The system
Value flows to whichever physical layer is the binding constraint. The map holds the candidate layers, including ones not yet named. Where and how hard the constraint binds is read from filings each quarter, never assumed — and when innovation routes around a constraint, the map is built to show it relaxing.
The constraint map
Rows are physical layers; columns are the five tells a filing can show. A cell is PRESENT only when found in a filing. Where more than one company has been read on a layer, the cell shows how many of them show the tell — "2 of 3" — never the union. Evidence chips: [F] filing-grade; [F-KR/PROV] a provisional Korean pattern set that a Korean-reading reviewer may overturn. Hover a cell for the names.
Corrected 2026-09-18. Originally read: “ASSESSED — four expressers: TSEM (prepayments PRESENT-provisional, reservation held), AAOI (five ABSENT), COHR (4 of 5, ruling), LITE (1 of 5 — lead_time on the supply-allocation sentence, ruling; one read pending). Intersection 0 of 5 · all 4 names; any 5 of 5.” The hand-written note stated an expresser count and intersection that disagreed with the cells; the counts were struck.
The five tells, defined
| signature | plain English | what it looks like in a filing | strength |
|---|---|---|---|
prepayments | Customers pay before delivery. | Customer deposits or contract liabilities growing faster than revenue; "advance payments from customers". | Descriptive, not a weight: strong — cash moves before product does. The score counts all five tells equally (PRESENT 1, WEAK 0.5). Corrected 2026-09-18. Originally read: “Strong — cash moves before product does.” The cells read as a ranking of weights; the score counts all five tells equally, so they now say they are descriptive. |
capacity_reservation | Customers book future output in advance. | "capacity reservation agreement", "long-term supply agreement" with committed volumes or reserved wafer starts. | Descriptive, not a weight: strong — commitment to a specific supplier. The score counts all five tells equally (PRESENT 1, WEAK 0.5). Corrected 2026-09-18. Originally read: “Strong — commitment to a specific supplier.” The cells read as a ranking of weights; the score counts all five tells equally, so they now say they are descriptive. |
lead_time | Order-to-ship time is stretching. | MD&A language: "supply constrained", "extended lead times", "allocation". | Descriptive, not a weight: the weakest reading of the five — can be inherited from upstream, not owned. The score counts all five tells equally (PRESENT 1, WEAK 0.5). Corrected 2026-09-18. Originally read: “Weakest of the five — can be inherited from upstream, not owned.” The cells read as a ranking of weights; the score counts all five tells equally, so they now say they are descriptive. |
take_or_pay | Customer pays whether or not it takes the product. | "take-or-pay", "minimum purchase commitments", "non-cancellable". | Descriptive, not a weight: the strongest reading — the buyer has accepted the supplier's terms. The score counts all five tells equally (PRESENT 1, WEAK 0.5). Corrected 2026-09-18. Originally read: “Strongest — the buyer has accepted the supplier's terms.” The cells read as a ranking of weights; the score counts all five tells equally, so they now say they are descriptive. |
gm_pricing_power | Margin rises on the same product. | Gross margin expanding without a mix shift; price increases disclosed in MD&A. | Descriptive, not a weight: strong but lagging — shows after the others. The score counts all five tells equally (PRESENT 1, WEAK 0.5). Corrected 2026-09-18. Originally read: “Strong but lagging — shows after the others.” The cells read as a ranking of weights; the score counts all five tells equally, so they now say they are descriptive. |
What the map knows, and what it does not
Updated 2026-09-23: six cells in the map above changed after a stricter, tiered reading of what counts as a signature. Compute: NVIDIA prepayments now present (it received $2.8 billion of customer advances); Intel capacity reservation now absent (its customer deposits are counted once, under prepayments); Cerebras take-or-pay now weakly present (stated, not sized). Optical: Tower prepayments and Corning capacity reservation now absent (each deposit is counted once, in one cell). Power: Constellation capacity reservation now present (Microsoft buys the full output of a restarted plant)
What it knows
- Compute — 3 names read (NVDA, INTC, CBRS); 1 of 5 signatures PRESENT on every name read Updated 2026-09-23: the count of signatures present on every name read on this rung rose from 0 to 1 — NVIDIA's own filing shows $2.8 billion of customer advances, so prepayments now reads present on all three names (3 names); on at least one name: 5 of 5 (1 only weakly).
- Memory bandwidth — 4 names read (SKHY, SSNLF, MU, SNDK); 4 of 5 signatures PRESENT on every name read (2 names); on at least one name: 5 of 5. provisional 5 of the five cells rest on a provisional Korean pattern set.
- Optical interconnect — 6 names read (TSEM, AAOI, COHR, LITE, GLW, MRVL); 0 of 5 signatures PRESENT on every name read (6 names); on at least one name: 5 of 5 (1 only weakly).
- Substrate — 1 name read (AXTI); 3 of 5 signatures PRESENT on every name read (1 name); on at least one name: 3 of 5; 1 absent on every name read; 1 not assessable from the filings read.
- Advanced packaging — 2 names read (TSM, AMKR); 1 of 5 signatures PRESENT on every name read (2 names); on at least one name: 2 of 5; 1 absent on every name read. Corrected 2026-09-18. Originally read: “Advanced packaging — 2 names read (TSM, AMKR); 1 of 5 signatures PRESENT on every name read (2 names); on at least one name: 2 of 5; 3 absent on every name read.” The clauses counted a rung-level summary instead of every name read; one called a cell no one had read untestable.
- Power — 5 names read (CEG, GEV, VRT, MHI, BE); 0 of 5 signatures PRESENT on every name read (4 names); on at least one name: 5 of 5 (1 only weakly).
- Physical embodiment — 2 names read (MP, HDS); 3 of 5 signatures PRESENT on every name read (1 name); on at least one name: 3 of 5. Corrected 2026-09-18. Originally read: “Physical embodiment — 2 names read (MP, HDS); 3 of 5 signatures PRESENT on every name read (1 name); on at least one name: 3 of 5; 1 absent on every name read; 1 not assessable from the filings read.” The clauses counted a rung-level summary instead of every name read; one called a cell no one had read untestable.
What it does not
- Compute — names disagree on capacity_reservation, lead_time, take_or_pay, gm_pricing_power — shown as a fraction Updated 2026-09-23: prepayments left the signatures the names on this rung disagree about — it now reads present on all three, never as the union.
- Memory bandwidth — names disagree on prepayments, capacity_reservation, lead_time, take_or_pay — shown as a fraction, never as the union; for SKHY, prepayments could not be tested from the documents on file; for SSNLF, prepayments could not be tested from the documents on file. Corrected 2026-09-18. Originally read: “Memory bandwidth — names disagree on prepayments, capacity_reservation, lead_time, take_or_pay — shown as a fraction, never as the union.” The clauses counted a rung-level summary instead of every name read; one called a cell no one had read untestable.
- Optical interconnect — names disagree on prepayments, capacity_reservation, lead_time, take_or_pay, gm_pricing_power — shown as a fraction, never as the union.
- Substrate — one name is not the rung — the cells say what that filer shows, not what every vendor on the layer shows; lead_time could not be tested from the documents on file.
- Advanced packaging — names disagree on capacity_reservation — shown as a fraction, never as the union; for TSM, lead_time, gm_pricing_power could not be tested from the documents on file. Corrected 2026-09-18. Originally read: “Advanced packaging — names disagree on capacity_reservation — shown as a fraction, never as the union.” The clauses counted a rung-level summary instead of every name read; one called a cell no one had read untestable.
- Power — names disagree on prepayments, capacity_reservation, lead_time, take_or_pay, gm_pricing_power — shown as a fraction, never as the union; for MHI, capacity_reservation, lead_time, take_or_pay, gm_pricing_power could not be tested from the documents on file. Corrected 2026-09-18. Originally read: “Power — names disagree on prepayments, capacity_reservation, lead_time, take_or_pay, gm_pricing_power — shown as a fraction, never as the union.” The clauses counted a rung-level summary instead of every name read; one called a cell no one had read untestable.
- Physical embodiment — names disagree on prepayments, capacity_reservation — shown as a fraction, never as the union; for HDS, prepayments, capacity_reservation, lead_time, take_or_pay could not be tested from the documents on file; for MP, lead_time has never been looked at. Corrected 2026-09-18. Originally read: “Physical embodiment — names disagree on prepayments, capacity_reservation — shown as a fraction, never as the union; lead_time could not be tested from the documents on file.” The clauses counted a rung-level summary instead of every name read; one called a cell no one had read untestable.
Hub and spokes
One live constraint chain; candidate chains wait behind four graduation gates (G1 chokepoint · G2 pure-play · G3 credible voice · G4 testable in filings).
Corrected 2026-09-18. Originally read: “analysis core — six factors, weakest link binds”. Six was the first rubric's factor count; the rubric in force binds on five, and the count is now read from it.
A row is a physical layer AI scaling passes through. A spoke is a separate constraint chain with its own chokepoint, voices, and ladder, sharing one core and one book. — this caption is a stored literal, not read from the record.
Drawn from core/ladder.json, data/architecture-nodes.json and the gate files at build; ladder generated 2026-09-07 08:54:42.
Superseded updates
SUPERSEDED 2026-09-23 — this update no longer describes the page: the value it reported changed again, and a newer update beside that value says so. It is kept here because a published note is never removed.
Updated 2026-09-20: the count of signatures present on every name read on this rung fell from 1 to 0 — the one signature all three names shared was gross-margin pricing power, and on one of them it is now absent
SUPERSEDED 2026-09-23 — this update no longer describes the page: the value it reported changed again, and a newer update beside that value says so. It is kept here because a published note is never removed.
Updated 2026-09-20: gross-margin pricing power joined the signatures the names on this rung disagree about — it used to read the same on all three
Retired corrections
RETIRED 2026-09-20 — this correction is no longer live: the text it corrected no longer appears on this page. It is kept here because a published correction is never removed.
Corrected 2026-09-18. Originally read: “Compute — 3 names read (NVDA, INTC, CBRS); 1 of 5 signatures PRESENT on every name read (3 names); on at least one name: 5 of 5; 1 absent on every name read.” The clauses counted a rung-level summary instead of every name read; one called a cell no one had read untestable.