Photonics & Chip-Design Research Brief — 2026-07-07 (v3)

Stage 2 run against bundle/2026-07-07_bundle.json (6 articles, 6 tickers). First run under the codified one-voice corroboration rule (ANALYSIS-PROMPT.md, committed 8d35c25). Supersedes briefs/2026-07-07_brief_v2.md. Inputs vs v2: +4 new Damnang posts (07-02 Meta-compute, 07-04 VCSEL-CPO, 07-05 CXL, 07-07 NVIDIA-racks); same six [F]/[M]/[C] on disk (EDGAR/valuation _2026-07-07); same bottleneck/2026-07-06_map.json; same ledger prior. Analysis run by hand.*

Headline result: net conviction change ZERO vs v2 — NVDA 4 · COHR 4 · LITE 3 · AXTI 3 · MU 3 · AEHR 2 (all held). But this is NOT a null run: the one-voice rule legitimately raised corroboration for the two InP names Damnang newly endorsed (COHR, LITE 6.5→7.5) and modestly for MU (5.0→6.0), while refusing to let four Damnang posts inflate NVDA. Every name remains bound below its improved corroboration by catalyst timing or fundamentals, so no conviction moved — the correct weakest-link outcome.

1. EXECUTIVE SUMMARY (~300w)

Four new Damnang articles arrived this run; all four are one voice. The rubric's newly codified one-voice rule did exactly its job: it let Damnang initiate as a second independent voice on COHR and LITE — names he had never covered before, now explicitly his "most solid position" in the light-source war [A4] — lifting their corroboration 6.5→7.5, while treating his other three posts as recency/depth on already-counted theses, not as fresh confirmations. NVDA, already two-voice, held at 6.5 despite two new Damnang pieces reinforcing it. Post-count did not masquerade as corroboration.

None of that moved conviction, because every name is capped below its corroboration score by a harder-binding factor: NVDA and MU by absent dated catalysts, COHR and LITE by catalyst timing, AXTI and AEHR by loss-making fundamentals. The map can only cap, never lift; corroboration here sits above the binding factor in all six. So the scorecard holds — and holding is the honest answer.

The analytically live material is in the falsifier confrontations. Damnang's VCSEL-CPO piece [A4] is the first serious challenge to the InP-substrate thesis, yet his own verdict — displacement odds "very low" within 2028, coexistence not replacement — strengthens AXTI's near-term CRITICAL chokepoint rather than starting its falsifier clock. His CXL piece [A5] answers the natural MU bear — "does HBM roll over?" — with a "clear no," reaffirming the HBM chokepoint while surfacing a different, subtler DRAM-cycle risk (pooling + DDR4 recycling flatten conventional-DRAM growth and deepen the next trough). And his Meta-compute piece [A3] independently reaches PhotonCap's tents conclusion [A2] — the July-1 "surplus/peak" selloff was a misread; Meta's buildout lifts, not caps, the demand that flows to memory and optics. Genuine convergence — but between two already-counted voices, so it adds zero to any independence count. The non-consensus read stands: the profitable incumbent COHR still outranks the pure-plays the authors most love.

2. TOP 10 TAKEAWAYS

1. One-voice rule, first live run: 4 Damnang posts + 0 new PhotonCap = one voice. It raised COHR/LITE (Damnang newly names them) but not NVDA (already counted). Zero conviction moved — corroboration sits above the binding factor everywhere.
2. Damnang crossed the aisle onto the InP names. VCSEL-CPO [A4] names Lumentum and Coherent as the "most solid position, holds both InP and VCSEL." COHR/LITE go from PhotonCap-only to genuinely two-voice.
3. The first real InP-thesis challenge resolves in the thesis's favor. VCSEL is a different-architecture (GaAs) bypass, not an in-production non-InP substrate path; Damnang rates displacement within 2028 "very low." AXTI's CRITICAL chokepoint is affirmed, not falsified — [A4] calls InP CW "a disaster-level bottleneck right now."
4. AXTI +12.2% on the day [M] into its 2026-07-30 AXT-earnings catalyst — the only dated catalyst in the bundle. Conviction still 3: loss-making fundamentals (Fu 3) override the bundle's highest structural exposure (Bn 9). Structural rank ≠ investability.
5. CXL does not threaten HBM. Damnang [A5]: "clear no" — HBM and CXL grow together (division of labor). MU's HBM-chokepoint basis holds; its Bn falsifier is not tripped.
6. But a new MU risk appears where the map wasn't looking: CXL pooling reclaims stranded memory and Vistara-style DDR4 recycling substitutes for new low-end DRAM, flattening conventional-DRAM growth and, per Damnang, building "optimistic bias" into next-downcycle-trough estimates. Sharpens the peak-cycle-trap valuation ceiling (3).
7. MU's own author now argues a structurally higher-floored cycle (take-or-pay RPO ~$100B, HBM die-intensity) [A3][A5] — a credible counter to the map's "cyclical-not-structural" flag, flagged for a future map revision, not applied now.
8. Cross-author convergence on Meta: Damnang [A3] and PhotonCap [A2] independently reject the July-1 peak read and affirm Meta's buildout as accelerating demand into memory + optics. Real convergence; adds nothing to the count (both voices pre-counted).
9. MU HBM4 supplier risk retired: SemiAnalysis had put Micron's Rubin HBM4 share at ~zero; Micron was since certified alongside SK hynix/Samsung with volume shipments [A6].
10. NVDA roadmap-scare deflated [A6]: the SemiAnalysis "Kyber MASSIVE DELAY" reads as a rack-mix (Kyber↔Oberon) adjustment, not a generation slip. Watch Taiwan monthly PCB/CCL revenue (2H-2026) for the real Kyber-vs-Oberon signal. NVDA held at 4.

3. CORROBORATION — NAMED-VOICES CHAINS (one-voice rule live)

Two independent voices exist this run: Damnang (posts [A1] 07-01, [A3] 07-02, [A4] 07-04, [A5] 07-05, [A6] 07-07 — ONE voice) and PhotonCap ([A2] 07-06 — no new post this run; already in the prior). Forgetting ≈ 0 for all (prior dated 2026-07-07; same-day re-score). Chains below name every voice behind each update.

None of these updates bind. In all six, another factor (catalyst or fundamentals) already sits at or below the corroboration score, so raising corroboration changed no Stage-1 minimum. The point of recording them is the ledger/prior for future runs.

4. FALSIFIER CONFRONTATIONS (requested)

(a) VCSEL-CPO [A4] vs the AXTI map falsifier ("a viable non-InP substrate path for SiPh reaches production")

Verdict: clock does NOT start; the chokepoint is strengthened.

• VCSEL CPO is a different-architecture bypass (GaAs-substrate laser that emits and modulates in one device), not a non-InP substrate for silicon photonics. SiPh in CPO still requires an InP CW light source [A4]. The falsifier is worded for a substrate swap within SiPh; VCSEL doesn't meet it.
• Nor is it "in production" for CPO: the industry has retreated to NPO (near-packaged), blocked by dense-packaging yield and the thermal wall (a directly-modulated VCSEL can't use InP's external-light-source heat-separation trick) [A4].
• Damnang's own conclusion reinforces CRITICAL/worsening: InP CW is "a disaster-level bottleneck right now," the NVIDIA roadmap is "locked to InP" through Feynman (2028), and VCSEL-displacement odds within 2028 are "very low." He even frames the InP shortage as the catalyst that slowly lets VCSELs in — i.e., the chokepoint is real and biting.
• [I] reassessment flag — raised, NOT applied: the map's absolute phrasing "no route-around at the substrate layer" now has a long-tail crack — a conceptual architecture-level route-around (VCSEL for scale-up short links) that gains ground "the deeper the shortage runs," beyond the 2027 falsifier horizon. Worth softening the map's absolutism in a future revision; AXTI Bn held at 9 this run (severity/trend/horizon all unchanged inside the falsifier window). [I: VCSEL becomes a genuine substrate-layer route-around — falsifier: a hyperscaler adopts VCSEL arrays for production scale-up CPO, or dense-packaging/heat-separation yield is demonstrated at volume, by end-2028.]

(b) CXL [A5] vs MU's "cyclical-not-structural" flag / HBM falsifier

Verdict: HBM falsifier NOT tripped; the cyclical flag is reinforced by a different mechanism, and separately contested by Damnang himself.

• MU's map falsifier requires "HBM demand-intensity eases — supply catches demand." [A5] says the opposite: "[… passage withheld: paid source …]." HBM and CXL are complementary tiers; "an HBM shortage and a CXL demand surge happening simultaneously… is itself the evidence." Chokepoint holds.
• New, subtler risk the map didn't price: CXL pooling reclaims stranded memory and Vistara DDR4 recycling substitutes for new low-end DRAM — Damnang: this "flattens the slope of growth… and deepens the trough when the cycle turns," building "optimistic bias" into downcycle-bottom estimates [A5]. This reinforces the −2 cyclical discount on MU's Bn (via conventional DRAM, not HBM) and sharpens the peak-cycle valuation trap (ceiling 3).
• Counter-tension (honest): the same author argues the cycle is structurally higher-floored — ~$100B take-or-pay RPO at guaranteed minimum prices, HBM consuming far more die per wafer — so "the odds the drawdown is the same halving cycle as before are low" [A3][A5]. That partially rebuts the map's "purely cyclical" framing.
• Net: MU Bn held at 6 (HIGH·worsening 8, −2 cyclical). The flag stands (Damnang concedes "the peak debate will come," guides slowing price increases), but I raise an [I] reassessment note — NOT applied: if the higher-floor holds through the next downcycle, soften MU's −2 discount. [I: MU's cycle is structurally higher-floored, not a classic halving — falsifier: the next memory downcycle prints a trough >prior-cycle bottom with take-or-pay volumes honored; check FY2028 as new-fab shipments begin.]

(c) Meta-compute [A3] as cross-author agreement with PhotonCap's tents thesis [A2]

Verdict: genuine convergence — but +0 to the independence count.

• Substantive agreement: PhotonCap [A2] — the more Meta pitches "tents" (rapid buildout), the faster InP runs dry (demand accelerates). Damnang [A3] — the July-1 selloff misread Meta's cloud/"surplus" move; compute is being rationed not oversupplied (Google rationing Gemini; SpaceX bridge capacity at $920M/mo), and Meta selling surplus lifts the demand ceiling — "not one optical module" of demand disappears. Two authors, two angles (InP-supply vs supply-chain-financing), one conclusion: Meta's buildout is accelerating, not peaking, infrastructure demand into memory + optics.
• Count impact = zero, per the one-voice rule. Damnang (author of [A3]) is already counted as this run's Damnang voice; PhotonCap ([A2]) is already counted in the prior. The agreement is between two already-counted voices, so it creates no third independent confirmation. It is recorded as a Bigger-Picture convergence (§6), not as corroboration arithmetic. (Note the sequence: Damnang's [A3] 07-02 predates PhotonCap's [A2] 07-06 — independent arrival at the same read, which is why it's a real convergence and not an echo; it still doesn't add to the count.)

5. RECOMMENDATIONS — full conviction breakdowns

Scorecard (Stage-1 Quality = min of applicable factors; Stage-2 = min(Quality, Val ceiling)):

#1 — NVDA (Conviction 4) — "Rack-scale scale-up is a production-proven monopoly the roadmap scares keep mispricing."

• Thesis 8 — [A6] decodes the scale-up moat (NVLink/NVSwitch, copper-reach → rack density) and shows the only production-proven rack-scale scale-up since late-2024 is NVIDIA's; [A3] adds the demand-not-peaking supply-chain read. Mechanism-specific.
• Corroboration 6.5 — two voices (Damnang [A1][A3][A6] + PhotonCap [A2]); +0 this run (both pre-counted; see §3).
• Fundamental 8 — [F] TTM revenue $253.49B, net income $159.61B (as of 2026-04-26). Carry-over earnings-quality note: prior-run Q1 FY2027 NI > operating income (~$4.8B non-operating) — monitor, not disqualifying.
• Catalyst 4 — no dated in-bundle NVDA event. The datable signal is Taiwan monthly PCB/CCL revenue (~10th monthly, 2H-2026) for Kyber-vs-Oberon mix [A6] — a read-through, not an NVDA print.
• Bottleneck 4 — map: MODERATE·easing(self-mitigated)·#5 → 4. Co-binds with catalyst. NVDA has partly bought its way out of the InP/HBM chokepoint (COHR/LITE stakes [A2]; InP-locked roadmap [A4]).
• Valuation ceiling 6 — [C] P/E 29.7 (= $4.74T [M] ÷ $159.61B NI [C←F]); [C] fwd P/E 21.8, fwd PEG 0.25 — the bundle's only clean (non-cyclical) forward PEG.
• Stage 1 = min(8, 6.5, 8, 4, 4) = 4 (Catalyst & Bottleneck co-bind) → Stage 2 = min(4, 6) = 4.
• Most important catalyst: Taiwan 2H-2026 monthly board-supply revenue (Kyber ramp evidence) [A6]. [I: the Kyber "delay" is a rack-mix shift, not a generation slip — falsifier: Rubin Ultra chip-generation slips out of 2H-2027, or Taiwan high-layer PCB/CCL orders show no Kyber ramp through 2H-2026.]

#2 — COHR (Conviction 4) — "Top of the InP stack: cheapest optical multiple, only profitable name, now two-voice."

• Thesis 8 — PhotonCap: first 6-inch InP fab (Sherman 2026-27), the capacity relief valve [A2]. Damnang now adds the light-source-war-agnostic angle: Coherent makes EML, CW and VCSEL and has demonstrated 1.6T across all three [A4] — robust to who wins.
• Corroboration 7.5 — Damnang joins as 2nd voice this run (+1.0) [A4]; PhotonCap pre-counted. Genuine cross-author lift (§3, Ruling #3).
• Fundamental 7 — [F] TTM revenue $6.602B, net income $468.85M (as of 2026-03-31), back to GAAP profit. Docked from 8: no in-bundle gross profit / stale operating income → margin quality unverifiable; P/E rests on net income alone.
• Catalyst 4 — no dated in-bundle catalyst (Sherman ramp is 2026-27, undated).
• Bottleneck 8 — map: HIGH·worsening(owns relief valve)·#2 → 8. Doesn't bind.
• Valuation ceiling 5 — cheapest optical name on sales: [C] P/S 8.06 (= $53.23B [M] ÷ $6.602B rev [C←F]) vs LITE 22.9; [C] P/E 113.5. Forward PEG still n/a — no TIKR estimate (the one profitable name lacking a forward denominator; per §22 the highest-value data-completeness gap, though catalyst-capped it can't lift COHR above 4 alone).
• Stage 1 = min(8, 7.5, 7, 4, 8) = 4 (Catalyst) → Stage 2 = min(4, 5) = 4.
• Most important catalyst: first Sherman 6-inch InP yield/qualification disclosure (2026-27) [A2]. [I: COHR's 6-inch first-mover lead is the InP relief valve and a pricing moat — falsifier: Sherman slips materially past 2027 or a rival brings equivalent 6-inch InP capacity first.]

#3 — LITE (Conviction 3) — "Same InP leverage as COHR, now dual-source-validated, but capped by a re-rated multiple."

• Thesis 8 (↑ from 7) — PhotonCap: >½ EML share, capacity tight to mid-2028 [A2]. Ruling #10: bumped 7→8 — Damnang [A4] adds a new mechanism dimension (Lumentum holds both InP and VCSEL → light-source-war resilience), a more robust thesis than single-source InP. Does not bind.
• Corroboration 7.5 — Damnang joins as 2nd voice this run (+1.0) [A4]; PhotonCap + [F] inflation pre-counted.
• Fundamental 7 — [F] TTM revenue $2.488B, net income $439.9M (as of 2026-03-28); thin trailing profit base historically (docked from 8).
• Catalyst 4 — no dated in-bundle catalyst; near-term 6-inch lever lags COHR (Greensboro mid-2028) [A2].
• Bottleneck 7 — map: HIGH·worsening·#3 → 8, −1 (relief lags to mid-2028) = 7. Doesn't bind.
• Valuation ceiling 3 — already re-rated: [C] P/E 129.3 (= $56.89B [M] ÷ $439.9M NI [C←F]), P/S 22.9, fwd PEG 0.30 UNRELIABLE — cyclical (recovery snapback, not secular).
• Stage 1 = min(8, 7.5, 7, 4, 7) = 4 (Catalyst) → Stage 2 = min(4, 3) = 3.
• Most important catalyst: Greensboro / EML-capacity disclosures through 2027-28 [A2]. [I: LITE's EML tightness is real but its relief lags COHR by ~2 years — falsifier: EML supply catches demand before mid-2028, or LITE's EML share falls below ~50%.]

#4 — AXTI (Conviction 3) — "The bundle's highest structural exposure and its only dated catalyst — but the weakest-link guard still binds on losses."

• Thesis 7 — PhotonCap: ~80% InP substrate share (AXT+Sumitomo duopoly), crystal-growth moat, furthest upstream, "no route-around" [A2]. VCSEL-CPO [A4] tests but does not break this (see §4a).
• Corroboration 6.0 — held (+0); full new-voice credit declined — Damnang affirms the substrate-shortage mechanism but doesn't name AXTI, routing InP exposure to IQE/AAOI/Sivers [A4] (§3, Ruling #5). Doesn't bind regardless.
• Fundamental 3 — [F] TTM revenue $95.89M, net income −$14.08M (as of 2026-03-31); sub-peak revenue (FY2022 $141M → FY2025 $88M); sharpest China export-license tail (produces in China). Binds.
• Catalyst 7 — owns the bundle's only dated catalyst: AXT Q2 results 2026-07-30 [A2] (23 days out). Note [M]: AXTI +12.2% on the day into it.
• Bottleneck 9 — map: CRITICAL·worsening·#1 → 9, the highest structural exposure in the bundle. Enters the min but cannot lift; Fu 3 overrides. Structural rank ≠ investability.
• Valuation ceiling 3 — [C] P/S 30.7 (= $2.945B [M] ÷ $95.89M rev [C←F]); P/E n/m (negative earnings). Forward PEG n/a (no TIKR estimate).
• Stage 1 = min(7, 6.0, 3, 7, 9) = 3 (Fundamental) → Stage 2 = min(3, 3) = 3.
• Most important catalyst: AXT Q2 earnings, 2026-07-30 [A2]. [I: structural primacy ≠ investability until AXTI converts backlog to profit — falsifier: the 2026-07-30 print shows backlog→revenue conversion + a resolved China export-license path, lifting Fu off 3.]

#5 — MU (Conviction 3) — "Real HBM supercycle, real value trap; the cheap multiple is the danger, not the opportunity."

• Thesis 6 — HBM demand-intensity + $100B take-or-pay RPO + HBM4 certification [A3][A6]; structural-vs-CXL complementarity [A5]. Not a named InP-stack thesis, hence 6.
• Corroboration 6.0 — Damnang initiates by name this run (+1.0) [A3][A5][A6]; no PhotonCap memory coverage. Tempered by his own cycle caveats (§3, Ruling #4). Doesn't bind.
• Fundamental 8 — [F] Q3 FY2026 revenue $41.456B (+346% YoY), GM 84.6%, TTM revenue $90.27B, NI $50.47B (accession 0000723125-26-000015). Advisory anomaly-flag retained (verified real in the 10-Q, §18/§19); flows into scoring.
• Catalyst 3 — no dated in-bundle catalyst; next-quarter guide (~$50B rev, ~86% GM) is [A3]-reported, not an in-bundle dated event. Binds.
• Bottleneck 6 — map: HIGH·worsening·#4 → 8, −2 (cyclical-not-structural) = 6. Flag reinforced by CXL/recycling risk and contested by higher-floor argument (§4b); net held.
• Valuation ceiling 3 — [C] P/E 22.0 (= $1.11T [M] ÷ $50.47B NI [C←F]) and fwd PEG 0.02 UNRELIABLE — cyclical peak: cheap on peak ~85% margins that normalize toward ~40%. The CXL trough-deepening risk [A5] sharpens the trap.
• Stage 1 = min(6, 6.0, 8, 3, 6) = 3 (Catalyst) → Stage 2 = min(3, 3) = 3.
• Most important catalyst: next 10-Q for pricing/lead-time normalization signals (undated in-bundle). [I: MU's 22x P/E and 0.02 PEG are a peak-cycle trap, not cheap growth — falsifier: a filing shows take-or-pay volumes honored through a downcycle with the trough above the prior-cycle bottom, converting "cyclical" to "structurally higher-floored."]

(6th, not recommended) — AEHR (Conviction 2): Th 7 · Cor 6.0 · Fu 2 (binds) · Ca 4 · Bn 4 · Val 2. [F] revenue −43.7% YoY (Q3 FY2026 $10.31M vs $18.31M), operating losses all FY2026 quarters, most expensive on sales (P/S 50.3). Bull case rests on one [A2] order, not the trailing numbers; Damnang doesn't name AEHR this run.

6. BIGGER PICTURE — convergences the writers didn't connect

• The Meta-compute demand read is now two-author (§4c) — but the corollary each author half-states is the real signal: if Meta's buildout is accelerating (not peaking), and the scale-up TAM is ~9× scale-out and doubling again in Rubin [A4], then the demand flows to whichever layer is the binding chokepoint. Damnang's own supply-chain map [A3] and his light-source map [A4] both point to the same place PhotonCap does [A2]: the InP/optical layer, not the GPU. The convergence isn't "AI is big" — it's that three of the six articles independently locate the value in the interconnect/substrate chokepoint.
• Damnang's "hold both light sources" thesis [A4] is a COHR/LITE quality argument the bottleneck map doesn't capture. The map scores COHR/LITE on InP-chokepoint severity. But [A4]'s point is orthogonal and additive: they're the two names that win regardless of the InP-vs-VCSEL outcome. That's a durability/optionality argument for the valuation-ceiling and thesis factors, and it's the strongest single reason the profitable incumbent COHR keeps outranking the pure-play substrate names.
• The memory story has two clocks running in opposite directions. HBM (structural, complementary to CXL, higher-floored) and conventional DRAM (structurally pressured by pooling + recycling). MU straddles both. The market's single "MU is cheap at 22x" read collapses these; the honest read is HBM-bull / commodity-DRAM-cautious, which is exactly why the peak-cycle-trap ceiling (3) survives the supercycle print.
• Notable people / signals: Ashkan Seyedi (ex-NVIDIA SiPh → ams OSRAM VCSEL) [A4] — the hire that "raises the question" without proving displacement. Watch-signals that are genuinely datable: AXT Q2 2026-07-30 [A2]; Taiwan monthly PCB/CCL revenue 2H-2026 [A6]; Marvell Structera S switch sampling Q3-2026 [A5].

7. UNVERIFIED / ASSERTED-BUT-UNGROUNDED

• InP/EML/substrate share figures (~80% AXT+Sumitomo; LITE >½ EML) are [A]-asserted (Reuters via [A2]), not [F]-verified. Carried as thesis color, not fundamentals.
• Forward EPS/growth for COHR, AXTI, AEHR — no TIKR estimate on disk; forward P/E and PEG are honestly n/a (not fabricated). COHR is the material gap (only profitable name without a forward denominator).
• MU next-quarter guide (~$50B rev, ~86% GM) and $100B take-or-pay RPO [A3] are author-reported, not [F] in this bundle. Directionally corroborated by the [F] Q3 print but not independently filed here.
• SemiAnalysis Kyber-delay specifics and NVDA's one-line rebuttal [A6] are reported/asserted; the resolving evidence (Taiwan monthly board revenue) is future.
• AEHR anchor order / ramp [A2] is [A]-only; [F] revenue trend contradicts it.

8. SOURCES

Authors [A] — voice count: Damnang = ONE voice ([A1],[A3],[A4],[A5],[A6]); PhotonCap = ONE voice ([A2]).

• [A1] Damnang, "The Inference Chip Architecture Map," 2026-07-01.
• [A2] PhotonCap, "The More Tents Meta Pitches, the Faster InP Runs Dry," 2026-07-06.
• [A3] Damnang, "Meta Compute Misread: Where the AI Infrastructure Opportunity Really Is," 2026-07-02.
• [A4] Damnang, "Can VCSEL CPO Shake Up the InP Optical Order?," 2026-07-04.
• [A5] Damnang, "The Memory Crisis: Can CXL Be the Answer?," 2026-07-05.
• [A6] Damnang, "A Complete Guide to NVIDIA's Racks, and Then SemiAnalysis," 2026-07-07.

Filings [F] (EDGAR, edgar/*_2026-07-07.json; TTM as-of dates from [C] roll):

• [F-NVDA] TTM rev $253.49B / NI $159.61B (2026-04-26).
• [F-COHR] TTM rev $6.602B / NI $468.85M (2026-03-31).
• [F-LITE] TTM rev $2.488B / NI $439.9M (2026-03-28).
• [F-AXTI] TTM rev $95.89M / NI −$14.08M (2026-03-31).
• [F-MU] Q3 FY2026 rev $41.456B (+346% YoY), GM 84.6%; TTM rev $90.27B / NI $50.47B (2026-05-28); accession 0000723125-26-000015 (advisory anomaly-flag, verified real).
• [F-AEHR] Q3 FY2026 rev $10.31M vs $18.31M YoY (−43.7%); TTM rev $45.26M / NI −$11.42M (2026-02-27).

Market [M] (valuation/*_2026-07-07.json, ts 2026-07-07):

• NVDA $195.55 / mktcap $4.736T. COHR $335.70 / $53.23B. LITE $731.25 / $56.89B. AXTI $63.52 / $2.945B (day +12.2%). MU $984.75 / $1.112T. AEHR $72.325 / $2.275B.

Calculated [C] (formulas cite [M] and [C←F]): P/E, P/S per name as shown in §5; forward P/E and PEG carry [LOW-CONFIDENCE forward][NEAR-TERM]; LITE/MU forward PEG carry [UNRELIABLE — cyclical]. Bottleneck [I]-grade structure from bottleneck/2026-07-06_map.json (gen 2026-07-06) — used for factor 5 only; no figure lifted into the brief.

Methodology: ANALYSIS-PROMPT.md (one-voice rule live, commit 8d35c25); bundle bundle/2026-07-07_bundle.json; ledger prior ledger/conviction-ledger.json (lastUpdated 2026-07-07). Supersedes briefs/2026-07-07_brief_v2.md.

9. NUMBERED RULINGS LOG (continuing; #1 = the one-voice rule itself)

1. One-voice rule codified (prior turn, commit 8d35c25): same-author articles = one corroborating voice; only cross-author agreement or an [F] filing lifts the count.
2. Damnang = ONE voice despite 5 posts. NVDA already two-voice → 2 new Damnang posts add recency/depth only; corroboration held 6.5. Post-count ≠ corroboration.
3. Damnang initiates as 2nd voice on COHR and LITE (previously PhotonCap-only; [A4] names both as "most solid position"). Genuine cross-author lift: +1.0 each → 7.5. (One voice added, consistent with #1.) Non-binding.
4. MU corroboration +1.0 → 6.0. Damnang now engages MU by name ([A3] "clearest misjudgment"/$100B RPO; [A6] HBM4 certified; [A5] HBM-structural). Tempered by his own peak-margin/cycle caveats. Non-binding (catalyst 3).
5. AXTI full new-voice credit DECLINED; held 6.0. Damnang affirms the InP-substrate mechanism [A4] but does not name AXTI and routes InP exposure to IQE/AAOI/Sivers. A case for +0.5 (mechanism-level) is noted and not taken. Non-binding regardless (Fu 3).
6. AEHR corroboration held 6.0. Damnang does not name AEHR; no new voice.
7. Falsifier (a): AXTI clock does NOT start; chokepoint strengthened. VCSEL is a GaAs-architecture bypass, not an in-production non-InP SiPh substrate. Bn held 9; long-tail route-around raised as an [I] flag, NOT applied.
8. Falsifier (b): MU HBM falsifier NOT tripped ("clear no" HBM roll-over). Cyclical flag reinforced via CXL/recycling and contested via higher-floor argument. Bn held 6; higher-floor softening raised as an [I] flag, NOT applied.
9. Convergence (c): Meta-compute ↔ tents is genuine but +0 to the count — both voices pre-counted. Recorded as a Bigger-Picture convergence, not corroboration.
10. LITE thesisStrength 7→8 — Damnang adds the dual InP+VCSEL light-source-war resilience mechanism [A4]. Non-binding (catalyst 4).
11. No silent map edits. All six Bn scores held at v2 values; the two [I] reassessment flags (#7, #8) are surfaced for a future map revision, explicitly not applied this run.

Net conviction vs v2: 0 across all six. Ledger-relevant sub-score deltas: COHR Cor 6.5→7.5, LITE Cor 6.5→7.5 & Th 7→8, MU Cor 5.0→6.0.