Upstream to downstream, from platinum-group by-products to agent identity — 131 scored chokepoints, where supply is rigid and demand is compounding, where the binding constraint goes next, and how this whole thesis dies. Built from fifteen primary-research dossiers. Written to be argued with.
Read Part XIX first. Since the v1 draft this document’s central instrument — EDGE = N − M — has been withdrawn, along with four other claims, on tests written before the evidence was collected. The map stands; the attempts to price it do not. Retraction register →
What this is. A scored map of 131 physical and institutional chokepoints in the AI value chain, built to be argued with. It answers where is supply rigid while demand compounds, and how long does that last. It does not answer what should I own, and Part XVII is a record of five attempts to make that conversion, all of which came back no.
Published by. Seeking Edge, independent research
Draft completed 24 August 2026; this edition 28 August 2026. Not yet published. Layer vintages differ and are itemised in the footer (the thesis status board, the start-here case and the anchor reconciliation are a status layer dated 21 September 2026 and change no score); the scoring data is as of 8 August 2026 and the market comparisons in Part XVII are as of 21–22 August 2026.
How much weight it bears. Of 1,050 distinct source URLs, roughly 21% are primary and ~58% are relay (trade press, channel research, aggregators) — full breakdown and an explicit correction to a harsher earlier estimate in Appendix A. 98 claims carry a verbatim anchor across 35 nodes, of which 57 are machine-verified against the live document and 48 carry a resolved content hash (Part XVIII); 16 of the 25 V=6 nodes carry no anchor at all (Part XVI). 86 of 131 nodes are self-graded ‘high confidence’, which is an author’s prior and not an external referent.
Known-wrong rows. Four nodes carry an ruling open flag — the author believes the score is wrong and the change has not been signed off. The original score is what is shown, everywhere.
Not investment advice. Nothing here is a recommendation to buy or sell any security. Named
companies appear as evidence of where scarcity sits. Figures are point-in-time and several are explicitly
single-source. Reproduction of the scoring is possible from data/ and scoring_v2.py; the
underlying dossiers are the author’s and are not published in full.
Each is a falsifiable claim with the evidence attached and, in Part IX, a pre-committed kill criterion. Where I disagree with consensus I say so explicitly rather than hedging — that is the point of the exercise.
The theses are shown as written. Part XII (15 August) applied a primary-disclosure rule and Part XVII (23–24 August) put ten nodes through primary-source triage, but the 28 August edition never stated what those findings did to the theses. This board, added 21 September 2026, does. Where a supporting leg did not survive, the status is stated here and next to the thesis, and the score is left unchanged until the adjudication is signed off. “Standing” means not contradicted, not verified: batch 1 of about nine triage batches is done, so roughly eight remain.
Twelve tiers. The useful observation is not that the chain is long — everyone knows that — but that severity is not monotonic in position. On tier means it peaks at T2 (75.1) and T5 (73.5) and troughs at T10 (53.2) — design and foundry sit mid-pack because they are toll roads, extremely profitable but not physically binding. On tier peaks the picture is flatter, and the two highest single nodes (94.8) sit at opposite ends of the chain, in T0 and T11. What changes across T8-T11 is not the level but the character, where the constraints stop being physical and become institutional: permits, apprenticeships, qualification lists, balance sheets, insurance capacity, and un-gameable graders.
131 nodes, each scored on six dimensions and carrying its evidence, its relief clock and its tripwire. Thirteen entries are marked DEBUNKED — they exist because a framework that cannot say no is a narrative generator, not an analytical tool.
Run the severity index over all 131 nodes and the top three are all the same thing: heavy rare earths (94.8), rare-earth magnets for robotics and drones (94.8), and the Chinese extraterritorial licensing rule that governs both (92.4). The single most severe chokepoint in the AI supply chain is not a semiconductor, a laser or a turbine. It is a Chinese administrative decision with a days-long fuse, an expiry date of 10 November 2026, and no substitute at any price inside a decade.
And heavy rare earths' Investable Rent Index is 11.4, ranked 100th of 131 — the licensing rule scores 12.9 and the magnets 22.8. That gap between severity and investability is the most important structural feature of this whole map: the hardest constraints in the chain are largely un-ownable, either because they are policy rather than property, or because they sit inside private German and Japanese firms. Anyone whose AI-supply-chain exposure feels comfortably liquid is, almost by construction, not holding the real chokepoints.
| Tier | Node / holders | BIND | DEM / MON | RCI | Binds | Direction | Conf |
|---|
IRI = Severity × (AI demand beta/10) ×
(rent capture/10) — multiplicative, so a severe node with no AI exposure or no pricing power
scores near zero. That is intentional.The two-sided map. Growth alone is not interesting — growth against rigid supply is. And the decay side is where most AI research is weakest, because it is long-biased and because the strongest decay signal is not falling revenue.
Terminal value dies years before revenue. Getty: TTM revenue +4.5%, market cap $183m, share $0.44. Gartner: revenue −0.6%, stock −71%, 7× ARR → 1.8× revenue. Concentrix: revenue +1.9% and a $1,479m loss quarter. TaskUs: revenue +11.4%, market cap −55.8%. Screening for revenue decline misses the trade entirely. Screen for impairments, guidance withdrawal and multiple collapse.
The single most useful thing in this document. The market reliably over-attends the previous bottleneck and under-attends the next one, and the next one is usually derivable from the arithmetic of the tier above rather than from any new information.
Ten technical directions, each with what is actually measured rather than claimed, and what each implies for the supply chain. I have separated high-confidence from "watch it, do not underwrite it" — the difference matters more than the list.
Where it actually stands: Opus 5 pays 2.2x tokens for +3.3pp; Opus 4.7 paid 4.8x for +16.5pp one quarter earlier. The curve is concave AND getting worse as base models improve.
Supply-chain consequence: Bandwidth- and capacity-bound hardware, not FLOP-bound. HBM capacity, KV-cache tiering, prefill/decode disaggregation, prefill-optimised silicon (Rubin CPX: 128GB GDDR7 at >50% lower cost/GB).
Chokepoints pulled: HBM, DRAM, KV-cache storage tier, DPU · Confidence: High
Where it actually stands: RL went from ~3-10% of pretraining compute (early 2025) to ~100% (Grok 4) to ~570% (Cursor Composer 2.5). Max documented 2026H1 run was 3.87e25 FLOP, 85% of it post-training.
Supply-chain consequence: A different hardware profile: inference-shaped, heterogeneous, with large CPU fleets running environments. Also a wholly new mid-scale compute buyer — application companies doing their own RL on open Chinese base checkpoints.
Chokepoints pulled: Un-gameable verifiers, expert trajectories, sandbox infrastructure, CPU capacity · Confidence: High
Where it actually stands: One instrumented coding agent (Claude Code, one week) ran 98.16% KV-cache reads and 0.30% output tokens, and ~80% of its spend was input/context. Claude Fable 5 on CursorBench: $18.02/task, 76 steps.
Supply-chain consequence: Inverts the 'decode is bandwidth-bound, buy HBM' heuristic for the fastest-growing workload class. Creates a priced storage tier inside inference and makes the DPU a chokepoint.
Chokepoints pulled: Flash/NAND for cache, DPU, network storage · Confidence: High
Where it actually stands: Spectrum-XGS, Google's OCS fabrics, Microsoft's 120,000+ fiber miles, coherent ZR/ZR+ pluggables. Anthropic's 1M+ TPU commitment rides Google's MEMS OCS.
Supply-chain consequence: The gating input is civil engineering: fiber draw 18-24 months, duct/trench 18-36 months. Also raises the value of optical circuit switching and long-haul coherent optics.
Chokepoints pulled: Fiber, trench/permits, coherent DSP, OCS · Confidence: High
Where it actually stands: Not shipped at frontier. The strongest evidence remains negative: PostTrainBench SOTA is 0.4%; frontier agents given six days on two unpublished NeurIPS submissions were unambiguously rejected by the authors.
Supply-chain consequence: If it lands, it collapses the train/serve distinction and makes fleet-wide weight updating an infrastructure problem — checkpointing, versioning, rollback, and eval-on-every-update.
Chokepoints pulled: Eval capacity, checkpoint storage, orchestration · Confidence: Low — do not underwrite
Where it actually stands: ~500,000 hours of high-quality real robot interaction data exist globally against 1-10 BILLION hours needed for baseline generalisation — a 2,000-20,000x gap. Figure's System 0 trains entirely in simulation across >200,000 parallel environments.
Supply-chain consequence: Simulation compute becomes a first-class demand category. But note the scale reality: global humanoid shipments in 2025 were 13,317 units — roughly $0.67bn at a $50k ASP, and even Morgan Stanley's doubled 2026 China forecast of ~50,000 units is only ~$2.5bn. BOTH ARE SMALLER THAN AI CODING.
Chokepoints pulled: Force/torque sensors, roller screws, rare-earth magnets, simulation compute · Confidence: Medium
Where it actually stands: Weather is the cleanest win (GraphCast/GenCast lineage now operational). Formal math is real (Lean/AlphaProof). Drug discovery is not delivering: Recursion revenue -60.1% as platform deals fail to renew.
Supply-chain consequence: The recursive case that matters for this atlas is AI designing its own supply chain — materials for magnets, superalloys, dielectrics. AlphaEvolve's best real contribution to date: 1% off a training run, 0.7% of a fleet.
Chokepoints pulled: Self-driving lab hardware, characterisation instruments · Confidence: Medium
Where it actually stands: Photonic/analog/thermodynamic compute remains pre-revenue. The one substrate change actually shipping is NVIDIA licensing Groq's LPU IP (~$20bn for a NON-EXCLUSIVE licence plus the team) into Vera Rubin.
Supply-chain consequence: Watch incumbent absorption, not startup displacement. That is the observed pattern for every orchestration and inference-optimisation company in 2025-26.
Chokepoints pulled: — · Confidence: Low
Falsifier: Dies if any photonic or analog accelerator publishes an MLPerf Inference result within 2x of contemporaneous GPU perf/watt at >100-unit deployment scale. Confirmed if a hyperscaler signs a >$500m non-von-Neumann compute order.
Where it actually stands: RLI at 0.161, growing 20x/yr from 1.6% — the only genuine exponential in the measurement set. But PostTrainBench SOTA is 0.4% and the AI-research-agent evidence is unambiguously negative.
Supply-chain consequence: Watch it; do not underwrite it. If RLI's trajectory holds it is the single largest regime change available, and it would arrive through chip design and kernel optimisation first — the two domains with the strongest oracles.
Chokepoints pulled: EDA, compilers · Confidence: Low but high-consequence
Where it actually stands: Complete protocol infrastructure and ZERO disclosed volume. Six standards, three networks, two hyperscalers, a Linux Foundation. OpenAI scaled back ChatGPT checkout on weak uptake (25 Mar 2026); Mastercard's CEO voiced 'grave concerns'.
Supply-chain consequence: The bottleneck is not rails, it is liability allocation — which is why identity, not payments, is the node to own. The Ninth Circuit's Perplexity v Amazon ruling (4-5 Aug 2026) left the vacuum open.
Chokepoints pulled: Agent identity, insurance capacity · Confidence: Medium on direction, Low on timing
Falsifier: Dies if no card network or top-20 merchant discloses agentic GMV by Q2 2027. Confirmed the quarter Visa or Mastercard breaks out agentic volume in a filing.
Four structural shifts that the supply-chain evidence supports, stated without either boosterism or catastrophism. I have tried to keep each anchored to something measured.
The industry is converging on tokens-per-joule as the operative efficiency metric because in the best markets the scarce input is no longer silicon but firm, permitted, fast power — which, as Thesis 3 argues, is a governance constraint rather than a shortage of electrons. Vera Rubin at ~10× tokens/watt versus GB200 NVL72 is a larger economic event than any single node transition. The consequence is that electricity price becomes a political variable attached to AI: US retail electricity is +47.8% since January 2020, 51 local moratoria are live, New Jersey has legislated a datacenter ratepayer class that curtails before residential customers, and Texas removed 20% of the national pipeline with a governor's letter. The unresolved question is not whether AI can get power. It is who pays for the grid it induces, and that is being decided in state utility commissions rather than in Washington or in any lab.
The honest arithmetic: the most AI-exposed occupations are still growing at +1.1%/yr against +2.0% for the least exposed. Damage concentrates in 22-25 year-olds in automation-exposed roles at −3.8%/yr. That intersection is ~2.8% of employment, implying roughly a 0.16pp/yr aggregate drag — about one month of normal payroll growth. Yale's Budget Lab finds no discernible aggregate disruption at 33 months. Both poles of the public argument are wrong. But the distributional point stands: the cost is falling on entry-level workers in exactly the roles that used to be the on-ramp to a professional career, and the mechanism that replaces that on-ramp does not yet exist. The most striking single datapoint in this entire atlas is that displaced professionals are now a supply input — laid-off scientists and lawyers writing the rubrics that train their replacements, at $125/hr against $300-800/hr day rates.
AI has saturated the domain with a perfect oracle — code, where a compiler tells you if you are wrong — and coding is 55% of all departmental AI spend. Everywhere accountability is irreducible (medicine, law, audit, engineering sign-off, safety cases), the binding input is a grader that cannot be gamed, and reward hacking is already measured at 67.0% prevalence in Frontier Science traces, with environment-trained skill retaining 55% one hop out of the training distribution, 31% at two hops and 20% at three. Layered on top: carriers have won approval to exclude AI from standard liability policies in all 50 states, ISO's standard CGL forms now carry gen-AI exclusion language, and affirmative capacity is trivial. So AI liability has migrated onto corporate balance sheets by default, and the CFO and general counsel become the deployment veto, not any regulator. That, not model capability, is the most likely brake on AI's move into the regulated economy.
Two dates one day apart — 9 November 2026 (the BIS affiliates rule) and 10 November 2026 (China's rare-earth suspension) — were explicitly traded for each other, and either can be reactivated with days of notice and no administrative process. The IEA puts full reimplementation of China's controls at $6.5tn/yr of ex-China downstream output at risk. Meanwhile Commerce holds $150m of equity in a challenger to the EUV chokepoint it simultaneously polices through ASML, and export controls have turned out to bind China's AI programme through memory rather than lithography: SMIC has ample ≤7nm capacity, but CXMT's HBM output supports only 250-300K Ascend 910C. The durable implication is that the physical AI supply chain is now a legitimate coercive instrument for both governments, which means every capacity plan in this atlas carries an unhedgeable political term. The market is currently pricing a rare-earth settlement that Beijing's own statements have twice declined to acknowledge.
N unmeasurable at 3 of 3 nodes tested, which closes
EDGE = N − M as an instrument — see Part XIX. What is left of the framework
is holder attribution, and one executable screen built after the v1 draft. The framework is left
standing rather than quietly rewritten, because the falsification is more useful than the original claim —
but it should be read as a hypothesis that has been substantially refuted, not as a live recommendation. Part XVII has the five cases and the arithmetic.H1 (core). In a capex super-cycle whose top-line growth rate exceeds the expansion elasticity of its physical inputs, excess returns accrue disproportionately to nodes that are (a) concentrated, (b) non-substitutable within the cycle's horizon, (c) capacity-rigid, and (d) able to price. The Investable Rent Index is the operational form of that claim: IRI = CS × (B/10) × (V/10), multiplicative so that a severe chokepoint with no AI exposure or no pricing power scores zero.
H2 (migration). The market systematically over-attends the previous binding constraint and under-attends the next one. The excess return is not in owning the current bottleneck — that is where sell-side attention and crowding already are — but in owning the node that becomes binding one to three quarters ahead, identified by the mechanical arithmetic of the tier above it. CoWoS in 2026 is the canonical over-attended node; substrate and low-loss laminate are the canonical under-attended ones.
H3 (rent evidence hierarchy). Not all scarcity claims are equal. In descending order of information content: customer prepayment or capacity guarantees (Modine's $165m cash upfront against a $4bn 2027-29 guarantee; Intel asking customers to prepay for substrate) > book-to-bill above 2.0x (Powell 3.0x) > operating margin expansion above 40% (Ajinomoto >50%, Lasertec 48.8%, FormFactor 37.3%→50.7%) > price increases > lead-time extension > management commentary > third-party market sizing. Prepayment and capacity guarantees are the only two that cannot be talked into existence.
H4 (the short side). AI deletes transaction-priced, single-deliverable, low-context work; it does not delete relationship-priced, accountability-bearing work. Fiverr −10.0% against Upwork +1.5% in the same quarter is the cleanest natural experiment available, and the rule correctly predicts the stock-media split (Shutterstock −17.0% vs Getty growing), the publisher split (how-to −20%, tech −28%, but sports +37%, travel +22%) and the BPO non-event.
Seventeen signals, chosen for being public, high-frequency and hard to talk into existence. Note how many are customs data, PPI series and monthly Taiwanese revenue filings rather than anything a vendor produces.
| Signal | Source | Frequency | Latency | Why it matters |
|---|---|---|---|---|
| DRAM/NAND contract & spot price | TrendForce, DRAMeXchange | Monthly / weekly spot | 1-5 days | The current binding constraint's price. Watch the LTA/non-LTA split from Q3'26 — the blended figure conceals hyperscalers paying near-zero increments while neoclouds and channel absorb the shock |
| Ajinomoto electronic-materials segment margin | Company quarterly | Quarterly | ~6 weeks | The single cleanest read on substrate scarcity anywhere. >50% OP margin is the proof no substitute is qualified |
| Korean CCL import unit value | Korea Customs (KITA) | Monthly | ~3 weeks | $20,728/t Mar-26 vs $11,880 a year earlier. A public, high-frequency proxy for a private oligopoly |
| Ibiden / Unimicron / AT&S monthly revenue | TWSE monthly filings, company IR | Monthly | ~10 days | Taiwanese monthly revenue disclosure is the highest-frequency substrate telemetry that exists |
| Powell / Eaton / Vertiv book-to-bill and backlog | Quarterly filings | Quarterly | ~5 weeks | Switchgear is the under-rated power node; 3.0x book-to-bill is the evidence |
| BLS PPI: transformers (PCU335311335311), switchgear, HVAC | BLS | Monthly | ~2 weeks | Cleanest public separation of real from narrative scarcity. Transformer PPI flat-lining while switchgear runs is the whole trade |
| Silicon Data GPU rental indices (SDH100RT, SDA100RT) | Bloomberg / Silicon Data | Daily | <1 day | Residual-value truth serum, and the marks that made GPU debt bankable |
| Secured vs unsecured AI credit spread | Bond/loan pricing; deal press | Continuous | <1 day | Nebius at SOFR+250 (6.15% all-in, ten-bank syndicate) vs CoreWeave unsecured ~10%. The best real-time gauge of AI credit stress |
| Hyperscaler OCF / capex ratio | 10-Q cash-flow statements | Quarterly | ~5 weeks | The master valve on every tier above. Alphabet's first-ever negative FCF quarter is the regime marker |
| ASML EUV shipment guidance & Zeiss capex | ASML quarterly, Zeiss AG report | Quarterly / annual | ~4 weeks | The 2029 supply curve is explicitly uncommitted. Watch for commitment to the 2028 +30% step |
| Lumentum / Coherent EML unit commentary; Veeco & AIXTRON AsP orders | Quarterly calls | Quarterly | ~5 weeks | Veeco/AIXTRON order books are an ~18-month leading indicator for optics relief |
| NVIDIA DSX / RVL qualification announcements | NVIDIA newsroom, vendor PR | Event | same day | Capex cannot relieve an administrative gate. New passers are the definitive entry signal for thermal suppliers |
| Colo vacancy, absorption, under-construction MW | CBRE, datacenterHawk, JLL | Quarterly | ~6 weeks | NoVA at 0.3% vacancy with under-construction MW falling YoY is the highest-quality rent in the atlas |
| PJM capacity auction clears & IRAS status | PJM | Annual + filings | same day | First RTO-wide shortfall in history; 1 Jun 2027 curtailment priority is a dated, mechanical event |
| Federal Register: BIS, Section 232, MOFCOM announcements | federalregister.gov API, MOFCOM | Daily | <1 day | The 9-10 Nov 2026 pair is the highest-density policy risk window in the calendar |
| Epoch AI compute & capability index; METR time horizons | epoch.ai, metr.org | Continuous | days | The only independent series for the demand function. Note METR p80 is decoupling from p50 — p80 is the enterprise-relevant number |
| Ramp AI Index; Menlo enterprise spend | Ramp, Menlo | Monthly / annual | weeks | Use with care: median paid-AI firm spends $11.38/employee/month while the top 1% spend $7,450 — a 654x spread that invalidates any TAM built off adoption rates |
Portfolio construction. Four sleeves, deliberately different in character, sized by evidence quality rather than by conviction narrative.
Sleeve A — Confirmed rent, prepayment-evidenced (largest weight). Nodes where a customer has put cash down or guaranteed volume: substrate and ABF, low-loss laminate, applied HVAC/chillers, nearline HDD, MV switchgear. Entry discipline: require at least one item from the top two tiers of the H3 evidence hierarchy. These are already partly discovered; the edge is in duration — the market prices 2-4 quarters of scarcity where the supply arithmetic says 6-10.
Sleeve B — Migration ahead of consensus (highest expected alpha, highest variance). Nodes whose bindingness is arithmetically implied by the tier above but not yet in the price: fiber-to-PIC attach, temporary bonders, burn-in boards and sockets, InP epitaxy, prefab factory throughput, agent identity. Most are private, inside conglomerates, or sub-segments — so implementation is frequently a second-order listed proxy, and position sizing must reflect that the exposure is diluted.
Sleeve C — Fade the over-attended (funding short). CoWoS scarcity, large power transformers as a price story, the 800VDC TAM, High-NA as a near-term constraint, UALink ecosystem, multi-vendor chiplets, standalone vector databases, agentic-commerce revenue. Each of these has a specific falsifier documented in the atlas; each is where sell-side attention currently sits.
Sleeve D — Duration short, not revenue short. Per H4: short the multiple and the contract duration in transaction-priced, single-deliverable, low-context businesses. Explicitly avoid shorting revenue in relationship-priced BPO, seat-based SaaS with accelerating growth, and editorial/rights-managed media — that is where the crowded short has repeatedly been wrong.
Signal construction. For each node, three orthogonal inputs: (i) a price/margin series (PPI, import unit value, segment operating margin); (ii) a quantity/commitment series (book-to-bill, backlog, under-construction MW, monthly revenue); (iii) a relief-clock series (announced capacity with a dated first-production milestone). Long the node when (i) and (ii) are both accelerating and (iii) is more than four quarters out. Begin exiting when (iii) crosses inside two quarters, not when (i) rolls over — price is the last thing to turn.
Sizing. Weight by IRI × source-confidence on the long book only. IRI is unsigned, which
is a real defect: all thirteen DEBUNKED nodes and all five RELIEVING nodes still carry positive IRI (CoWoS 2.5D scores
43.7, top-quartile, while sitting in the fade sleeve). Screen on phase and dir first, then
size on IRI within the surviving long set. Cap any single node at a level that survives the low-confidence register being wrong. Twenty of the atlas's load-bearing claims are single-source; three are flagged for outright rejection.
The honest framing first: this cannot be conventionally backtested, and anyone who claims otherwise is selling. The AI capex cycle has one observation. A chokepoint-rent factor fitted on 2023-26 is fitted on the single most extreme capex acceleration in semiconductor history. What follows is therefore a validation protocol, not a performance claim.
1. Out-of-cycle historical validation (the only real test). Reconstruct the chokepoint framework on three prior capex cycles where the outcome is known and the analyst's memory is not the data source: the 2000-01 telecom optical build, the 2004-08 China commodity super-cycle, and the 2017-18 memory cycle. Score nodes using only information available at the time — vintage trade press, filings, PPI series — and measure whether high-IRI nodes outperformed. Critical: the 2000-01 optical case is the designed failure mode, because it is the cycle where genuine physical chokepoints (EDFAs, DWDM components) delivered catastrophic losses when demand reversed. If the framework does not flag that, it is a momentum factor wearing a supply-chain costume.
2. Point-in-time discipline. Every input must be as-reported with its actual publication lag, not restated. PPI series get revised; TrendForce figures get revised; Supermicro's Q4 FY26 gross margin print used in this atlas is explicitly preliminary and unaudited. Build the panel from vintage snapshots or do not build it.
3. Known biases, stated rather than buried. (a) Survivorship — the named chokepoint holders in this atlas are the firms that survived to hold the position; the ones that were second-sourced out are invisible. (b) Narrative selection — nodes get researched because they are already in the discourse, which correlates with being in the price. My partial mitigation is the thirteen DEBUNKED entries, which exist to test whether the framework can say no. (c) Look-ahead through search — the research is dated 8 Aug 2026 and knows how Q1-Q2 2026 resolved. (d) Reflexivity — publishing a chokepoint list is a mechanism for closing it. (e) Interested-party sourcing — every "we are sold out" statement is made by someone who benefits from you believing it; prepayment is the only self-authenticating form.
4. The falsification test that actually matters. Rather than fitting returns, test the framework's predictive claim directly: does high-IRI predict subsequent fundamental outcomes — gross margin expansion, positive estimate revision, book-to-bill persistence — at a 2-4 quarter horizon, controlling for momentum and for the AI-capex beta? If the alpha vanishes once you control for a naive "AI supply chain" basket, then the atlas is an elaborate factor-loading exercise and should be treated as such. That is the null I would want a colleague to try hardest to confirm.
5. Capacity and liquidity. Several of the highest-IRI nodes are unlisted (Ajinomoto is listed but the exposure is one segment; Zeiss SMT, Trumpf, IMS Nanofabrication, Stäubli, Shinko post-delisting and Surge are not accessible at all). Estimate implementable capacity before conviction, not after. A framework whose best ideas are unbuyable is a research product, not a strategy — and roughly a third of this atlas is exactly that.
I ran the obvious diagnostics on my own scoring and they are not flattering. Publishing them is cheaper than having someone else find them.
None of this invalidates the map — the evidence, relief clocks and tripwires in Part III stand on their sources, not on the arithmetic. It does mean the indices should be used to rank and group, and the actual decisions should be made off the evidence hierarchy in H3 and the migration logic in Part V. Anyone who wants to defend the scoring as a factor should first run the test in step 4 of the backtest plan.
Written as a red team against everything above. If you only read one section, read this one — it is where the framework is most likely to be wrong, and the kill criteria are pre-committed and dated so they cannot be renegotiated later.
m_why has been audited, its positive edge is an upper bound, not an estimate; 33 of 36 rows remain unaudited.| Cadence | Action | Owner artefact |
|---|---|---|
| Daily | Silicon Data GPU indices; Federal Register BIS/232 docket sweep; MOFCOM announcements; secured/unsecured AI credit spread | Tripwire dashboard |
| Weekly | DRAM/NAND spot; 800G vs 1.6T module list prices; NVIDIA DSX/RVL qualification announcements; datacenter moratorium tracker | Node status sheet |
| Monthly | Taiwanese monthly revenue (Ibiden proxy set, Unimicron, GUC, Alchip, KYEC, AVC, Auras); Korean CCL import unit value; BLS PPI (transformers, switchgear, HVAC); TrendForce contract prices | Migration model refresh |
| Quarterly | Re-score every node's C/S/R/G/B/V; hyperscaler OCF/capex ratio; ASML EUV guidance; Lumentum/Coherent EML commentary; Veeco/AIXTRON AsP orders; colo vacancy and under-construction MW; backlog and book-to-bill across the switchgear/HVAC/prefab complex | Full atlas revision |
| Event-driven | Any prepayment, capacity guarantee or take-or-pay disclosure anywhere in the chain — highest information content available; treat as an immediate re-score trigger | Ad-hoc note |
Every known decision point, expiry and scheduled event that can move the chain. Two entries dominate: the 9-10 November 2026 policy pair, and PJM's 1 June 2027 curtailment priority.
| Date | Event | Mechanism of impact |
|---|---|---|
| 17 Aug 2026 | RTO/ISO responses due to FERC's six unanimous §206 show-cause orders on large-load interconnection | Sets the cost of capital for every interconnection request; ≥50MW / ≥69kV definition |
| 27 Aug 2026 | Comments close on adding electric conductor cables and heat-exchanger parts to §232 metals duties | Would hit fiber/cable, switchgear and chiller nodes directly — the three most cost-sensitive build inputs |
| Sept 2026 | Largan fiber-array-unit pilot line starts (<0.3µm alignment) | First capacity response in the fiber-to-PIC attach chokepoint |
| Sept 2026 | Synopsys Investor Day — detail promised on the pre-revenue IP royalty model | Would convert EDA into an Arm-like unit-levered royalty business. The tier's biggest unmodelled item |
| H2 2026 | BESI's promised clarity on whether HBM4E adopts hybrid bonding | Decides whether 2028 hybrid-bonding volume is memory or logic-only |
| ~3 Nov 2026 | US midterms; electricity prices live in GA, MI, MD, VA, TX | Ratepayer politics is the least-hedged risk in the atlas |
| 9 Nov 2026 | BIS affiliates / '50 percent' rule stay EXPIRES | Auto-listing of ≥50%-owned Entity List affiliates; snaps back absent extension |
| 10 Nov 2026 | China's rare-earth suspension EXPIRES — 0.1% extraterritorial rule returns | IEA: full reimplementation puts $6.5tn/yr of ex-China downstream output at risk. One-man veto, days-long fuse |
| 2 Dec 2026 | EU AI Act new Art. 5 prohibition (no grandfathering); Art. 50(2) watermarking grace ends | €15m or 3% of global turnover; enforcement powers already live since 2 Aug 2026 |
| 4 Dec 2026 | §232 polysilicon tariff: 15% plus minimum import prices ($21/kg poly, $100/kg wafers) | Direct cost on the wafer chain; note the semiconductor tariff itself exempts data-centre use |
| Late 2026 | TCLC hafnium-precursor core patents expire | Opens second-sourcing in a two-supplier-per-fab node under every DRAM capacitor |
| 1 Jan 2027 | Colorado CADMA and Illinois SB 315 take effect | First substantive US state AI compliance obligations to bind |
| 1 Jun 2027 | PJM IRAS: loads without their own generation are curtailed FIRST | The single most actionable dated event in the power tier. Reprices behind-the-meter generation instantly |
| 2027 | DFARS ban on Chinese-origin rare earths vs a 300,000-drone requirement by early 2028 | ~98% Chinese magnet supply against a 3-7 year domestic build-out. The arithmetic does not close |
| 2 Dec 2027 | EU AI Act Annex III high-risk obligations bind (deferred 16 months) | The first broad high-risk compliance wall |
| 31 Dec 2027 | §232 Annex III rate on grid/industrial equipment steps 15% → 25% | Adds 1-2pp to a 100MW build on top of the current 8-13% tariff burden |
| 1 Jan 2028 | Illinois mandatory third-party frontier-AI audits (>10²⁶ FLOP, >$500m revenue) | First external audit mandate anywhere; creates a licensed-auditor capacity question |
| 2028 | Kyber / 600kW rack step (slipped from 2027 on PCB midplane manufacturability) | Pushes the immersion-cooling take-off to ~2028-29. Anyone underwriting 2027 is a year early |
Green → Amber. Any single kill criterion partially met, or any tripwire firing outside its historical range. Action: halve the affected sleeve, do not reverse. Re-score the node and its two nearest neighbours in the tier map.
Amber → Red. Two kill criteria met, or the secured/unsecured AI credit spread widening more than 200bp in a month, or a single-counterparty credit event at OpenAI/Oracle scale. Action: exit Sleeve B entirely (it is the highest-variance and most illiquid), hold Sleeve A only where prepayment evidence remains current, and increase Sleeve C.
Red → Stop. Demand-side confirmation: two consecutive quarters of declining aggregate token volume, or hyperscaler capex guidance cut by more than 15% in aggregate. Action: the cycle thesis is invalidated; the atlas reverts to a reference map and the strategy is retired rather than adapted. Adapting a capex-cycle framework into a capex downturn is how supply-chain analysts became momentum traders in 2001.
Standing discipline. Re-read the low-confidence register before every re-score. Twenty of the load-bearing claims here are single-source, three are flagged for rejection outright, and several key quantities — per-rack UQD counts, gallons-per-MW water use, goodput at 100k+ GPU scale, agentic GMV, and every SaaS NRR series — do not publicly exist. A framework that forgets what it does not know will fail on the thing it never measured.
v1 scored six 0-10 fields and multiplied three of them. Diagnostics run against my own data killed that design on three counts, so it has been rebuilt. v1 is retained alongside rather than overwritten — the same discipline this atlas adopted in Part XII applies to my own scores.
Rank-space softened the one property v1 got right. The first cut promoted "OpenAI counterparty concentration" — MON 2, meaning no ownable expression whatsoever — by twenty places. That is precisely the error the framework exists to prevent.
The fix was to stop asking one number to do two jobs. MON ≤ 3 is now not ranked at all; those 26 nodes go to a systemic-risk register. They are constraints to underwrite against, not positions to hold — the rent accrues to a state, a queue, or nobody. The remaining 105 are ranked by RCI. This makes structural what the severity-versus-monetizability chart was already showing.
The Tier-1 fix: per-integer anchors tied to filings rather than impressions, so a second analyst lands within ±1 and a disagreement can be adjudicated against a document. Note level 6 — it was v1's default when uncertain, and v2 requires margin evidence to sit there.
| MON | Observable anchor required to score it |
|---|---|
| 10 | Sustained >50% segment operating margin AND documented customer prepayment, take-or-pay or capacity reservation AND price increases accepted without volume loss. |
| 9 | >40% segment operating margin with multi-quarter expansion, OR book-to-bill sustained above 2.0x, OR a disclosed capacity guarantee paid for in cash. |
| 8 | 30-40% segment operating margin and expanding, OR documented price increases passed through to customers and visible in reported gross margin. |
| 7 | Margin expanding but the exposure is diluted: the relevant business is a minority of parent revenue, or the holder is private so realization is inferred, not observed. |
| 6 | Pricing power asserted by management or implied by lead times, but NOT yet visible in reported margin. (v1 default when uncertain — v2 requires the margin evidence.) |
| 5 | Genuinely scarce, but long-term contracts, regulated returns or customer LTAs cap what the holder can realize in the current cycle. |
| 4 | Rent leaks to commodity or by-product economics: the holder cannot price the scarce attribute separately from a larger fungible output. |
| 3 | Rent accrues to labour, to a state, or to a queue — not to an ownable enterprise. |
| 2 | No ownable expression at all: the constraint is a policy, a permit, a standard or a physical law. |
| 1 | Scarce in principle but too cheap to matter, or fully commoditized. |
| RCI | Node | BIND | DEM | MON | v1→v2 rank |
|---|---|---|---|---|---|
| 87.9 | HBM (High Bandwidth Memory) | 83.3 | 10 | 9 | 2 → 1 |
| 84.6 | 200G/400G-per-lane EML die | 84.9 | 10 | 8 | 6 → 2 |
| 82.7 | SoIC / logic hybrid bonding capacity | 81.2 | 10 | 8 | 7 → 3 |
| 82.5 | ABF build-up film | 96.4 | 9 | 9 | 1 → 4 |
| 81.4 | High-power CW lasers for CPO | 78.8 | 10 | 8 | 10 → 5 |
| 80.1 | Permitted shell with firm power (colo) | 68.1 | 10 | 10 | 3 → 6 |
| 79.7 | High-layer FC-BGA substrate | 85.7 | 9 | 9 | 4 → 7 |
| 78.5 | TSMC N3/N2 leading-edge capacity | 84.2 | 9 | 9 | 5 → 8 |
| 76.4 | Frontier training at 10k+ GPU scale | 70.6 | 10 | 8 | 11 → 9 |
| 75.2 | Custom HBM base die on logic nodes | 68.1 | 10 | 8 | 12 → 10 |
| 71.9 | InP substrates | 93.8 | 9 | 7 | 13 → 11 |
| 71.5 | DRAM wafer-input trade ratio (the cannibalisation) | 76.8 | 9 | 8 | 14 → 12 |
| 71.5 | HBM stacking yield | 73.3 | 10 | 7 | 17 → 13 |
| 71.1 | Blind-mate quick disconnects (UQD) | 62.7 | 10 | 8 | 15 → 14 |
| 70.5 | NVIDIA component qualification (RVL / DSX validation) | 70.6 | 10 | 7 | 19 → 15 |
| BIND | MON | Node — severe, and deliberately NOT ranked |
|---|---|---|
| 100.0 | 3 | Heavy rare earths (Dy, Tb) |
| 96.6 | 2 | China rare-earth extraterritorial rule (10 Nov 2026) |
| 84.6 | 3 | State / RTO large-load tariff and siting regime |
| 82.2 | 3 | ArF immersion lithography (second source) |
| 79.7 | 2 | BIS affiliates / '50 percent' rule (9 Nov 2026) |
| 77.0 | 2 | Helium |
| 77.0 | 2 | Water rights and cooling-water permits |
| 68.8 | 2 | OpenAI counterparty concentration |
| 59.4 | 2 | Remote-access / cloud export loophole |
| 57.4 | 2 | Section 232 metals tariffs on full entered value |
| 57.4 | 2 | EU AI Act high-risk obligations |
| 38.6 | 3 | Cancellable 'backlog' |
N, the number of quarters until physical relief lands. A
pre-registered pilot tested whether N can be measured from primary filings at all. It cannot:
3 of 3 nodes returned NOT_QUANTIFIABLE. TSMC has never published CoWoS wafers per month;
Ibiden’s capacity-addition column is — in every row of every year; Ajinomoto’s
production-volume disclosure stops after FY2016; no switchgear maker publishes a numeric lead time.
M is measurable — commitments are disclosed. N is not — capacity and yield
are the competitive secret. A difference with one unmeasurable term is unmeasurable, and the pilot’s own
pre-committed stop condition fired. The scatter, the rankings and the −0.07 orthogonality result
below are retained as a record of what was built and how it failed, not as a live instrument. Do not trade
from them. Part XIX has the full retraction register.Everything up to here measures physical supply and demand. Opportunity is imbalance minus what is already discounted, and this atlas has had no way to compute the subtrahend. This section closes that, using a principle that needs no market feed at all.
You do not need the price to estimate what is expected. You need the behaviour of the parties who know more than the price does — the supplier, who knows its own capacity plan, and the customer, who knows its own order book. Both reveal their expected duration of scarcity by committing capital to it. An announced expansion is a forecast. A signed take-or-pay is a forecast. Both sit in filings.
N = quarters until the arithmetic says relief lands.
M = quarters of scarcity that committed capital implies the informed parties actually expect.
EDGE = N − M — quarters of shortage nobody has yet acted as though they expect.
M defaults to 4 quarters unless a dated expansion or multi-year commitment justifies more.
That default is the engine: it forces you to name the specific commitment, or concede that none exists.
| EDGE | N | M | RCI | Node — and the fact behind each number |
|---|---|---|---|---|
| +12q | 20 | 8 | 82.5 | ABF build-up film N: Brownfield +50% by 2030; the third plant breaks ground 2028 and operates 2032. Physical relief is five years out. M: Ajinomoto has committed ¥25bn and guided a 2030 brownfield step. Nobody — supplier or customer — has committed capital consistent with the 2032 date being the binding one. [M-AUDITED 2026-08-15, M HELD AT 8 — EDGE +12 STANDS: swept the Ajinomoto FY26/3 results deck, the Palliser Capital activist deck (Mar-2026) and the DigiTimes/Commercial Times chain. No customer prepayment, LTA or capacity reservation at the FILM layer found anywhere — and the activist deck is affirmative evidence in the same direction: Palliser's whole thesis is that Ajinomoto UNDER-invests (the ¥25bn 'compares unfavorably' to peers) and under-prices, demanding >=30% hikes and segment separation. An adversarial party incentivized to surface lock-ins surfaced none. First audited row where EDGE holds: audit tally 2 moved down, 1 held.] |
| +12q | 20 | 8 | 46.1 | Physical-design engineers (2nm + HBM4 + CoWoS-L) N: 5-8 years to create one; complexity grows ~50%/yr against ~20%/yr productivity. M: The Arm royalty ladder IS committed capital against this scarcity and it is multi-year. But it is sized to current programmes, not to the widening gap. |
| +10q | 16 | 6 | 49.0 | Skilled trades (electricians, pipefitters) N: 4-5 year apprenticeships against openings driven by retirement. Arithmetically unfixable this decade. M: Contractor backlogs are committed but priced on 1-2 year horizons. Nobody is underwriting the decade-long shape. |
| +8q | 16 | 8 | 71.9 | InP substrates N: JX is spending ¥120bn for 7-10x capacity THROUGH FY2030. That is the dated relief and it is four years out. M: MIGRATED 2026-08-15: M 5 -> 8, EDGE +11 -> +8. The prior claim — 'no customer has committed capital against the substrate layer itself; the prepayments went to the laser makers one tier up' — is FALSIFIED by disclosure: Coherent's MDSA with Beijing Tongmei (2026-06-26, 6-inch, 3-year) carries a $22.29M prepayment, and Lumentum holds a 6-year Tongmei capacity reservation with $43.5M + $43.5M deposits. Two named customers with dated, multi-year, cash-backed commitments at the substrate layer is precisely what M exists to count. Held below 9-10 because both commitments secure allocation at one supplier (AXTI's 85.5%-owned Tongmei) rather than funding the industry-wide build the way Modine's guarantee does. |
| +8q | 12 | 4 | 68.4 | Un-gameable verifiers for non-formal domains N: No oracle exists for non-formal domains; the substitute is licensed-professional time, which does not scale with capital. M: No disclosed capacity commitment anywhere. Nobody has put dated capital against this constraint at all. |
| +6q | 14 | 8 | 61.5 | Carl Zeiss SMT EUV optical columns N: Oberkochen's expansion took four years to deliver its first building; the 2029 EUV supply curve is explicitly 'under investigation', i.e. uncommitted. M: ASML has committed €4.41bn of purchases, €1.91bn of loans and two €1bn facilities. Enormous committed capital — but explicitly NOT extended to the 2029 step, which is the point. |
| +6q | 10 | 4 | 53.0 | EUV mask blanks N: No credible third entrant; the market is too small to attract one. M: No announced capacity response. Default — and here the default is arguably the truth. |
| +6q | 10 | 4 | 41.3 | Multi-beam EUV mask writers N: No qualified parity alternative; mask shops do not dual-source. M: No disclosed capacity expansion by anyone. Default. |
| +5q | 8 | 3 | 29.1 | Actinic EUV mask inspection N: Sole-source capability; orders lead revenue ~2 years and are violently lumpy. M: Orders DOWN 60% YoY. Behaviour implies participants expect near-term slack — the opposite of the structural position. Largest disagreement between behaviour and structure in the atlas. |
| +4q | 10 | 6 | 82.7 | SoIC / logic hybrid bonding capacity N: ~$6.8-7.0bn capex per 10,000 wpm against a 10-15K wpm target; gates Rubin Ultra 2027 and Feynman 2028. M: TSMC capex is committed and dated, but the disclosed programme is sized to 2027-28 demand, not to the Feynman-generation step. |
| +4q | 12 | 8 | 80.1 | Permitted shell with firm power (colo) N: Land plus power plus permit cycles run 3-5 years; NA under-construction MW FELL year-on-year. M: Pre-leasing and long leases with escalators are committed capital on both sides. But supply formation decelerating while absorption accelerates means nobody has committed to the back half of the gap. |
| +4q | 10 | 6 | 79.7 | High-layer FC-BGA substrate N: Ibiden staged from FY2027, Unimicron H2 2027, Shinko Malaysia late 2027 — the last capacity lands ~10 quarters out. M: Customers are being asked to prepay and four Taiwanese plus two Japanese suppliers are seeking binding commitments. That is real committed capital, but sized to the announced 2027 tranches, not beyond. |
| +4q | 9 | 5 | 60.6 | Low-loss copper-clad laminate (M8/M9) N: Doosan's Thailand capacity mass-produces H2 2028; CCL machinery order books run to 2028. M: One Seoul PCB maker placed 5x its normal monthly volume as an advance order. Customer commitment exists but is a single disclosed instance, not an industry-wide lock-in. |
| +4q | 8 | 4 | 40.6 | Agent identity and scoped delegation N: The technical substrate has settled but concentration is still low, so the rent has not formed. M: One consolidation transaction. Almost no committed capital against the constraint. |
| +3q | 8 | 5 | 84.6 | 200G/400G-per-lane EML die N: Lumentum's Greensboro contributes nothing material until early 2028; Veeco epitaxy deliveries peak 2027 so wafers arrive 2027-28. M: NVIDIA prepaid $2bn each into Lumentum and Coherent — a serious commitment, but sized to secure allocation rather than to fund the multi-year InP substrate build beneath it. |
| +3q | 8 | 5 | 64.9 | Burn-in boards, sockets and load boards N: Same 2027+ substrate timelines as FC-BGA, because the same three firms supply both. M: KYEC is spending NT$50bn for +30-50% high-power burn-in — real committed capital, but on the oven side, not the board side where the constraint actually sits. |
| +3q | 10 | 7 | 53.5 | Nearline HDD (HAMR) N: Sold out into CY2028 and — the load-bearing fact — ZERO unit manufacturing capacity is being added. Relief comes only from areal density. M: Long-term agreements cover 2027-28 and customers have accepted price increases. But no participant has committed capital consistent with 'unit capacity never expands', which is what the disclosures actually say. |
| +3q | 10 | 7 | 51.6 | Fiber draw, duct and trench (scale-across) N: Fiber draw 18-24 months and duct/trench 18-36 months. You cannot expedite a trench. M: Corning is spending ~$2bn in 2026 and hyperscalers have laid 120,000+ fiber miles — heavy, dated, two-sided commitment. |
| +3q | 7 | 4 | 50.5 | DPU / SmartNIC as the inference storage controller N: Merchant DPUs exist but are not qualified for the context-memory role; qualification is the clock. M: No dated third-party capacity or customer commitment. Default. |
The more useful half, and the one that validates the method: it independently reproduces three calls made earlier in this atlas from entirely different evidence. CoWoS at −5q is the fade. TSMC N3/N2 at −4q is capacity arriving ahead of plan. Gas turbines at −6q is the capital-cycle top with slots booked to 2031. Applied HVAC is the interesting one — a take-or-pay guarantee implies the informed parties expect longer scarcity than my own clock, so either their information is better than mine or the commitment is a negotiating artefact.
| EDGE | N | M | RCI | Node — and the fact behind each number |
|---|---|---|---|---|
| -6q | 6 | 12 | 58.3 | Heavy-duty gas turbine slots 2029-31 N: GEV to 30GW/yr by 2030; orders peaking 2026 — a textbook capital-cycle top. M: Slot reservations booked to 2031 and backlog+reservations went 100->116GW in a quarter. Committed capital implies three years more scarcity than the capacity cycle supports. |
| -5q | 2 | 7 | 58.1 | CoWoS 2.5D capacity N: Supply-demand gap narrowing 20%->10% by end-2026; capacity roughly doubling. M: Years of committed capex and the most-attended constraint in the complex. The clearest negative-edge node in the atlas: everyone is acting as though this binds long after the arithmetic says it stops. |
| -4q | 4 | 8 | 78.5 | TSMC N3/N2 leading-edge capacity N: N3 hit 180K wpm AHEAD of plan; N2 tracks to ~100K wpm by year-end. Physical relief is arriving early. M: Enormous committed capex on multi-year horizons. Behaviour implies far longer scarcity than the execution data supports. Strongly negative edge. |
| -2q | 8 | 10 | 43.7 | Applied HVAC / large chillers N: Chiller plant capacity 18-30 months. M: Modine took $165m CASH UPFRONT against a $4bn 2027-29 capacity guarantee. That is the single strongest commitment signal in the atlas — and it implies expected scarcity LONGER than my physical clock. Negative edge: the informed parties expect more than the arithmetic supports. |
m_why audit tally (2026-08-15): 3 rows audited — InP and CPO lasers moved down, ABF +12q
held (Ajinomoto IR, the Palliser activist deck and trade press all show zero customer capital
at the film layer). The bias is real but not mechanical. Read every remaining positive edge as an
upper bound until its M has been audited; 33 of 36 rows remain unaudited.None of these exist in this session. Each is ordinary buy-side infrastructure, and the first alone would do most of the work.
The direct read on what is already in numbers. If consensus is revising up, the physical story is being priced. Wide dispersion with a rising physical signal is the classic setup. Source: any estimates feed. This is the single highest-value addition and it alone would convert EDGE from a proxy into a measurement.
Absolute multiples are useless across this chain — a Japanese materials segment and a US semicap name are not comparable. Self-relative percentile is. It answers whether the market is already paying up for the scarcity.
Distinguishes 'nobody has noticed' from 'everybody who matters already owns it', which is the difference between an edge and a crowded trade. EDGE cannot see this at all.
This section exists because a separate, independently-built graph of the same industry applies harder evidentiary rules than I did, and testing my atlas against them cost me a headline finding. Recording that is more useful than quietly fixing it. Each item below states what the atlas did before, what it does now, and the price of the change.
What the atlas did before: This atlas treated a cessation claim as fact because several trade-press outlets carried it consistently.
What it does now: Cessation, bankruptcy and M&A claims now require an exchange filing or a company statement. Trade-press or research-firm relay is insufficient regardless of how many outlets repeat it, and for a listed supplier the absence of a legally-required disclosure counts as disconfirming evidence rather than as neutral silence.
What it cost: It cost me a headline finding. The WF6 tungsten-hexafluoride exit — ~25% of global capacity leaving on 1 July 2026 — was in my original twelve theses' supporting evidence and is now marked DISCONFIRMED, with expansion rigidity and substitutability cut as the mechanical consequence. Node 6's rent index fell from 33.0 to 19.3.
What the atlas did before: I had a single three-level confidence flag (H/M/L), which cannot express 'two incompatible hypotheses, both with evidence stacks'.
What it does now: A claim can now be carried as disputed with both hypotheses and both evidence stacks attached, rather than being silently resolved in favour of whichever side the analyst finds more plausible. The graph project models this at category level with an explicit exit-scenario object whose status is disconfirmed — the counterfactual survives as data.
What it cost: Nothing, and I should have had it. A framework that can only say 'high/medium/low confidence' quietly forces a verdict on every contested claim.
What the atlas did before: My dossiers carry ~1,100 inline URLs and no content hashes, so link rot degrades them invisibly and a page edited after I read it is indistinguishable from one that was not.
What it does now: The right unit of provenance is {url, a ≤15-word verbatim clause that actually carries the claim, a hash of the fetched page text, the fetch date}. The hash is what makes the citation falsifiable later.
What it cost: I cannot retrofit hashes — this session's search budget is spent and I cannot re-fetch ~1,100 pages. So the field structure is adopted and the hashes are owed. That is a real, dated debt, not a completed upgrade.
What the atlas did before: Requalification times were scattered through my prose — 2-3 years for a high-purity-HF second source, 12-18 months per HBM generation, 18 months for a specialty gas — and were therefore unusable for anything but reading.
What it does now: qual_time_months is the delay constant: a shock at the supplier reaches the buyer as margin or volume pain for at least that long. Once it is a field rather than a sentence, shock propagation becomes computable instead of narrated.
What it cost: Low. It also exposed that several of my prose estimates were ranges I had never committed to a single number.
What the atlas did before: My Investable Rent Index blends everything into one number, and my own red-team showed the blend is dominated by demand-exposure (ρ 0.917) rather than by severity (ρ 0.566).
What it does now: The graph keeps them apart on purpose: how much the buyer bleeds without the input class; how much of that input comes from this seller; how fast the bleeding can be stopped; and whether the relationship is believed at all. Its schema says the merged version is unusable for propagation, and on the evidence of my own diagnostics that criticism lands.
What it cost: This is the most uncomfortable item here. It says my headline index is the wrong shape for anything beyond ranking, which is roughly what my own framework-limitations section concluded independently.
What the atlas did before: A decayed node in my atlas simply reads as decayed; the counterfactual is gone.
What it does now: An ended relationship keeps its record with an explicit end date and renders as severed. History is signal — and severances are precisely what a propagation engine reasons about.
What it cost: None. My thirteen DEBUNKED entries were already an instinct in this direction; this makes it structural.
Applying the primary-disclosure rule retroactively. None of these were fabricated; all were carried by sources I considered adequate and now do not. The most consequential is the first.
| Claim | New status | Why it fails the test |
|---|---|---|
| The WF6 permanent-exit claim | DISCONFIRMED | A cessation claim carried only by trade-press relay. Corrected in this revision; scores reduced accordingly. |
| Ajinomoto's reported +30% ABF price increase effective Q3 2026 | STRUCK, then UPGRADED TO [M] 2026-08-15 | Originally struck as aggregator-sourced. The 2026-08-15 audit found trade-press carriage: DigiTimes and Taiwan Commercial Times (13 May 2026) report substrate makers were formally notified of ~30%, effective Q3-26 — one tier below a filing, so it stays out of the evidence field and the margin remains the carrier. Timeline: Palliser Capital's activist deck (Mar-2026) demanded ≥30% two months earlier; the deck is also the likely origin of the garbled aggregator variants, including a ChemNet claim of a '30% China supply cut' that no credible outlet carries and which is REJECTED. |
| Ajinomoto's >95% share of ABF film | DEMOTED | Trade-press only. Retained as a directional claim in prose, but it must not become a numeric share field anywhere. The >50% segment operating margin — company-disclosed — is what actually carries the concentration argument. |
| Intel "asking customers to prepay" for substrate capacity | DEMOTED | Trade-press reporting of a company statement, not a disclosed prepayment. It is still the strongest scarcity signal in the substrate tier, but it belongs one tier below a filing in the evidence hierarchy. |
| Every M&A price in the optics section (Credo/DustPhotonics $750m, Molex/Teramount, Ciena/Nubis $270m) | NEEDS PRIMARY | M&A is an extraordinary lifecycle claim. For the listed acquirers an 8-K or equivalent carries it; until cited, the transactions are directionally usable and the prices are not. |
| Lumentum's "InP tightness exceeds DRAM and NAND" | TAGGED | A supplier describing the scarcity of its own input, i.e. management commentary — the second-lowest tier in my own evidence hierarchy. Already flagged in the atlas; now formally classed rather than merely noted. |
My original red-team pass caught eight arithmetic and consistency errors and one inverted statistic. It did not catch the WF6 exit, because the claim was internally consistent, widely repeated and arithmetically sound. Consistency checking cannot detect a well-formed claim with insufficient carriage — only a source-class rule can.
So the two verification passes are orthogonal and both are necessary: one asks "do the numbers agree with each other?", the other asks "is anyone actually on the hook for this claim?" I ran the first and not the second, and the second is the one that found the real error. The debt I still owe is the content hashes — I cannot retrofit ~1,100 of them in this session, and until they exist every citation in the dossiers is one silent page-edit away from being unverifiable.
Sixteen parallel research workstreams produced fifteen dossiers totalling ~192,000 words and 1,448 citations (1,050 distinct URLs, 323 hosts), executed 8 August 2026. I then scored 131 nodes on six dimensions and derived severity and rent indices from those scores. Every number in this atlas traces to a dossier; every dossier carries its own low-confidence register, and the dossiers themselves are indexed in Appendix A.
sec.gov appears eight
times across a 192,000-word corpus about listed companies. The regulatory workstreams did comply —
federalregister.gov at 19 is among the largest genuinely primary hosts — and the company-level
workstreams largely did not. This is disclosed rather than corrected, because correcting it means re-sourcing the
claims, not re-labelling them. Appendix A carries the full breakdown, including a correction
to a harsher first estimate of this same ratio.conf = L in the database (7 of 131) and should be verified before sizing. They are: the EMIB-M 8-12x reticle claim and the merchant-2.5D allocation asymmetry built on it; the ~$6.8-7.0bn capex per 10,000 wpm SoIC figure; SUSS's ~12 temporary bonders per year; Supermicro's preliminary unaudited Q4 FY26 gross margin; the ~23-35% estimate of NVIDIA revenue traceable to funded entities; water rights (no public gallons-per-MW figure exists at all); and model-weight security. Separately, three claims carried in conf = H or M nodes are weaker than their host: Ajinomoto's widely-quoted +30% Q3-26 price increase is aggregator-sourced rather than company-disclosed (the operating margin is better evidence of the same thing, and the price claim has been removed from the evidence field); Lumentum's 'InP tightness exceeds DRAM and NAND' is a supplier describing its own scarcity; and Intel's substrate prepayment request is trade-press reporting of an Intel statement, not a disclosed prepayment.Five corrections surfaced by reading Chinese primary disclosures against claims this atlas carried. Documented the way Part XIII documented the WF6 error: what was said, what the filings say, and the mechanical consequence. Analyst-entered values are migrated in the open, never silently overwritten.
Beijing Tongmei is qualified at datacom volume: a Coherent MDSA dated 2026-06-26 (6-inch, 3-year) with a $22.29M prepayment, and a Lumentum 6-year capacity reservation with $43.5M + $43.5M deposits. Tongmei is 85.5%-owned by NASDAQ:AXTI; its STAR IPO was withdrawn 2026-06-26/07-08. The deeper consequence is in the edge layer: this atlas's M for InP substrates rested on "no customer has committed capital against the substrate layer itself" — falsified by those same two disclosures. Migration: M 5 → 8, EDGE +11q → +8q. By the atlas's own evidence hierarchy (prepayment ranks first) this was not an optional re-score.
Previously carried: "San'an 70–100mW, 4–8× too weak." San'an confirmed — but Shijia 688313 discloses a 400mW CW DFB at 小批量 (small-batch production) and Yuanjie 688498 has 300mW at sampling. The paper power gap is ~1–2×; the real gap is Telcordia GR-468 qualification, a named CPO customer, and volume. Migration: N 8 → 6 (the relief clock is a qualification timeline, not the Western epitaxy build), M 4 → 5 (a disclosed small-batch line is dated committed capital), EDGE +4q → +1q. This was the highest-priority re-score flagged by the qualification pass, and it demotes the node from the top-edge list. Portfolio consequence: the COHR thesis-break trigger ("Chinese 400mW CW DFB with a named CPO customer and GR-468 data") is now partially met — power yes, qualification no.
Shengyi's 235-page audited FY2025 annual report contains zero occurrences of M9/M8/M7/英伟达/超低损耗/低损耗 — while the consensus narrative carries it as the qualified Chinese name. Shengyi remains in holders as a volume CCL maker only. The better-evidenced Chinese entrant at AI grades is Dongcai 601208, which names 英伟达 in an audited AR under securities-law liability. Same lesson as WF6: keyword-verify the audited filing before carrying a qualification claim.
Two errors from the same pass: Big Fund III's AKM Meadville investment dated "January 2026" (correct: December 2025 — the 2 Jan TrendForce piece was the reporting date), and Shennan's FC-BGA fab placed in "Wuxi" (correct: Guangzhou; Wuxi is the older BT/memory-class site). Both claims were checked against every data file and built artifact in this tree and appear nowhere — they lived in session prose only. Recorded here so they cannot regress into the CL6 build, where both names are expected to appear.
This atlas carried no provenance layer. Its sister repo (ai-supply-graph) carries 257
verbatim anchors with sha256 and a carriage verifier; this one carried a free-text evidence field and a
self-graded confidence letter. Measured: of the 25 nodes scoring V=6, 22 named no carrier at all,
and 16 of those carried conf = H. That is a citation gap, not a thinking gap — the claims are
specific and falsifiable, they were simply unsourced. An anchor store and validator now exist
(data/anchors.json, verify_carriage.py). Coverage at this pass: 9 of 92
decomposed claims. Everything below is a migration shown in the open; no score was changed.
A demote of AI-grade 300mm wafers V 6 → 5 was proposed on the basis that “SUMCO honours LTAs to end-FY27 and raises price only from FY28.” That claim cannot be pinned. No SUMCO document contains FY2027 or FY2028 in connection with LTAs — not the Q1 deck, Q2 deck, Q2 tanshin, or the 2026-08-07 半期報告書. SUMCO publishes its call as audio only, so every “2027/2028 expiry” figure in circulation is a paraphrase of that audio. What the company does say (2026-08-06): “Long-term contract prices were honored. Negotiations have begun on spot price revision.” — which is nearer the V=6 anchor than V=5. Demote withdrawn. V=6 stands. This is the 2026-07-11 staleness rule working as designed: a weaken/demote finding must survive a current-state check first.
This atlas read SUMCO’s weakness as the node’s weakness. The primaries say otherwise. 立昂微’s H1 report (2026-08-08) states the global top three raised prices twice in 2026, with high-end AI/HPC wafers up as much as 20%. GlobalWafers ran 12-inch fully utilised at a 9.4% operating margin (2026-08-04). Shin-Etsu’s Electronics Materials OP grew 23% (2026-07-24). SUMCO’s breakeven is a balance-sheet condition — H1 depreciation ¥64.4bn exceeded H1 EBITDA ¥53.9bn — not an industry law.
Four carried figures are wrong or stale. “¥5.2bn operating loss” was 1Q-2026; 2Q was ¥1.1bn and 3Q is guided to exactly ¥0.0bn — the record overstates the run-rate roughly fivefold. Yoshinogari was deferred, not cancelled (METI supply-plan change 2026-03-27; maximum subsidy cut ¥75.0bn → ¥19.3bn). Siltronic’s capex cut is −40% to −51% (FY26 guidance €180–220m vs FY25 €369m), not −38% — that figure matches only H1 cash capex. And the disconfirming claim introduced during this pass — “China is 28% of global wafer supply and already serves ~60% of demand” — is itself relay, tracing to a sell-side note via a think-tank, pre-cutoff. 立昂微’s audited filing says the overseas top three hold >70% of global share and China’s 12-inch localisation rate is ~10%. Chinese volume is real (ESWIN shipped >1.0m 12-inch wafers in June 2026; NSIG is at 1.0m/month) but its prime-wafer ratio is still ramping and it sells at ~20% discount, with NSIG’s H1 loss widening to RMB −965m on ASP −17%. Defensible on nameplate count; indefensible on qualified prime supply. The “60% of demand” leg has no source and is dropped.
This atlas carried “JEDEC is relaxing HBM4E height to ~900µm (vs HBM4 ~775, HBM3E ~720)” with no source. An independent sell-side note carries the identical figures, which reads as corroboration and is not. JEDEC’s JC-42.2 committee document list, current through August 2026, holds five items — none of them HBM4E. A height specification for a generation with no standard cannot be in the standard. Every carrier of the figure names Korean media, not JEDEC, and the number is still moving: 825–900µm in March, 900µm in April, “as much as 1,000µm” for 20-high HBM5 by July. A published spec does not drift 175µm in four months; a committee negotiation does. 720/775 are real standard content (JESD270-4A v1.1, Dec 2025) and stay. 900 is downgraded to “reported direction of travel, magnitude unstable.” The node’s underlying call — that HBM hybrid bonding is roughly two years later than consensus — survives on other evidence, including Besi’s own 2026-07-23 disclosure that evaluations are running at only one of the three memory makers.
CVD-SiC and graphite chamber consumables. The node is scored on pricing power implied by lead times. Tokai Carbon’s H1 FY2026 bridge (2026-08-05/06) shows price as a headwind: 売価 −¥330m in the revenue bridge and −¥336m in the OP bridge, with the entire gain coming from volume (+¥4,387m) and FX (+¥1,245m). Fine Carbon grew revenue 19.1% and OP only 6.9%. That is volume, not rent. Note also the earnings-exposure asymmetry: Fine Carbon is 19.0% of group revenue but 31.2% of group operating profit.
Temporary wafer bonders. SUSS MicroTec’s H1 2026 report (2026-08-06) states plainly that “the Bonding Solutions product line recorded a significant sales decline,” with group margin hit by “a lower proportion of high-margin temporary bonders.” ABS EBIT fell to 5.6% from 11.1%. Order intake nearly tripled — but the €115m OSAT order was coaters, not bonders. The carried “~45% temporary bond/debond share” is a company self-estimate last restated in the FY2025 annual report and not repeated in the H1 release.
Actinic EUV mask inspection. Lasertec’s order recovery is confirmed and large — ¥105,226m → ¥237,504m, 2.26× (FY6/2026 tanshin, 2026-08-06), semiconductor equipment orders +208.5%. But the gate named in this atlas (“A14 qualification”) is wrong: management names A200HiT customer evaluation at N2/N3, and says explicitly that some customers “may not wait for the completion of the evaluation.” More consequentially, the carried share of 100% is eroding at the edges: KLA disclosed on 2026-03-12 that it is “developing an actinic inspection system and print check” with prototype imaging complete, and has commercially launched TeraBeam 8XX, a multi-column e-beam EUV reticle inspector co-developed with TSMC; ZEISS shipped AIMS EUV 3.0 in February. Lasertec’s own CEO conceded in May that competitors “eventually will” launch. All competitive evidence predates 2026-07-01 and is marked stale; the 100% share needs a review date, not a re-score.
The V rubric’s anchors are financial observables — segment operating margin, book-to-bill, price pass-through visible in reported gross margin. Those are properties of a company. V is scored on a node. When a node’s holders diverge, the node inherits whichever holder the analyst happened to read. The wafer node is the clean case: GlobalWafers at 9.4% OP fully loaded and Shin-Etsu at +23% segment OP sit in the same node as SUMCO at breakeven under a depreciation wall. Hanmi is the mirror image — H1 2026 operating margin 45.95%, Q2 alone 51.9%, which is V=9 territory on the rubric, inside a node scored 6 because hybrid bonding has printed nowhere. Both readings are defensible and they cannot both be a single number. V must be scored per holder and aggregated, with the aggregation rule stated. This is the largest open item this pass produced, and it is a schema question, not a data one.
verify_carriage.py --strict fails on any relay-tier carrier), and it flags an honest
PENDING hash rather than accepting a fabricated one.EDGE +Nq figure does not — see Part XIX. The
arithmetic on prices, cash flows and margins in these cases is unaffected and was the point of the exercise;
the EDGE annotation was decoration on top of it.The atlas scores nodes. It cannot be acted on until a node is converted into a holder and a holder into a price. Five conversions were attempted on 22–23 August 2026 using reverse-DCF and implied duration against the industrial evidence — explicitly not backtesting or out-of-sample factor validation, which this dataset is far too short to support. All five came back no. The four different ways they came back no are the finding.
| Name | Node (RCI rank) | How it failed |
|---|---|---|
| Coherent · COHR $289.52, EV ~$60.1bn |
EML die #2 (84.6) also CW lasers #5 |
Composition. Third-party channel work puts the EML top three at Lumentum / Broadcom / Mitsubishi ≈72% combined — Coherent is not in it; its leadership is in CW-DFB, a different row of this atlas. The price requires ~20.6% revenue CAGR for ten years at management’s own margin target, against a node measured at EDGE +3q. FY26 free cash flow −$1,023M against +$1,097M non-GAAP net income; 212 days of inventory financed by stretching payables from 82 to 156 days. Incremental segment profit 28.0% against Lumentum’s 54.0% incremental operating margin. |
| SK hynix · 000660.KS ₩1,730,000, net cash ₩69.4tn |
HBM #1 (87.9) | Duration. The 76% Q2’26 operating margin is verified — and the quarter was a miss (₩60.54tn vs ₩64tn consensus), stock −9.6% on the print. The price discounts ~5.5 years of that margin, or a perpetual 39.5%, against a node carrying EDGE 0q. At the late-June peak it required a perpetual ~70%; ₩916tn of market value has been erased since. The 2018 analogue is the load-bearing evidence: SK hynix printed the best year in its history and the stock fell 23%, with the P/E troughing at 2.54x at the earnings peak. HBM share 62%→58% after NVIDIA qualified all three suppliers for HBM4 on 5 June 2026. |
| Ajinomoto · 2802.T Ibiden · 4062.T |
ABF film #4 (82.5, EDGE +12q) FC-BGA #7 (79.7, +4q) |
Already capitalised. This is the node that best satisfies the corrected selector — >95% share, the highest EDGE on the board, and Functional Materials at 7.0% of Ajinomoto’s sales but 32.4% of group business profit on a 57.7% margin. Taking the food stub at 18x, the market already pays ~47x EV/business profit for the film. Ibiden is on 41.0x EV/OP with the ¥500bn programme depreciating declining-balance, implying ¥50–86bn of incremental annual D&A peaking FY2028–29. And 76% of Ibiden’s FY2026 operating-profit upgrade is ASP/mix, which has an expiry date the multiple does not reflect. |
| Siltronic · WAF.DE €81.05, P/B 1.31x |
wafers #72 (33.2) retained as a control |
The node itself does not earn. Depreciation is straight-line over 4–10 years, so the roll-off mechanism that makes the SUMCO story work does not exist here as accounting policy — it is a plateau, with €2,218.9m of assets under construction not yet depreciating. At its all-time-best 2022 margin applied to 2026E sales, EBIT is still −€24m. A June 2026 equity raise at €91, three months after disclosing a possible 2027 covenant breach, now sits above the market price. |
| SUMCO · 3436.T ¥3,384, P/B 2.08x |
wafers #72 (33.2) | Entry point gone, and a relay figure hid it. The setup was built on P/B ~1.32x taken from a sell-side note; actual P/B is 2.08x. 1.32x maps to Q1 2026 — the note described an entry that had already closed. The stock is +186% off the December-2024 trough close. The depreciation relief is real (machinery is 定率法 declining-balance, mainly 5-year life; the CFO stated in writing that D&A peaks FY2026 and declines from FY2027) and it does not rescue the case: annualising H1 and flowing price straight to EBITDA, a +10% blended price rise still yields only a 3.3% operating margin. |
Four distinguishable failure modes, not one repeated complaint: composition (Coherent), duration (SK hynix), already capitalised (the ABF complex), and the node does not earn (Siltronic, and SUMCO’s closed entry). The uncomfortable common factor is that in none of the five was the market wrong about whether a chokepoint exists. Identifying chokepoints is consensus. The residual edge, if there is one, is in duration and in holder attribution.
The corollary is testable and points at the part of this atlas that is least finished: the map should be most valuable where BIND is high and V is low — rent that is structurally real but has not yet printed in reported margin, and therefore cannot already have been capitalised. That is the V=6 cohort, whose provenance debt is documented in Part XVI. The audit queue was pointed at the right cohort for entirely the wrong reason.
holders is an unordered set with no share attribution. Node #1 carries shares
(“SK hynix 55–62% bit share”); node #2 carries none, and Coherent — named there — is not
in the EML top three. A reader, including the author, infers “named holder of #2” as “holds the #2
position.” Proposed:
{name, ticker, share_pct, share_conf, share_src, node_share_of_holder_EV}. Until it lands, the standing
rule is that a named holder is a lead, never a position — the same epistemic class as a relay source.V is scored per node while every V observable is holder-level. Inside the single wafer node:
GlobalWafers 9.4% operating margin, SUMCO around breakeven, Siltronic negative at its best-ever margin. One score
cannot describe all three. The holder–ticker work found only 2 of ~45 holders report the named business
as a reportable segment with revenue and operating profit, which is the harder half of the problem.All three are flagged for hand entry. data/chokepoints.py holds
analyst-entered values and is never machine-written, so none of this has been auto-applied — the faults are
disclosed here and the scores are unchanged.
Appendix A discloses that ~58% of this atlas’s evidence base is relay. This is the work of fixing that, node by node, and it is the only step that changes what this document is. Ten nodes were put through primary-source sourcing on 23–24 August 2026: RCI #1, #2, #4, #6, #7, #8, #11, #18, #24 and #28. Every returned anchor was then re-fetched by machine and checked as a verbatim contiguous span before being accepted. Nothing here rests on an analyst’s report of what a document says.
| 22 Aug | 24 Aug | |
|---|---|---|
| Anchors in the store | 26 | 98 |
| Nodes with any anchor | 9 | 35 |
| Verified verbatim (CARRIED) | 0 | 57 |
| Nodes with a verified anchor | 0 | 17 |
| Resolved content hashes | 0 | 48 |
| Anchors that failed on re-fetch | — | 0 of 72 |
Roughly fifty atlas claims were put to primary sourcing. Thirty-two came back CANNOT PIN — no issuer, regulator, standards body or statistical agency publishes the fact. That is not a failure of searching; it is the measurement this atlas most needed, and it falls in a pattern. Contract-price indices, market-share splits, wafer prices, per-node capacity, lead times, industry-revenue aggregates and vacancy/absorption series are simply not disclosed by anyone. They are the product of TrendForce, Omdia, CBRE and JLL, and this atlas has been carrying them as if they were facts of the same class as a filed margin. They are not.
The Coherent case in Part XVII turned on a composition argument: that third-party channel work puts the EML top three at Lumentum / Broadcom / Mitsubishi and excludes Coherent. That rests on a relay source, and the primary record does not support it. Coherent’s own newsroom describes “200G EML solutions for 1.6T transceivers” and 400G/lane differential EMLs; its FY2026 10-K lists EMLs among its products. All five holders named on node #2 make first-party EML claims.
What survives is narrower and more useful: no primary source anywhere establishes EML market share. The only rank statement in existence is Mitsubishi Electric’s “No. 1 global market share in EML chips for data centers” — and its own footnote says “according to Mitsubishi Electric estimate”. So the ~72% top-three figure could be neither corroborated nor refuted from filings. The Coherent case’s other legs — ten-year implied duration, the −$1,023M FY26 free cash flow against +$1,097M non-GAAP net income, and 28.0% incremental segment profit against Lumentum’s 54.0% — stand on their own and are unaffected.
Reconciled 21 September 2026 against the anchor store (data/anchors.json, per-anchor carriage and checked fields). The store holds 98 anchors. The 24 August pass re-fetched 72: 47 CARRIED (verbatim span found in the live document) and 25 UNVERIFIED — not rejected — because their hosts are unreachable from the verification environment and the documents (a 1.5MB Form 20-F, a 10-Q) exceed the fetch tool’s content window, so carriage is unproven in either direction; 0 NOT CARRIED. The 23 August pass covered the other 26: 10 CARRIED, 11 UNREACHABLE, 2 UNREADABLE, 2 VIEWER_SHELL and 1 NOT CARRIED (document readable, anchor text absent). Totals: 57 CARRIED, 25 UNVERIFIED, 11 UNREACHABLE, 2 UNREADABLE, 2 VIEWER_SHELL, 1 NOT CARRIED = 98. The 28 August text read as if 57 + 25 = 72; the correct decomposition of the 72 is 47 + 25, and the 57 CARRIED span both passes. An UNVERIFIED anchor is a limit of the verifier, not evidence against the claim, and
recording it as a failure would manufacture exactly the false demotes this layer exists to prevent. The four-way split
introduced on 23 August earned itself immediately: of that day’s first four NOT CARRIED verdicts, exactly
one was a real anchor failure — the others were a viewer-shell URL and two extraction failures.
The honest summary of this document’s evidentiary state: 57 claims now rest on a verbatim span in a primary document that a machine has re-read, against 131 nodes and several hundred assertions. That is the beginning of a provenance layer, not the possession of one.
The v1 draft of this document was completed 24 August 2026. In the two days after that its central instrument was destroyed by its own pre-registered tests, along with four other claims. Nothing above has been quietly rewritten. This section is the record, and it is the first thing a reader should use to discount everything else.
M, the anticipated duration, is measurable: commitments are disclosed. N, the physical
duration, is not. A pre-registered pilot tested three nodes and returned NOT_QUANTIFIABLE 3 of 3:
TSMC has never published CoWoS wafers per month; Ibiden’s capacity-addition column is — in
every row of every year; Ajinomoto’s production-volume section disappears after FY2016; no switchgear maker
publishes a numeric lead time. A difference with one unmeasurable term is unmeasurable. Part XII is retained as a
record of the construction and its failure.
This document predicted its own failure and the warning was not acted on. Part XII already
disclosed, on 15 August 2026, that corr(EDGE, N) = +0.88 against corr(EDGE, M) = −0.24
— that EDGE was “overwhelmingly the physical clock” and “substantially mechanical.” If
EDGE is almost entirely N, then EDGE is only as measurable as N is. That inference was
available eleven days before the pilot ran, in a caveat this document published about itself. The lesson recorded
is not that the caveat was missing; it is that a disclosed weakest joint was left as a caveat instead of being
tested.
EDGE figures also appear inline throughout Parts I and XVII and in several node annotations. All of them are withdrawn, wherever they appear.
Tested on a cash-flow panel rebuilt from primary tanshin filings (162 rows, fiscal years derived from
quarter_end rather than the guidance year). The sign flip survived a materially doubled panel, which
was the pre-registered kill condition. Advantest showed ρ = +0.006 at T+1 on 13 clean quarters.
0 of 30 nodes have both segment revenue and segment operating profit. The finding that “no map field explains the cross-section” was therefore measured on whole-company financials for 23 of the 30 rows. It is void: it did not fail, it was never a valid test of the map.
Expected twelve-month return −3.2% gross, −15.7% net of borrow and beta drag, with a 44% chance of simply being up. What survives is narrower and is not a recommendation: guided FY2026 opex flat in dollars on +32% revenue implies 2.31× earnings leverage, and only $2.1bn of the $5.0bn backlog is non-cancelable RPO.
Intended as the replacement for the unmeasurable N. It requires a corpus of dated public technology
commitments. That corpus does not exist in machine-readable form, for a structural reason: dated forward
commitments in SEC filings and earnings calls carry securities-litigation exposure, so counsel strips them, while
the identical claim in a keynote or press release does not. Measured both ways. NVIDIA returns zero
filings containing “co-packaged optics” across all of EDGAR 2016–2026, and Broadcom zero, with a
control query confirming the filter works. A pre-registered recall test on five known commitments scored
0 of 5 against the corresponding earnings calls. In every case where the technology did reach the
call, the commitment survived and the date did not.
An earlier audit stated that Ajinomoto’s functional-materials margin “cannot exist” and that the inputs came from an activist deck. Both statements are false. ¥1,206 億 sales and ¥655 億 business profit are Ajinomoto’s own disclosed full-year forecast, published quarterly in its supplementary results material. The error was searching the annual report’s IFRS 8 segment note, not finding it, and reporting “I could not find it” as “it does not exist” — then escalating that to an accusation about the source. The row is company guidance.
Four scoring rulings remain outstanding and require the author’s judgement rather than a script: the skilled-trades and PJM velocity scores, a Credo/DustPhotonics mapping error, and HG Genuine’s listing status. Separately, the Largan node assignment is refuted on the evidence — a 254-page review found no fiber, PIC, FAU, waveguide or transceiver language anywhere — but the scored row has deliberately not been machine-edited, because analyst-entered values in this dataset are never overwritten by a script.
Every URL behind this atlas, grouped by the workstream that produced it, plus the composition disclosure that tells you how much weight the evidence base can bear. Per-node source lists are inside each row of the chokepoint atlas — click a row to expand it.
| Class | URLs | Share |
|---|---|---|
| primary | 225 | 21.4% |
| relay | 601 | 57.2% |
| unclassified | 224 | 21.3% |
Weighted by citation occurrence (1,448 in total) the split is materially the same: relay 59.8%, primary 21.0%, unclassified 19.2%.
Primary is the entity that owns the fact publishing it: issuer IR and newsroom pages, regulatory filings (SEC, DART, TDnet, cninfo), standards bodies, statistical agencies, and academic preprints by the authors. Relay is someone reporting on somebody else's fact: trade press, channel research firms, aggregators, encyclopedias. Under this atlas's carriage rule — ratified in the Kanto/Central-Glass ruling of 10 July 2026 — a relay source may tell you where to look; it may not itself carry a rent-capture or lifecycle claim.
The practical consequence for a reader: the single most-cited host is TrendForce (229 citations), which is relay. Where a claim in this atlas rests on a relay source alone, it is corroboration, not proof, and should be treated as a lead to verify rather than a fact to underwrite.
| Citations | Host | Class |
|---|---|---|
| 115 | trendforce.com | relay |
| 58 | datacenterdynamics.com | relay |
| 47 | convergedigest.com | relay |
| 35 | digitimes.com | relay |
| 33 | stockanalysis.com | relay |
| 28 | stocktitan.net | relay |
| 27 | theregister.com | relay |
| 25 | epoch.ai | primary |
| 19 | prnewswire.com | relay |
| 16 | federalregister.gov | primary |
| 15 | techcrunch.com | relay |
| 13 | investing.com | relay |
| 13 | tomshardware.com | relay |
| 12 | futurumgroup.com | relay |
| 11 | blog.cloudflare.com | primary |
| 10 | newsletter.semianalysis.com | relay |
| 10 | datacenterfrontier.com | relay |
| 9 | power-eng.com | relay |
| 8 | bls.gov | primary |
| 8 | utilitydive.com | relay |
| 7 | arxiv.org | primary |
| 7 | semiengineering.com | relay |
| 7 | cnbc.com | relay |
| 7 | jdsupra.com | relay |
| 7 | semi.org | primary |
| 7 | techtimes.com | relay |
| 6 | developer.nvidia.com | primary |
| 6 | fred.stlouisfed.org | primary |
a URL attaches to a node only on an exact node-name or holder-name mention within the preceding 1200 characters of dossier text; unattached URLs remain in the appendix rather than being guessed into a node No fuzzy matching and no topic inference. The rule under-attaches on purpose: a missing citation is recoverable, a wrong one is not.
Fifteen dossiers, ~192,000 words. Colour marks class: primary · relay · unclassified.