The Bottleneck Playbook: what was observable before memory, Nvidia, semis and shipping exploded — and where the signature points now

Supply-Chain Bottleneck Intelligence — Late October 2026

Forensic timing edition: what was observable, when, before the stocks moved

Date context: 26 Oct 2026. Method rule applied throughout: an indicator is only credited at the date it was publicly observable — transcript, auction result, trade-press price series, or filing available that day. Retrospective aggregations are flagged [retrospective] and never counted as leading evidence. Analyst speculation is flagged [speculation].

Confidence: High for dated transcript/auction/price-series facts cited below; medium where a fact rests on a single trade-press aggregator; low where flagged.

EXECUTIVE SUMMARY

1. The repeatable pattern is not “demand is strong.” It is: (i) a demand shock hits a supply base that cut capacity 12–24 months earlier, (ii) the binding constraint turns out to be a step nobody headlines (packaging, film, glass cloth, transformers), (iii) contract/spot prices turn before earnings, (iv) management language escalates in a fixed order — discipline → tight → allocated → sold out → take-or-pay/prepayment — and (v) the stock’s explosive move usually starts at the pricing + sold-out stage, not at the demand headline. In memory, the demand headline (ChatGPT, Nov 2022) preceded the memory re-rating by ~2½ years; anyone who bought the demand story in memory in 2023 sat through a loss-making year.
2. Shortest reliable lead times came from prices, not transcripts. TrendForce contract prices turned in Q4 2023, ~15–20 months before Micron’s April 2025 trough-to-peak run and ~21 months before the Q4 2025 +45–95% quarterly price explosions. GPU rental/lead-time signals preceded Nvidia’s 24 May 2023 guidance shock by ~3–5 months, but the stock had already risen +109% YTD before that print (SiliconANGLE) — the print was confirmation, not discovery.
3. Energy (Part 3B) is the strongest current analogue — and it is NOT early for the obvious tickers. PJM capacity prices did exactly what DRAM contract prices did: $28.92 → $269.92/MW-day in one auction (Jul 2024), then cleared at the regulatory cap in the next three auctions ($329.17, $333.44, $325.00), with the last two auctions short of PJM’s reliability target for the first time in its history. Transformer lead times are 128 weeks, gas-turbine slots are being sold for 2032, switchgear is 50–104 weeks. Yet GEV is already +37% over 1 yr after a much larger 2024–25 run and is –32% off its peak, CEG is –31% over 1 yr, VST –27%, ETN –2%, VRT +6% — i.e. the equipment bottleneck is mid-stage and the merchant-power trade is late/post-peak. The least re-rated energy-linked names in our data are the regulated utilities signing the load (AEP, Dominion, Duke) — but their upside is rate-base arithmetic (5–9% EPS growth guided), not bottleneck pricing power, because regulators cap it. Verdict: energy shortage = REAL and FIRED outright on 7 of 10 playbook indicators (two partial, one split); investable bottleneck stage = MID for grid equipment, LATE for merchant nuclear/power, EARLY-MID only in second-order consumables (ABF film, T-glass) that feed all of the above.

PART 1 — ANATOMY OF THE MEMORY BOTTLENECK (TEMPLATE CASE)

1a. What became scarce, and why

• HBM first, then everything. HBM (stacked DRAM on a CoWoS interposer beside the GPU) was the original scarce item; its production cannibalises conventional DRAM: Micron CEO Sanjay Mehrotra, Q4 FY2024 call, 25 Sep 2024: the constraint is “lower wafer capacity compared to the peak of 2022, as well as the HBM 3-to-1 trade ratio” — roughly 3 wafers of conventional capacity consumed per wafer of HBM — “keeping the industry in a tight supply, and tight supply, not just for HBM, but also for non-HBM” and management remaining “extremely disciplined about CapEx.” Source: https://www.MarketBeat.com/earnings/reports/2024-9-25-micron-technology-inc-stock/
• Capacity discipline preceded the shock. Micron Q4 FY2023 call (Sep 2023): “total wafer start reductions in both DRAM and NAND are approaching 30% versus peak 2022 levels… our DRAM and NAND wafer starts will remain significantly below 2022 levels for the foreseeable future… Lead times to increase this wafer capacity will be long.” Same call: “The ramp of HBM3 and 3E production will reduce overall DRAM bit supply growth industry-wide with particular supply impact on non-HBM products.” Source: https://strike.market/stocks/MU/earnings-call-transcripts/226225
• The cuts began even earlier. Micron Q1 FY2023: FY23 capex cut to $7–7.5B from $12B in FY22 (~40%), WFE down >50% YoY, wafer starts cut ~20%. Source: https://www.alphaspread.com/security/nasdaq/mu/investor-relations/earnings-call/q1-2023
• NAND/HDD followed via the same wafer reallocation plus AI storage demand, culminating in sold-out NAND (see §1c) and sold-out nearline HDD — a product most investors had written off.
• Advanced packaging was a co-constraint throughout: every HBM stack must pass TSMC CoWoS (see Part 1B, Episode 2, and Part 3, CoWoS).

1b. Who controls supply — concentration

Source: https://en.sedaily.com/finance/2025/12/19/samsung-reclaims-no-2-in-hbm-narrows-dram-gap-with-sk-hynix — Note: shares move quarter to quarter (SK hynix HBM was 64% in Q2 2025 in the same series); the structural fact is the 3-supplier HBM/DRAM oligopoly, not any single quarter.

1c. Earliest observable signals — verbatim, dated (full forensic timeline in Part 1B)

Escalation ladder, in the order it actually appeared:

1. Dec 2022–Sep 2023 — Discipline: capex –40%, wafer starts –30% vs peak (above).
2. Q4 2023 — Prices turn: TrendForce: DRAM contract prices rose in Q4 2023 and were forecast +13–18% in Q1 2024, NAND +15–20%; enterprise SSD contract prices +>15% in Q4 2023, forecast +>25% in Q1 2024, eSSD industry revenue +47.6% QoQ in Q4 2023. Sources: https://hkhdelectronics.com/ic-storage-dram-news/ ; https://DRAMexchange.com/WeeklyResearch/Post/2/11930.html
3. 18 Dec 2024 — First clean “sold out”: Micron FQ1 2025 prepared remarks: “our HBM is sold out for calendar 2025, with pricing already determined for this time frame” and HBM TAM raised to “exceed $30 billion in 2025.” Source (transcript PDF): https://files.quartr.com/reports/cf188cb63db425a18ed5ffdf6a982546-2026-09-30-20-39-40.pdf?ref=U0E=
4. 29 Oct 2025 — SK hynix: 2026 HBM supply negotiations concluded; on the Q3 call the company said HBM has “remained sold out since 2023” and supply will stay tight in 2027; all DRAM+NAND output for 2026 effectively secured/sold out. Sources: https://threadreaderapp.com/thread/1941493202873352400.html ; https://www.blocksandfiles.com/ai-ml/2025/10/30/hbm-gold-mine-sends-sk-hynix-revenues-to-a-new-high-point/1587272
5. Jan 2026 — HDD joins: Western Digital CEO Irving Tan, Q2 FY26 call: “We’re pretty much sold out for calendar 2026. We have firm POs with our top seven customers” plus LTAs for 2027 (two customers) and 2028 (one). Source: https://www.tomshardware.com/pc-components/hdds/western-digital-is-already-sold-out-of-hard-drives-for-all-of-2026-chief-says-some-long-term-agreements-for-2027-and-2028-already-in-place — Seagate CEO Dave Mosley, 28 Jan 2026: nearline “fully allocated through calendar year 2026”; by 28 Jul 2026: “the vast majority of our nearline exabytes are now allocated into calendar 2028.” Source: https://pcserverandparts.com/news/hard-drive-shortage-2026-what-to-do/
6. 2026 — Take-or-pay + prepayment (the terminal stage): SanDisk, 5 Aug 2026 call / 13 Aug 2026 investor day: 8 “New Business Model” customers, $93.9B minimum revenue at floor pricing, $16.5B in customer guarantees/deposits, covering >50% of bits in FY27 and ~⅔ in FY28. Sources: https://fourweekmba.com/ai-micron-and-sandisk-have-225b-in-memory-sales-under-contract/ ; https://www.trefis.com/stock/sndk/articles/617682/what-could-send-sandisk-stock-soaring/2026-10-06 — Micron FY26 Q4 (30 Sep 2026): summaries disagree — Particle reports 26 Strategic Customer Agreements, ~$150B remaining performance obligations and ~$32B commitments; TechTimes reports 16 SCAs, ~$100B minimum and $22B deposits ($18B cash + $4B letters of credit). Treat the exact SCA count/$ as unverified between sources; the direction (multi-year take-or-pay with cash deposits, >75% of CY27 output committed in one summary) is consistent across both. Sources: https://particle.news/story/micron-posts-record-results-as-hbm-capacity-sells-out-and-backlog-swells ; https://www.techtimes.com/articles/328376/20261001/micron-posts-record-54b-quarter-locks-hyperscalers-100b-memory-contracts.htm

1d. Pricing 2024–2026 and duration guidance

• Q4 2023–Q1 2024: DRAM +13–18%/qtr, NAND +15–20%/qtr (above) — the turn.
• Late 2025: conventional DRAM contract +45–50% QoQ; Q1 2026: +90–95% conventional DRAM, +55–60% NAND; Q2 2026 forecast +58–63% DRAM, +70–75% NAND (TrendForce, via secondary reporting). Server DRAM cumulative +64% in 2H 2025, ~+270% expected in calendar 2026; enterprise SSD +35% in 2H 2025, ~+235% in 2026 (TrendForce, 25 Aug 2026). HBM contract prices could rise 70–140% in 2027; some LTAs signed from 2Q26 include price ceilings. Source: https://www.trendforce.com/presscenter/news/20260825-13198.html
• Duration: Micron guided tightness past 2026, saying it could meet only half to two-thirds of demand from several key customers (Reuters summary in Dec 2025 coverage); by Sep–Oct 2026 management and SK hynix were framing tightness through 2027–28, with new fabs not at volume until late 2027/2028. Caveat: at least one Sep 2026 trade report expected price increases to moderate in Q4 2026 (DRAM +10–15%, NAND +15–20%) and Counterpoint was cited expecting DRAM price declines in 2028 — the late-cycle debate is live. Source: https://aistockwire.com/blog/memory-stocks-2029-glut-substrates-helium-shortage-october-2026

1e. Stock outcomes vs signals

Finance data: data/mu.json, data/sndk.json, data/wdc.json, data/stx.json (fetched 2026-10-26T13:42Z) plus sources below.

Lesson embedded in the table: by Oct 2026 every memory name is far off its peak despite shortages intensifying. “Shortage worsening” ≠ “stock rising” in the late stage.

PART 1B — PRE-EXPLOSION INDICATOR TIMELINES (FORENSIC)

Categories: (a) demand, (b) supply/capacity, (c) pricing, (d) transcript language, (e) insider/industry data, (f) stock.

Episode 1 — MEMORY, 2022–2026 (pre-explosion window: Dec 2022–Apr 2025)

Measured lead times (Episode 1): capex/wafer discipline → explosion: ~28 months (too early to act alone; it fired in a falling market for 2 years). Price turn (Q4 2023) → trough: 16 months; → explosive phase: ~20–24 months. “Sold out” (Dec 2024) → trough: 4 months, with a further drawdown in between (false-precision trap). Qualification-race calendar (mid-2023) correctly identified who (SK hynix first) ~9 months before its stock’s main 2024–25 run.

Episode 2 — NVIDIA GPU SHORTAGE, 2022–2024 (pre-explosion window: Nov 2022–May 2023)

Measured lead times (Episode 2): demand shock → guidance shock: ~6 months. Rental/lead-time stress → guidance shock: ~3–5 months. CoWoS-as-chokepoint identification → its own multi-year re-rating trade (TSMC/Amkor/ASE): 12+ months and still running in 2026. Hyperscaler language led hyperscaler capex dollars by ~12 months (flat 2023, inflecting 2024) — language alone is a weak sizing signal.

Episode 3 — BROAD SEMICONDUCTOR SHORTAGE, 2020–21

Earliest reliable indicators, in order:

1. Q2–Q3 2020 (a)(b): Automakers cancelled chip orders in the COVID shutdown, then demand snapped back into capacity already reallocated to consumer electronics/5G/gaming (IHS Markit, 2 Feb 2021, reconstructing the sequence). Observable contemporaneously via auto production restarts vs foundry utilisation commentary in H2 2020. https://www.plasticstoday.com/automotive-mobility/automakers-pause-production-as-chip-shortage-escalates
2. Jan–Feb 2021 (e)(d) — first hard casualties, ~9 months before peak lead times: GM idles 4 plants for the week of 8 Feb 2021; Ford cuts F-150, warns the shortage “could lower its earnings in 2021 by up to $2.5 billion”; GM guided up to $2B hit. Customer shutdowns are a lagging physical proof, but they arrived while chip stocks had not yet fully re-rated. Same source as above.
3. Apr 2021 (e)(c): Susquehanna lead times hit 17 weeks (vs ~13 pre-2021), NXP >22 weeks — first series break to a record since tracking began in 2017. https://appleinsider.com/articles/21/05/19/chip-shortage-hoarding-forcing-record-wait-times-for-technology-suppliers
4. Jul–Oct 2021 (e): Lead times 20.2 wks (Jul) → 21 wks (Aug) → 21.9 wks (Oct); auto/home-appliance chips 26.5 wks vs 6–9 normal. https://www.gizmochina.com/2021/09/22/chip-shortage-worsens-wait-time-hits-21-weeks/
5. Stock reality check (f) — the anti-pattern: “It’s been a good year for the manufacturers – the SOX is up 22% this year” (Oct 2021). The shortage’s biggest P&L impact was negative for the customers (autos: ~$110B lost sales cited; 7M+ vehicles cancelled) and only moderately positive for diversified chipmakers. Lead-time extremes fired perfectly — and there was no memory-style 5–30× bottleneck-owner re-rating, because supply was diffuse (no top-3 = 100% chokepoint) and much of the shortage was in mature-node parts with many owners. https://www.electronicsweekly.com/news/business/chip-shortage-may-peaked-2021-10/

Lead time given: plant-shutdown + lead-time signals gave ~6–9 months warning of peak scarcity, but concentration — absent here — determined whether that scarcity converted into equity explosions.

Episode 4 — CONTAINER SHIPPING, 2020–21

1. H2 2020 (b): Years of cautious newbuild ordering meant the cellular fleet would grow only 4.2% in 2021 and 2.8% in 2022 vs ~10% throughput growth (Drewry). Supply inelasticity was published before rates peaked. https://maritimetodayonline.com/?p=1236
2. Late 2020–Q1 2021 (e): LA/Long Beach congestion records, ~20 ships at anchor daily in Mar 2021; Ever Given Suez blockage 23–29 Mar 2021 (~350 ships waiting) — visible, dated, physical. https://www.hellenicshippingnews.com/container-shippings-stranger-than-fiction-first-half-of-2021/
3. Q1–Q2 2021 (c)(d): Carriers earned more in Q1 2021 than in all of the prior year; “the entire trans-Pacific fleet is effectively sold out for a month or more ahead.” Same source.
4. Price series (c): Drewry WCI: $5,472/40ft on 6 May 2021 (+278% YoY) → $8,883 on 15 Jul (+339%) → peak $10,375 on 16 Sep 2021. https://www.hellenicshippingnews.com/drewry-world-container-index-now-339-higher-than-same-week-of-2020/ ; https://www.hellenicshippingnews.com/drewry-world-container-index-keeps-climbing-higher-now-323-higher-than-last-year-and-2-9-on-week/
5. Stock (f): ZIM IPO’d ~$15 in Jan 2021, ~$20 end-June in the source’s month-by-month (with $46 cited by end-May in the same narrative — internal inconsistency in that source; use “~3× IPO price by late spring 2021”), i.e. the re-rating ran with the rate series, roughly 4–6 months after congestion/sold-out signals and before the Sep rate peak. Rates stayed extreme into 2022, then normalised — and carrier equities gave back most gains as 2023 deliveries arrived, the same late-stage trap as SK hynix in 2026.

Lead time given: congestion + fleet-growth data: ~6 months before the rate peak; price series itself: coincident-to-1-quarter leading for earnings.

PART 2 + 2B — THE BOTTLENECK SIGNATURE & INDICATOR PLAYBOOK

Ranked pre-explosion checklist (all episodes consolidated)

Concentration gate (apply before any of the above matters for equities): top-3 share ≥ ~90% (HBM 100%, DRAM ~93%, CoWoS ~1 supplier, ABF film ~95%, T-glass ~90%, EUV 100%) → scarcity becomes pricing power for listed owners. Diffuse supply (2020–21 auto semis) → scarcity becomes pain for customers and only moderate gains for suppliers. This gate is the single biggest discriminator between Episodes 1/2/4 (explosions) and Episode 3 (no comparable explosion).

Before vs after the biggest moves — summary: BEFORE: #1, #2, #3, #4, #5. AT the move: #6, #7. AFTER / late: #8 (partially), #9, #10.

PART 3 — SCAN FOR THE SAME SIGNATURE NOW (OCT 2026)

Ranked candidates

Finance data for stage calls: data/gev.json, data/lite.json, data/etn.json, data/vrt.json, data/mp.json, data/asml.json, data/tsm.json, data/cohr.json, data/leu.json, data/mod.json, data/lhx.json, data/noc.json (fetched 2026-10-26).

Evidence per candidate (a–e, condensed)

1. ABF film / T-glass — (a) Ajinomoto supplies >95% of ABF insulating film used in every advanced AI package; it notified customers in May 2026 of a ~30% price increase effective Q3 2026, was running at full capacity (~2M m²/mo in Q2 2026), and its next major (third, Gifu) plant does not produce until 2032. Nittobo holds ~90% of low-CTE T-glass; prices +20–30%, Fukushima tripling not delivering until mid-2027, downstream CCL lead times 8–10 wks → >20 wks, in places “suppliers are no longer providing lead times.” Unimicron chairman, 20 Oct 2026: equipment lead times up to 30 months, critical-material gap ~20% in 2027; ABF supply gap ~10% H2 2026 → ~21% 2027 → >40% 2028 in supply-chain analyses [modelled estimates, not company guidance]. Sources: https://www.trendforce.com/news/2026/10/20/news-unimicron-warns-of-ic-substrate-crunch-up-to-30-month-equipment-lead-times-20-material-gap-in-2027/ ; https://asiadaily.org/news/16505/ (b) Owners above; Ajinomoto is a diversified food/chemical group (ABF is a small, >50%-operating-margin segment) — purity is the problem, not concentration. (c) Hard evidence = the May 2026 price notice + 2032 capacity date (above). (d) Ajinomoto ADR +~17% 1 yr (Finnhub snapshot in search) — by far the least re-rated true monopoly in this report. EARLY-MID. (e) Falsifier: Sekisui/Chinese second-source film or E-glass substitution qualifies at Nvidia/AMD at volume; Ajinomoto pulls the Gifu timeline forward materially; substrate makers report lead times normalising for 2 quarters.

2–3. Transformers / switchgear / gas turbines — see Part 3B for the full indicator audit. Headlines: standard transformer lead 128 wks, generator step-up 144 wks, prices +77% since 2019; GEV total backlog $176B at Q2 2026, gas backlog+slot reservations 116 GW, mostly sold out through 2030, selling 2032 slots (16 Sep 2026), H1 2026 prices +>20%; Hitachi Energy CEO: sector “overwhelmed,” wait up to 4 years without a reservation; Eaton Q2 2026: Electrical Americas rolling orders +41%, electrical backlog +43–103% by segment, switchgear 52–104 wks by voltage class. Sources: https://savrn.com/blog/data-center-delivery-lead-times-2026 ; https://www.datacenterdynamics.com/en/news/hitachi-energy-ceo-warns-of-transformer-supply-shortage/ ; https://www.stockmoguls.com/2026/10/21/eaton-owns-the-bottleneck-stalling-ai-data-centers/ Stage: GEV $815 close 23 Oct 2026, +37% 1 yr but –32% from $1,196 peak and –29% in 6 mo (data/gev.json); ETN essentially flat 1 yr (–2%, data/etn.json) despite the backlog — the market has priced the equipment bottleneck partially and is now debating conversion/margins, exactly the memory H2-2026 pattern. Falsifier: lead times compress below ~52 wks, GEV slot cancellations rise, PJM auctions clear below cap with no reliability shortfall.

4. Optical — Lumentum CEO Michael Hurlston, Bloomberg, 9 Oct 2026: some optical capacity “effectively sold out through early 2029,” unable to meet ~70% of demand for some products in 2027 and ~30% in 2028; new capacity takes 3–5 years; Nvidia invested $2B each in LITE and COHR in 2026 with multi-year commitments [investment terms as reported]. Ciena backlog $8.5B, FY26 guide raised to $6.42B. Source: https://www.semimedia.cc/lumentum-says-some-optical-capacity-is-sold-out-through-early-2029-as-ai-demand-outruns-supply/ Stage: LATE — LITE +349% 1 yr to $756 (data/lite.json), COHR +72% 1 yr but –38% in 6 mo. The Oct 2029-extension headline arrived after a ~10× move. Falsifier: silicon-photonics/CPO substitution cuts InP laser content per link; Chinese module ban relief restores Innolight supply.

5. CoWoS — TSMC CEO C.C. Wei, 16 Jul 2026 call: “packaging capacity is so tight that now it’s limiting my customers’ growth” (via Silicon Analysts’ transcript review, which notes TSMC did not use the words “sold out” in 2026 calls — the “sold out through 2026” phrasing comes from the 4 Jun 2026 AGM as reported elsewhere; keep the two separate). Lead times 52–78 wks [trade reports, not TSMC]; capacity ~75–80k WPM end-2025 → ~120–130k end-2026 target, still booked; Nvidia ~60%. Source: https://siliconanalysts.com/analysis/foundry-allocation-status-q1-2026 Stage MID-LATE: TSM +44% 1 yr (data/tsm.json). Falsifier: Intel EMIB / Samsung packaging qualifies at a top-3 accelerator vendor at volume.

6. ASML/EUV — Q3 call, 14 Oct 2026: management said planned +30% capacity in both 2027 and 2028 will not meet demand, customers discussing 2029–30 deliveries, large down-payments “enough to hurt” if cancelled (CFO Roger Dassen, via Finimize/Reuters summary). Sole EUV supplier; each system $200–400M, 12–18 mo delivery + 18–24 mo calibration [trade summary]. Sources: https://finimize.com/content/asml-says-chipmaking-tool-shortages-could-last-through-2028 ; https://www.techtimes.com/articles/329442/20261023/intel-cto-euv-tool-shortage-caps-ai-chip-supply-growth-roughly-double.htm Stage MID: ASML +52% 1 yr, –10% 1 mo (data/asml.json). Falsifier: order push-outs appear in bookings before revenue (watch bookings, not sales).

7. HDD — see §1c(5). LATE: both stocks 5×+ trough-to-peak and well off peaks (table §1e). Falsifier: QLC SSD price/TB converges enough to displace nearline for warm data — the opposite was happening in 2026 (SSD prices rising faster).

8. Rare earths — China: ~69% of mining, ~90% of refining, ~99% of heavy-REE refining; 4 Apr 2025 licensing on 7 medium/heavy REEs; suspended Oct 2025 measures expire 10 Nov 2026 with no renewal in sight at reporting; DoD $110/kg NdPr floor for MP (Jul 2025) matched for Lynas (Mar 2026); European Dy oxide $1,250–1,675/kg on 27 Aug 2026 vs $270–330 in China (~4.9×). Source: https://startupfortune.com/chinas-rare-earth-export-truce-expires-november-10-with-no-renewal-in-sight/ Stage: physical bottleneck real and unresolved (no US Dy/Tb at scale as of Aug 2026), but MP stock –44% 1 yr, –61% from Oct 2025 spike (data/mp.json) — the equity already had its explosion in 2025. Falsifier: truce extended/licences flow freely; ex-China heavy-REE separation reaches commercial scale.

9. Liquid cooling — Vertiv backlog $15B at end-2025, +109% YoY (company figure, whole portfolio, not cooling-only — do not over-attribute); Modine: one customer paid $165M upfront for guaranteed capacity to supply >$4B of Airedale product 2027–29 (explicitly not a booked order); Ecolab bought CoolIT for $4.75B (Jul 2026), Schneider took 75% of Motivair. Source: https://aistockwire.com/blog/ai-data-center-cooling-stocks-vrt-nvt-mod Fails the concentration gate (at least 6 credible suppliers + hyperscaler in-sourcing) → treat as a growth trade, not a bottleneck-rent trade. VRT +6% 1 yr, –40% 6 mo; MOD –13% 1 yr, –47% 6 mo (data files). Falsifier: already partially falsified as a monopoly thesis by supplier proliferation.

10. Solid rocket motors / energetics — US SRM makers consolidated 6 → 2 (Northrop, L3Harris/Aerojet since 2023); CSIS warned the bottleneck threatens missile-defence goals; Navy awarded entrant Ursa Major $137M on 5 Oct 2026; Ursa executive: “Production is the bottleneck in missile programs right now.” Anduril opened a third US line (Mississippi, Aug 2025). Source: https://defensewarroom.com/2026/10/06/ursa-major-navy-solid-rocket-motor/ Single-source claim flagged [speculation-grade]: one Substack asserts a single qualified US ammonium-perchlorate producer (AMPAC, owned by NewMarket/NEU) — not corroborated in primary sources in this research; do not trade on it without verification. NOC –9% 1 yr, LHX –13% 1 yr (data files) — no re-rating yet, and no disclosed SRM pricing power. EARLY-MID structurally, UNPROVEN financially. Falsifier: entrants qualify at scale and multiyear procurement fails to materialise.

11. HALEU — Centrus: sole public US HALEU producer; backlog reported as $4.5B in an Aug 2026 investor presentation (Fool) but $3.0B at Q2 2026 in another Fool piece — conflicting; use “$3–4.5B, unverified range.” Stock: $464 high 16 Oct 2025 → ~$102 Oct 2026 (–70% 1 yr) (data/leu.json). Textbook post-explosion. Falsifier (for a re-entry thesis): DOE awards convert to commercial LEU revenue before Russian-import-ban timing gap (2028) bites.

12. GLP-1 — The 2023–24 bottleneck (Novo buying Catalent for ~$16.5B, closed Dec 2024, to own fill-finish) has eased: industry reporting in 2026 says supply is better than 2023–24 though still short of demand, peptide synthesis remains the slowest step, and oral GLP-1 approvals (Lilly, Apr 2026) reduce dependence on sterile injectable lines. Sources: https://www.pharmanow.live/knowledge-hub/market-trends/glp-1-manufacturing-capacity-pharma-bottleneck ; https://pharma.nridigital.com/pharma_oct26/glp-1s_manufacturing_bulk-up_key_focus_amid_drug_class_s_ongoing_growth_explosion Verdict: do NOT present as a forming bottleneck in Oct 2026; it is a maturing one. Upside surprise would require injectable demand re-accelerating faster than the multi-$B capacity now arriving.

Semi consumables (watchlist, not ranked): photoresist/specialty gases and gallium/germanium refining (China 99% of primary gallium production 2024–25 per USGS via summaries, >90% refining share for gallium/graphite per IEA Global Critical Minerals Outlook 2026) score maximally on concentration but lack a clean listed pure-play and a dated 2026 price/lead-time series in this research — insufficient evidence to rank; flagged as a gap, not a recommendation. Source: https://insights.amphora.net/market-views/critical-minerals-refining-risk-2026.html

PART 3B — ENERGY SHORTAGE UNDER THE PLAYBOOK LENS (OCT 2026)

Question: is a US energy shortage forming the way memory formed in 2023–24? Answer: it has already formed in the wholesale/forward markets; the remaining question is stage by layer.

Energy verdict — stage, honestly

• Physical shortage: FIRED outright on 7/10 core indicators (two partial, one split) (guidance-shock form differs). This is not a forming thesis; in forward power markets it formed in July 2024, the exact analogue of memory’s Q4 2023 price turn.
• Merchant power / nuclear (CEG, VST, TLN): LATE / post-peak. CEG +125% in 2024 on the Microsoft deal, then –31% over the last 1 yr to $253 (data/ceg.json); VST –27% 1 yr (data/vst.json). The Oct 2026 Google/Vistra headlines produced +11–12% single-day relief rallies in stocks down double digits YTD (Barron’s) — that is late-stage behaviour, not discovery.
• Grid equipment (GEV, ENR, Hitachi, ETN, HD Hyundai Electric): MID. Backlogs and slot scarcity are still extending (2032 slots, $176B backlog) while stocks have de-rated 30%+ from peaks — the phase where, in memory, the correct question switched from “is it scarce?” to “does margin conversion beat the multiple?” GEV passes that test on guidance so far; Eaton does not yet on delivered margins.
• Regulated utilities (AEP, D, DUK, SO): EARLY-MID but structurally capped. They own the interconnection chokepoint, not a pricing chokepoint: returns are set in rate cases, residential rates are already a political flashpoint (PJM states saw up to ~20% rate rises 2024–25; Dominion notes Virginia residential rates still ~9% below the national average), and EPS guides are 5–9%, not bottleneck rents. They are the lowest-risk, lowest-upside expression.
• The genuine “early” energy-adjacent expression in this research is not a utility at all: it is the same second-order materials layer as in semis — GOES electrical steel, HV bushings (2-yr leads), and transformer components — where we found lead-time evidence but no clean, verified listed pure-play with dated pricing disclosure. Flagged as an evidence gap, not a recommendation.
• What would falsify the energy-shortage thesis: (i) a PJM BRA clearing materially below cap with no reliability shortfall; (ii) transformer lead times <52 wks or turbine slot cancellations; (iii) ERCOT/PJM queues converting at <5% while hyperscaler capex guides are cut (proving ghost demand); (iv) data-centre efficiency (inference optimisation) bending LBNL’s 2030 range toward its 9.5% low case. None of these had occurred as of the latest dated evidence above (Sep–Oct 2026). The biggest real risk is the opposite timing trap from memory: demand forecasts double-counted across queues (AEP Ohio alone reported >30 GW of inquiries vs 8–10.5 GW system peak, Feb 2026) — if even the contracted (not queued) load disappoints, 2029+ capacity arrives into a hole. Source: https://www.globaldatacenterhub.com/p/unbuilt-data-centers-electric-bills

COULD NOT VERIFY / GAPS

• Micron SCA count and $ (16/$100B/$22B vs 26/~$150B RPO/~$32B) — sources conflict; direction agreed, exacts unverified.
• Centrus backlog $3.0B vs $4.5B — conflicting secondary reports.
• H100 rental $8/hr Q1 2023 series is a retrospective compilation; contemporaneous direction confirmed, precise level not.
• ZIM spring-2021 price path has an internal inconsistency in the month-by-month source; only “~3× IPO by late spring” is asserted.
• Ammonium-perchlorate single-producer claim (NEU/AMPAC) — single Substack source, unverified.
• Gallium/germanium and GOES/bushing pure-plays — concentration evidence strong, dated listed-owner pricing evidence insufficient to rank.
• Lithium and MLCC episodes were not included: the two required additional episodes are covered by semiconductors (2020–21) and shipping (2021); no lithium claims are made in this report.
• Several 2026 stock “peaks” are 52-week highs from the finance-data files, not all-time intraday highs; multiples are trough-to-peak on that basis unless a source states otherwise.

SOURCES (finance data files)

data/mu.json, data/sndk.json, data/wdc.json, data/stx.json, data/gev.json, data/lite.json, data/etn.json, data/vrt.json, data/mp.json, data/asml.json, data/tsm.json, data/cohr.json, data/leu.json, data/mod.json, data/lhx.json, data/noc.json — quotes/OHLCV, fetched 2026-10-26T13:42Z. data/nvda.json, data/ceg.json, data/vst.json — fetched 2026-10-26T13:45Z. All other sources are cited inline at the claim, with the date the underlying event/call/auction occurred distinguished from the publication date where the source allows.