication — and publish each claim with its surviving counter-argument. No human desk can match that, and it directly addresses the reader's actual question ("what would these analysts say if they were wrong?").

4. The architecture that falls out

Four layers plus an evaluation harness. The organizing unit is the claim, not the section, the company, or the model — because the warrant (§1) is a property of claims.

4.1 The claim ledger — a derivation DAG under version control

Everything the system knows is one of four node types in a single graph, stored as human-readable files (YAML/Parquet + DuckDB views) in a git repository:

• Facts — observed leaves. {value or interval, unit, as-of, sources[≥1 URL], corroboration status, confidence}. Example: "TSMC 1Q26 call: 'doesn't leave any business on the table', CoWoS 2026F target ~130kwpm — transcript §, dated."
• Assumptions — analyst-grade priors, explicitly marked. {value/interval, rationale, owner, review-by date}. Example: compute-share-of-load = 70%; chips-per-CoWoS-wafer: GB300 16, VR 9.
• Derivations — deterministic code (Python) whose inputs are other nodes. Example: gw_to_wafers(gw, tdp, load_share, dies_per_wafer). Rerunnable, testable, diffable.
• Publications — rendered claims that shipped in some artifact, pinned to the exact vintage of their ancestry.

Properties this buys, each mapped to an operation:

• Provenance is structural (O3): every published number traces to leaf URLs by graph walk. "Source: Nomura estimates" becomes an expandable tree.
• Footing is redundant paths (O4): the rule "customer shares × output = per-customer wafers = per-customer revenue ÷ ASP" is three paths through the DAG that must agree. Footing checks are graph assertions, run on every commit like CI.
• Incremental update is dirty-propagation (O6): when a leaf changes (TSMC revises guidance), the graph names exactly which published claims are now stale — which is precisely the "what changed and why" narrative, generated rather than remembered.
• Vintages are git history: every publication is a tag; old-vs-new tables are git diff over state, rendered. The revision discipline the report performs manually becomes the storage layer's default behavior.
• Intervals propagate (O5): derivations do interval arithmetic natively; a point estimate is just a degenerate interval with better provenance.

4.2 Sensors — and the channel-check substitution stack

Sensors are ingestion agents, each owning a source class and a cadence, each emitting facts into the ledger (never derived claims). The interesting design work is not the scraper list — it is the substitution stack for the one input the desk has and agents don't. For each thing a channel check delivers, the public proxy chain:

The MOPS monthly-revenue monitor deserves star billing: Taiwan mandates monthly revenue disclosure for all listed companies. That is a free, 12×/year, ground-truth read on the entire supply chain the report covers — TSMC, ASE, ASPEED, KYEC, EMC, TUC, ZDT, Unimicron, and a hundred smaller names. A monitor that ingests it within hours of each print, normalizes it, and diffs it against the system's own implied trajectories is the single highest-value/lowest-cost sensor in the whole design, and it is the honest backbone of channel-check substitution.

Cadences layer as: continuous (news/deal wire), daily (prices, EDGAR/MOPS filings), monthly (MOPS revenue sweep — a fixed calendar event, ~day 10), quarterly (earnings-season transcript sweep — the big recalibration), event-driven (GTC/Computex/symposia, tape-out rumors, project-halt reports).

4.3 Models — code with assumptions factored out

The report's quantitative spine is surprisingly few distinct models. Each becomes a small, tested Python module whose every tunable input is an assumption node in the ledger (so recalibration is a data change, never a code change):

1. GW-deployment bridge — project ledger → GW by year → chips → wafer demand (the Fig. 3 engine).
2. Capacity/output bridge — nameplate capacity by step → binding-step solve → realizable output (the 2,000 vs 1,800kpcs engine). The solver takes the chain as data (steps, owners, capacity intervals, lead times), so when the constraint migrates to optics in 2028 the chain grows a node — no new model.
3. Allocation solver — constrained estimation of share vectors given the output interval + per-customer evidence; residual infeasibility emitted as an overbooking measurement (§3).
4. Unit×ASP revenue bridge — per customer per generation → the Fig. 35-class master table.
5. Server-market rollup — GPU supply → module mix → racks (with yield/bottleneck discount) → units/revenue by segment.
6. Three-statement company model — driver-based quarterly P&L → BS/CF with identity checks; one parameterized engine, N tickers.
7. Valuation engine — historical multiple bands, TP = multiple × basis, consensus-gap ladder, band-percentile flags. (It should flag the report's own tricks: "TP basis rolled forward from 2027F to 2028F" is a disclosure the system makes automatically.)
8. TAM/content bridges — attach-rate stacks (BMC per rack, SiC per Feynman, CCL layer content per platform) for the per-name kickers.

Eight models, each under ~500 lines, each with golden tests pinned to the June-2026 report's published numbers.

4.4 The hypothesis book — theses as objects under test

Research is distinguished from aggregation by carrying theses, and a thesis is only honest if it can die. The hypothesis book is a directory of standing claims, each a file:

id: H-2026-03            # "WoS, not CoW, is the 2027 binding constraint"
status: active           # active | resolved-true | resolved-false | retired
claim: >
  Wafer-on-substrate and small components bind 2027F AI-chip output below
  TSMC's CoW nameplate; owners of those steps capture the price hikes.
evidence_for: [fact-ids...]        # auto-appended by sensors
evidence_against: [fact-ids...]    # auto-appended — see refuter below
resolution:
  criteria: CoWoS output/capacity gap at 4Q27 print; substrate ASP trajectory
  date: 2028-01-31
falsifiers:                        # each becomes a scheduled watch
  - substrate spot ASP rolls over two consecutive months
  - TSMC reports CoWoS output ≈ capacity for 2 quarters
tradeable_expression: [ASE, Unimicron, TUC, EMC positioning]

Two agents service every hypothesis: a curator (routes new facts to the right evidence list) and a standing refuter whose only job is to find disconfirming evidence — searched for as actively as the confirming kind. Falsifiers compile into the scheduled-watch list (§5). When a resolution date arrives, the hypothesis must resolve or be explicitly re-dated with a stated reason — no silent evergreen theses.

The report's ~10 major calls (cycle-not-over; WoS bottleneck; Google-share-rising; agentic-CPU renaissance; EMIB-T coin-flip; FCF squeeze 2027; shortage spillover; neocloud bridge; re-rating persistence; memory-cost double-edge) seed the book on day one.

4.5 Publication passes — views over the ledger

Rendering is terminal and cheap once the state is right. Three artifact classes:

• Anchor pass (quarterly, or on demand): the full report — thesis, tracker updates, allocation, per-company chapters, appendices. Assembled by a workflow that renders each chapter from the current ledger vintage, with a hard footing gate before assembly (any DAG assertion failure blocks publication and names the offending nodes).
• Flash note (event-driven): when a falsifier fires or a material fact lands (a halt, a tape-out, a guidance change), dirty-propagation computes the blast radius and a short note renders only what moved and why — which is exactly what a client wants within hours of an event.
• Company update (per ticker, on earnings): model rerun, revision table, TP check, hypothesis-book deltas relevant to the name.

Every artifact embeds its git vintage tag; every number is hyperlinked (in the HTML render) to its provenance subtree.

4.6 The evaluation harness — the report as a labeled dataset

This is the step most designs skip and the one that makes the system trustworthy.

• Recoverability audit (backtest). The June-2026 report contains hundreds of dated numeric claims. Run the system against only pre-June-2026 public sources and measure which claims it recovers, at what interval width, and which it cannot. The output is a moat map: the precise subset of the report that genuinely required field checks vs the (I expect large) subset recoverable by exhaustive public triangulation. This tells you where to spend on data licensing and where the desk's mystique was just diligence.
• Forward scoreboard. Every dated forecast the system publishes goes into a resolution ledger and gets scored when its date arrives (interval calibration, directional hit rate, revision-vs-outcome). Published. An autonomous analyst with a public track record is a different product category from sell-side, not a cheaper copy of it.
• Golden regression. The eight models pinned against the report's published tables; any refactor that moves a golden number fails CI.

5. The concrete inventory

Mapped to this harness's actual primitives: tools = deterministic code invoked by agents; skills = codified procedures (slash-invocable prompt programs); workflows = multi-agent orchestrations; scheduled agents = cron-driven autonomy.

Tools (deterministic, testable — the ledger and the models)

Skills (procedures an agent follows; each ends by writing facts/updating the ledger)

Workflows (fan-out orchestrations; where parallel agents earn their cost)

Scheduled agents (the autonomy layer)

6. What stays human, and the degradation contract

The system should be honest about its boundary rather than paper over it:

• Field intelligence: not replicable. The contract: wherever the moat map (§4.6) shows a load-bearing field-check input, the system publishes an interval with stated provenance quality, never a false point. If a human analyst joins the loop, their checks enter as high-confidence facts through the same ledger door — the architecture is indifferent to whether a sensor is a scraper or a person.
• Consensus data: options are (a) license an estimates feed, (b) synthesize a "visible consensus" from public analyst notes and media (coarser but free), (c) publish vs own prior vintage only. Recommend (b) initially, upgrade to (a) when the product justifies it; never fake (a) with (b) unlabeled.
• Judgment calls that are actually risk appetite — the choice to put the multiple at the high end of the band is not analysis, it is a stance. The system computes the band and the narrative support; a human (or an explicit policy file) picks the point. Same for ratings.
• Sign-off: publication passes gate on human approval until the forward scoreboard (§4.6) earns autonomy. The scoreboard is the promotion criterion — that's the point of building it.
• Compliance: an autonomous system that publishes buy calls is a regulated activity in most jurisdictions; treat the artifacts as internal research/decision support unless and until that's engineered properly.

7. Build order — compounding assets first

The correct ordering criterion is not architectural elegance; it is time-in-market of the assets that compound. Two things in this design get more valuable every week they run, and cannot be backfilled later: the deal/project ledger (P1's census) and the MOPS monitor's history (the supply-chain heartbeat). Start them first even while everything else is a stub.

Phase 0 (weeks 1-2) — the spine. ledger tool with git vintaging + footing assertions; seed assumptions and the hypothesis book (10 theses from the report); stand up the daily deal watcher and /mops-sweep. The system is already useful here: it is a living tracker with provenance.

Phase 1 (weeks 3-6) — measurement. /ingest-call + earnings-season-sweep; bridge-gw and bridge-capacity with golden tests against Figs 3/18-22; two-sided deal corroboration; first flash notes fire off real events.

Phase 2 (weeks 7-10) — the calls. solve-allocation (with the overbooking residual), model-company for the nine tickers, value, /red-team. First allocation-refresh produces the system's own Fig. 35-class table with intervals.

Phase 3 (weeks 11-14) — the proof. moat-audit against the June-2026 report: publish the recoverability map internally. First full anchor-report pass, human-reviewed, diffed against Nomura's own next update when it ships — that comparison is the system's first real exam.

Steady state. Quarterly anchors, event flashes, monthly heartbeat, weekly hypothesis hygiene, scoreboard accruing. Headcount equivalent: the desk that wrote this report is ~6 analysts; the goal is not to fire them — it is a system where one analyst-operator plus the fleet covers what six did, with better footing, total provenance, a public track record, and no claim it can't defend.

One-line summary: build a git-versioned claim graph fed by exhaustive multilingual sensors, with eight small models whose assumptions are data, a hypothesis book that is forced to resolve, adversarial refutation at publication, and an evaluation harness that treats the Nomura report itself as the labeled test set — then the 120 pages become a render, and the research becomes the thing that runs every day.