Erica · PM Research MemoAll memos

PM RESEARCH MEMO

Title: Q3 Tech Divergence — Certainty Premium, Semis Washout, and Capelight’s Rotation Map
Author / source: Jeff Keller (founder & PM, Capelight Partners), guest on Other People's Money / The Monetary Matters Network with host Max Wiethe
Source title: This Market Beating Hedge Fund Manager Explains Recent Tech Sector Rotation | Jeff Keller
Source URL: https://www.youtube.com/watch?v=zaI0QdiAmjs
Video ID: zaI0QdiAmjs
Published: 2026-08-31 (YouTube upload calendar day; exact clock not exposed by yt-dlp — treat as calendar day 2026-08-31 UTC / same calendar day America/Toronto)
Duration: 1:01:36
Memo date: Tuesday, September 8, 2026 (America/Toronto)
Transcript: /workspace/youtube-transcripts/zaI0QdiAmjs.md (~12,791 ASR words) · Brief: /workspace/youtube-transcripts/zaI0QdiAmjs_brief.md
Caption source: YouTube English automatic captions (ASR) only (en-orig). No manual/official English track. Heavy proper-noun/ticker garbling locked from context + brief.
Source type: Tech-sector hedge-fund PM interview on Q3 rotation, AI complex, process/portfolio construction. Channel / guest framing: process views as of recording (~late Aug 2026 publish). Not investment advice.
Product: Regime / rotation map for allocation research. Not advice. No buy/sell from this desk.
Live levels: Q3 sector % (software/services ~+20%, tech hardware ~flat, semis/semi equipment ~−7.5% “as of yesterday’s close”), cyber/infra ~25–30× revenue, EWY ~73 / SOXX ~55 one-month realized vol, ~$500B labs+SpaceX+Google raise YTD — all as of source tape. Live quotes / Nvidia post-print / Anthropic S1 timing = External check needed.
Source discipline: Primary source is this transcript (+ brief for ASR locks only). Companion desk memos (Gavin Baker, Jordi Visser, Ben Hunt, Cboe, etc.) are different sources — do not silently merge numbers. Flag external facts. Keller cites Gavin Baker on AI-chain multiple dispersion; treat that as Keller’s citation of Baker, not a second primary.

ASR name locks:
| ASR heard | Likely intended |
|---|---|
| Capeite / Capelite | Capelight Partners |
| Maxi | Max (Max Wiethe) |
| socks / Socks / Socks V | SOXX (and SOXX realized vol) |
| IGV | iShares Expanded Tech-Software ETF (IGV) |
| Kathy Wood arc | Cathie Wood / ARK |
| Core Core Weave | CoreWeave |
| Nebus | Nebius (neocloud) |
| Cliff Sawson | Cliff Asness |
| EWI / EWY | EWY (iShares MSCI South Korea ETF) in vol discussion |
| Microso… / “one of the other hypers” | Microsoft / hyperscaler earnings call |
| Spaxs | SPACs |
| Jeeoff | Jeff |
| situational awareness and deleveraging in Korea | Situational Awareness (essay/theme) + Korea margin-call delever |
| meggaap | mega-cap |
| longshore | long/short |
| Demisabus / Sebastian Malby | Demis Hassabis / Sebastian Mallaby (book context — ASR soft) |
| Nickeola | Nikola (2021 SPAC-era example) |
| JeffLlight.co | Capelight email / site (spoken .co) |


EXECUTIVE SUMMARY


SOURCE-ACCURATE SUMMARY

Chapter-ordered with approximate timestamps. Quotes ≤20 words where useful. ASR locks applied in brackets where needed.

Cold open & setup (0:00–1:54)

  1. (0:00–0:27) Cold open. Keller (montage): “long-term uncertainty is so high that anything with short-term certainty is trading at a huge premium.” Heavy retail in a trade → “the clock is ticking.” Disconnect between “financial analysis” and West Coast “religious analysis.” “There’s going to be a lot of pain… if this ever rolls over… I just don’t think it’s yet.”

  2. (0:27–1:22) Intro. Host Max Wiethe (Other People's Money / Monetary Matters); guest Jeff Keller, Capelight Partners (ASR: Capeite), tech-focused HF since 2021. Host praise for navigating 2022 and subsequent undulations vs many 2021-birthed tech HFs.

  3. (1:22–1:54) Q3 divergence (host numbers, “yesterday’s close”). Since start of Q3: software & services ~+20%; tech hardware & equipment ~flat; semiconductors & semi equipment ~−7.5%. “Massive divergence” vs start of 2026. Asks for Keller’s read + background.

Background, S-curve, software split (1:54–8:15)

  1. (1:54–2:50) Ops background & AI inflection. Salesforce + MongoDB ops experience → theme-oriented, longer horizon, extreme dislocations. Early-2026: Iran war distracting from “massively inflecting” AI adoption; Q1 takeoff. Q2: not calling blow-off top, but ~100% SOXX gain + heavy retail → chop expected. “Situational Awareness” + Korea deleveraging → vol “obvious in hindsight.” Still “moving upwards on an S-curve.” Bulls’ color: Apple services / Google search still double-digit; Meta “growing in the 20s”; labs could have 10–15 years double-digit growth → “lift all boats.” Late June not end of mega-trend — “normal volatility.” Software “thrown out with the bathwater” → bounce makes sense.

  2. (3:11–4:12) Disruption narrative vs technicals. Hard to say software-disruption narrative disproven (newspaper analog: market priced pain before earnings). Extremely strong results vs narrative. Rally = fundamentals re-rating or technicals (software shorts paired with AI longs; degrossing still happening across HF complex after big delever event).

  3. (4:12–5:31) SOXX / IGV −1; app vs cyber/infra. “Socks and IGV basically have a negative one correlation” → technical element (bad AI news → ServiceNow / Salesforce up several %). Split software: application still under question; initial crash = recalibration from premium multiples / decade of double-digit growth → low-double / high-single; five-to-fifteen percent growth zone hard to “dream the dream” on revenue multiples / non-GAAP EPS. Salesforce / Workday (acquisition chatter) framed as coming back to earth with left-tail risk, not full AI wipeout. Bounce in cyber and infra more pronounced.

  4. (5:31–6:48) Certainty premium & cyber multiples. Usage-based / accelerate-able names get premium. Cyber: people “rushing to build up their defenses” vs AI threats → near-term number visibility. Infra: more software adoption → some flow into existing tools. Multiples “come a long way” — ~25–30× revenue for large companies; “success rate on those… paying up there is low.” Next couple quarters good; cyber not immune (labs think they’ll disrupt cyber too). Momentum + “beat for a little bit.” “Lofty territory… more so than AI, frankly.” Aside: “real AI winners are CrowdStrike and… rather than maybe some of the semis” (spoken comparative, not a desk rec).

  5. (6:48–8:15) Infra examples. MongoDB (ex-employer, “biased”): AI → software proliferation (vibecoded apps / more efficient org development) → every app has a DB line → end-market explosion. Selection now via LLMs not humans — mix shift unclear; “rising tide may lift that boat.” CDNs / Fastly / anything delivering compute resources → usage uplift short-term. Infra higher chance of beating near-term vs Salesforce / Wix-class.

Theme vs picking; AI allocation; valuation (8:15–15:25)

  1. (8:15–9:36) Theme > within-theme early. Bottoms-up still required, but “a lot more money gets made being in the correct major theme.” Cloud-migration analogy: sector first, names later. AI: still open-ended but GDP-scale → dispersion / sustainability starting to matter vs pure rising tide.

  2. (9:36–10:09) Host: S&P sector context. Tech third-worst among 11 sectors into the bounce; only industrials and utilities worse — those “thrown into the AI trade.” Where to stay exposed?

  3. (10:09–11:54) Risks of short-term extrapolation; multiple dispersion. Don’t extrapolate very short-term dynamics — e.g. extremely high price of compute, neoclouds; Elon entering may move “quicker to the glut.” Bottlenecks / semi names “trade on the same factor” — hyperscaler capex + lab ARR — yet “hugely different multiples.” Cites Gavin Baker: low-multiple AI-chain names may do better vs high-multiple extrapolators. Behind-the-meter / services on out-year numbers rich; Nvidia / Micron “really not extrapolating… telling you we’re at peak.” Contrarian lean: low-multiple names across the AI trade likely better even at different earnings-cycle stages. “Hidden AI winners” pricing rosier future than many semis.

  4. (12:24–13:38) Valuation horizon. Four-to-five-year projections dangerous — five years ago was NFTs, SPACs, metaverse; AI “not on the radar.” Cyclical businesses at massively inflated earnings + more growth/tech investors → multiple dangerous even where “very cheap.” Prefer ~3y visibility; beyond that hard in tech. Market puts “very high premium on one to two quarters.” Keller’s game: generally 2–3y through-cycle earnings multiple with decent downside protection.

  5. (13:38–14:54) Bridging long view vs market’s certainty premium. L/S beauty: both sides of book; isolate factors. Anecdote: time horizons of people he talks to “shorter than I can remember” — catalyst paths, beat sell-side consensus. When people that short-term, take a longer look. Shorting dangerous; near-term imbalances can get “juicy.” Opportunity: sit where people extrapolate and wait with 1–2 year horizon.

Sobriety, retail clock, 2021 analogy, rates (15:25–27:37)

  1. (14:54–17:20) Dampened announcement feedback loop. Host: partnership/deal catalysts (semi×Nvidia, bond offerings for capex) used to rip 10–20%; still announced, weaker reaction. Keller: “a lot more sobriety” — encouraging vs 2021 euphoria “multiples higher.” Many DC lease announcements have unclear economics (who’s on hook for extra costs / delays). OpenAI/Anthropic association was “gold” earlier; more scrutiny on sustainable profits. Crypto miners converting to DCs: euphoria “totally come off” (then host notes Bitcoin ripping → dual beta). Prefer environment where you “sit and earn” vs 20–50% on press releases.

  2. (17:45–19:37) Retail / holder base. Heavy retail involvement → “clock is ticking” (experience rule). Examples: gold/silver earlier this year; crypto; 2020–21 software (paid newsletters on MongoDB, Elastic, Fastly). “Boring hardware companies for 10 years… now all the rage among retail” = dangerous. Flip side: AI more open-ended; “shorting an open-ended growth story can be deadly” — no coffin nail. Adoption up S-curve; capex up for years. Not a 2022 environment; not “June was the blowoff top… semis bleeding for years.” Couple months = “information vacuum”; people latch onto every ARR leak. Cleaning leverage healthy. Likes short selling; not piling into shorting AI.

  3. (20:29–23:08) 2021 vs now (host probe; Keller quibble). Host: 2021 Archegos / ARK delever marked top of “arc” names even as market powered higher until 2022 inflation/rates; technical pain for “tourist” semi investors possible. Keller: ironic similarities (levered HFs); but “real businesses” (Tesla, software) peaked late-21; SPACs/GameStop cratered earlier. 2022 damage to real businesses = inflation/higher rates + ~50× revenue software after pandemic pull-forward. AI: some pull-forward, more legs, “more reasonable” multiples. Trade breaks when adoption slowing — not yet; complex seeing “explosive” adoption. Don’t want to short; rip again unclear.

  4. (23:36–27:12) Rates, fiscal, funding AI. Host: Fed hike risk next few months / by end-2026 priced; not 2022 zero-bound cycle; fiscal/long-end / credit-funded spend crowding Treasury; rate-sensitive real businesses. Keller: won’t forecast bonds; bond “freakouts” often reverse (10y could be −30–40 bps in two months). Capital crunch framing: hundreds of billions → near trillions that used to be buybacks now also needing debt. “More dangerous way to finance this buildout” — “not that worried.” Post-Anthropic IPO, labs + SpaceX + Google “about $500 billion year to date”; markets fine; lots of debt too. Cash businesses still growing double digits / 20s; cost base could be cut (2022 precedent). Silicon Valley leaders: existential race (Sergey/Larry back; Elon pivoting to DCs). Extra 150 bps borrow / aggregate $10–20B/yr more → won’t slow them. “Disconnect between the financial analysis… and the religious analysis… West Coast.” Grew up West Coast — can do both. Credit markets might try to slow; unless yields “really blowing out,” enough money for a couple years. If ROI never materializes → problem later.

Mag7, ROIC optionality, Meta (27:37–39:53)

  1. (27:37–29:34) Mag7 / hyperscaler profile. Host: Google first negative FCF quarter; Mag7 no longer clear market leaders. Keller: owns “some”; stocks “reasonably cheap”; different profile. Amazon ~10× since ~2016–17 (check him); can still make 15–20%/yr — but prior multi-bagger association “probably behind us.” Hyperscaler/cloud model (ex-Apple, ex-Meta among Mag7 drivers): ~3 vendors with thousands of customers → ~9–10 scaled compute providers (Oracle, SpaceX, labs direct, neoclouds CoreWeave, Nebius). Scale advantages remain; fewer customers; more uncertainty/competition; GDP-level market caps → 2/5/10× golden period of last 5–7 years over; still can be good risk/rewards — different profile.

  2. (30:05–32:37) Capex as optionality / ROIC. Host: Google adj. FCF estimates spoken ~$18.8B (’27) then positive ~$28.5B (’28) — too much certainty? Keller: ROIC debate; frame as buying an option — defensive (existential if you don’t play) + offensive (cloud numbers, new products, open-ended upside if leading compute). Hyperscaler call (Microsoft or “one of the other hypers”): buy powered shells, don’t fill with chips yet → years of optionality to ramp capacity. Low-but-positive ROI on that option still rational. Alternative (not playing) too risky. Implications: (1) keep spending to accumulate options; (2) 4–5% ROIC not great for shareholders but not incinerating capital; (3) spend driven more by optionality need than near-term ROI. Hyperscaler non-participation partly = uncertainty around that ROIC (ties to certainty-premium tape).

  3. (33:07–35:20) 2027–28 visibility risk. Host: what if 2027 arrives without ROIC visibility into 2028? Keller: three-horse → eight/nine/ten-horse business; profit locus in chain unclear; Anthropic/OpenAI building own power/DCs. AWS growing “40s” vs two years ago praying for ~18% — stocks “incrementally higher but not materially” vs what that growth would have implied. Queasiness on ROI + concentrated customers. Meta “on the ropes” with it. Andy Jassy laid out positive ROI case — “on board… for now”; can change quickly. Overestimate exec sensitivity to stock price (religious zeal); Meta took long into 2022 to cut. 10–20% drawdowns won’t necessarily flinch capex; hard to see market forcing stop if ROIC “good enough.”

  4. (35:20–39:27) Meta bear case / bull AI tension. Host: Meta signaled overbought compute could be sold/subleased → market rewarded; if Zuck can’t use compute, isn’t that bear case for AI? Keller: renting excess “a little shortsighted”; messaging = give market rope (“we will stop or release capacity if need be”). Three-month contracts / couple billion profits helpful but secondary to signaling on a trillion-dollar-plus company. Meta internal dynamics, hiring, open-source bounce; winners/losers within foundation models — aggregate progress more relevant to broader AI trade. Zuckerberg tendency (Mallaby/Hassabis book anecdote): equally enthusiastic AI / crypto / NFTs — go big on trends not obvious for core business; as shareholder “binary… not one I can underwrite.” Meta troubles ≠ broader AI troubles. Host: hard to hold “bull AI / compute = new oil” and “Zuck overbuilding.” Keller: Zuck’s view aligns with optionality + existential; biggest relative spender gets most questions; rational to tell investors “playing… will back off if need… not backing off yet” even if public markets whip the stock.

Book construction, net short bar, three pillars (39:53–46:33)

  1. (39:53–41:30) Theme exposure / contrarian pockets. Want big themes; contrarian streak. Data-center politics / midterms concern overstated; stocks implying challenged DC starts (labor, power-site applications) — unlikely; look past midterms/moratoriums; tech moral panics usually lose to market strength. Looking at things pricing in DC-start deceleration. Wary of extrapolation wherever high near-term certainty → back slope of mini-cycles. Idiosyncratic growth “left for dead” (e-commerce, some recent IPOs) compelling vs crowded “is software dead?” debate.

  2. (41:30–43:12) Long-biased L/S philosophy. “Long biased” = more than ~50% net, can be higher. Pod/momentum/market-neutral relative-value capital creates opportunity for unconstrained model. Capture beta for LPs without being paid for beta; more duration; not perfectly factor-balanced. As more money must be factor-balanced quarter-to-quarter or delever at wrong time (e.g. July), “old school long/short… renaissance.” Tough business (LPs want cheap long-only or high-fee market-neutral) → opportunity for old-school stock picking.

  3. (43:46–46:08) Net short bar; three break pillars. Unlikely ever net short. Late-2021 = once-in-10–15-year (gov support, retail, COVID cyclicality in secular industries). Markets faster now → maybe more often. AI eventually has ingredients (huge activity change, enthusiasm, curing-cancer/money-obsolete talk, cyclicality in secular-growth investor base). Pillars: (1) lab ARRs / adoption S-curve — no crack yet; (2) capex — concerns on forever growth rates, but leveling high still OK; capital markets tightening but still provided hundreds of billions this year — not worried markets stop it; (3) price of compute — signals overbuild; not there yet; DC build times not suggesting glut (Elon might move that). If/when cracks: short-selling skill helpful — “highly cyclical businesses trading peak on peak” + even some names he owns → “a lot of pain… if this ever rolls over… don’t think it’s yet.”

Echo chamber, factors, July/Korea playbook, year-end (46:33–1:01:36)

  1. (46:33–48:42) AI tools / echo chamber. Host: FinTwit/X + LLM sense-check = bulletproof-feeling theses markets may ignore. Keller: surprising AI-generated content in sell-side / Substacks / chatter; job is own thoughts; consensus from LLMs low value. AI good for ramp/facts; enables shortcuts + overconfidence → more behavioral inefficiency. Disagrees with “AI removes fear/greed” podcast claim — expects faster, more crowded narrative cycles + phones/terminally-online → markets less efficient (quant models maybe exception).

  2. (49:07–51:38) Factor awareness vs LP management. Must know factor exposures even if knowingly taking more than a minimizer. Speed of bucketing company alpha into themes is extreme. Keller: spent last 1–2 years much more factor-aware (wasn’t ~4y ago). Ultimately time horizon + LP management: monthly results “completely insignificant”; quarters often noise; software LPs asking “what’s wrong with process” when AI funds levered and shorted software — two-month thing, stocks “round-tripped over four months.” Be conscious of factor/sector bets. Extremes as opportunity: washouts, “uninvestable,” largest HF shorting day in seven years → pay attention.

  3. (52:02–55:00) July sense-check / Korea / optionality. Life cycle: “stock’s down I can buy more” → worry → “bigger factors.” Then margin calls / Situational Awareness → clearer. Pain concentrated Asia / Korea; notable % of Korean population margin-called (rumor caveat); huge retail + HF delever into secular trend still OK for “basically two large companies in that country.” Want to be buyer when people are getting margin called and liquidated. Maintain optionality: cash and/or lower gross — “not the one delevering or maxed out.” 3σ / once-in-20y events “every several months.” Longer horizon; don’t trade too much; wait until something “hits me in the face.” Money made on “one or two good ideas a year… couple big ideas every couple years” — optionality > maxed great stocks always.

  4. (55:26–56:55) Vol below the surface. Host: EWY realizing ~73, SOXX ~55 one-month (peaks ~185). How much leverage needed? Keller: runs comparatively very low leverage; leverage reduces optionality. Headline indices tranquil (VIX) while sector/stock vol spikes — “duck kicking below a surface.” Year of rotations = opportunity at extremes. Guess: shorter horizons, more uncertainty, more money into less-fundamental strategies → behavioral advantages increase.

  5. (57:22–58:55) Close-out 2026 watchlist. Anthropic S1 “probably next week or the week after” (as of tape) — maybe little new info (numbers broadly leaked) but public metrics hugely helpful vs third-hand ARR. Most important for AI trade: (1) ARR / adoption (labs + open source); (2) capex (strong for ’27; people look at ’28; feeling better on hyperscalers vs 1H); (3) forward prices of compute (very tight now; overbuild shows here). Stocks: would not surprise if chop; June retail/investor enthusiasm usually doesn’t rip straight to new highs; waiting for next breakthrough/acceleration (recursive self-improvement “floating”); labs public → less narrative vol.

  6. (58:55–1:01:05) SpaceX-effect / Anthropic IPO. Host: SpaceX IPO sucked capital from also-ran space names — Anthropic pure-play lab risk to AI proxies? Keller: possible; “don’t have a ton of concern.” Equity issuance signals euphoria differently when into real businesses at realistic prices vs 2021 SPACs / Nikola-class. Space stocks doubled then halved → back. AI stocks could run into Anthropic IPO then come down — sure. Semi euphoria/optimism/positioning “kind of cleaned out”; market “ready to digest” Anthropic; will want lab exposure; expects crazy price / lots of enthusiasm — “not saying I’ll buy it.”

  7. (1:01:05–end) Outro. Find him: Twitter Jeff Keller1; email at Capelight .co. Thanks / goodbye.


SYSTEMS MAP / VALUE CHAIN ANALYSIS

Key players and flows

Bottlenecks and leverage points

  1. Certainty premium — 1–2 quarter visibility dominates tape; penalizes long-dated ROIC stories (hyperscalers, some semis).
  2. Shared driver, dispersed multiples — one factor (capex + lab ARR), many prices → relative-value research surface inside AI.
  3. Adoption S-curve / lab ARR — pillar #1; IPO disclosure reduces narrative vacuum.
  4. Capex path — pillar #2; leveling high ≠ stop; funding via credit is the soft constraint.
  5. Forward price of compute — pillar #3; glut tell.
  6. HF pairing / degrossing — mechanical SOXX vs IGV; still running after July.
  7. Retail enthusiasm peaks — June-type clocks; don’t expect straight rip to highs.
  8. Portfolio optionality — cash/gross headroom to buy forced liquidations; leverage destroys that option.
  9. DC politics / midterms — Keller: overstated vs inertia; market may be overpricing start deceleration.
  10. LP reporting clock — monthly noise vs multi-month factor round-trips; process risk if LPs force bad delever.

Upstream / downstream

Where constraints create research opportunities (not recs)


SECOND AND THIRD-ORDER EFFECTS

Chain 1 — High long-term uncertainty → certainty premium → software/cyber outperformance vs semis/hyperscalers.
[Primary] Q3 soft/services +20% / semis −7.5%; SOXX–IGV ~−1; cyber 25–30× rev; hyperscalers punished on ROIC opacity (source).
→ [Second] Tape rewards usage beats and near-term cyber/infra; penalizes GDP-scale capex optionality even if S-curve intact.
→ [Third] Mean-reversion risk if 1–2q certainty disappoints or if ROIC/ARR visibility arrives and reallocates back into semis/hyperscalers.
→ [Relevance] Hypothesis: treat Q3 divergence as factor/certainty regime, not AI thesis kill. Research map only — no rec.

Chain 2 — Shared AI driver + multiple dispersion → low-multiple vs extrapolator relative game.
[Primary] Same capex/ARR factor; Nvidia/Micron “not extrapolating”; hidden winners rich; Baker citation (source).
→ [Second] “Creative” behind-the-meter / services can underperform even if AI continues, if priced for perfection.
→ [Third] Glut scare via compute prices hits extrapolators first; semis that already signal peak may be relative shelters or still beta to the factor.
→ [Relevance] Hypothesis: dispersion research > binary AI on/off. Not a ticker call.

Chain 3 — Retail clock vs open-ended growth → conflicting risk clocks.
[Primary] Heavy retail = clock ticking; shorting open-ended growth deadly; June enthusiasm; not 2021 1:1 (source).
→ [Second] Positioning washouts (Korea, HF degross) without adoption crack = buy-the-liquidation setups for long-biased books with dry powder.
→ [Third] If retail returns before pillar cracks, chop/false breakouts; if pillars crack with retail still in, “lot of pain.”
→ [Relevance] Hypothesis: watch retail share of flow + three pillars jointly, not retail alone.

Chain 4 — Existential West Coast spend + debt funding → persistence until yields blow out or ROI falsified late.
[Primary] Religious/existential framing; +$10–20B/yr won’t stop; ~$500B raised YTD; debt more dangerous (source).
→ [Second] Equity drawdowns of 10–20% insufficient to stop capex (Meta 2022 lag).
→ [Third] Financial analysis that underweights “must play” optionality systematically underestimates spend duration — until compute glut or adoption crack.
→ [Relevance] Hypothesis: capex stop is a pillar event, not a vibe from Mag7 stock charts. Diligence.

Chain 5 — Pod/factor capital + LLM echo chambers → mechanical rotations + overconfident consensus.
[Primary] Factor-balanced money delevers wrong; AI content flood; markets less efficient (source).
→ [Second] Idiosyncratic stories (e-commerce, left-for-dead IPOs) and DC-politics overshoots become hunting grounds; LP reporting design preserves alpha when two-month factor pain looks like process failure.
→ [Relevance] Hypothesis: process edge = optionality + LP horizon, not more LLM sense-checks.

Chain 6 — Lab IPO disclosure → less narrative vacuum; possible pure-play rotation.
[Primary] Anthropic S1 imminent; SpaceX analog possible but secondary; semis already cleaned (source).
→ [Second/Third] Public metrics replace ARR leaks; crazy IPO price could recreate retail clock in lab equity without falsifying S-curve.
→ [Relevance] Hypothesis: separate “lab equity IPO froth” from “adoption pillar.” Watchlist only.


SCENARIO FRAMEWORK

Horizon: remainder of 2026 into 2027 capex/ARR prints; weeks for Anthropic S1 / near-term Fed/bond freakouts. Odds qualitative from Keller’s tone (he does not assign numeric probabilities).

Base / “Intact S-curve, choppy rotation” (Keller’s center of gravity)

Assumptions: Adoption S-curve unbroken; capex stays strong into ’27; forward compute prices stay tight (no glut signal); Q3-style factor divergence can persist or partially mean-revert; Mag7 remain “different profile” not multi-bagger era; retail not fully re-melted; HF degrossing residual fades slowly; Anthropic/lab disclosures arrive without SPAC-like capital incineration.
Market shape: Chop around post-June enthusiasm; possible acceleration if tech breakthrough (recursive self-improvement “floating”); software certainty names stay bid until multiples fail or cyber disruption narrative bites; semis choppy but not multi-year bleed.
Process expression (Keller, not desk recs): Long-biased; theme exposure; dry powder for liquidation extremes; unlikely net short.
Leading indicators: Lab ARR / open-source adoption; hyperscaler capex guides; forward compute prices; SOXX–IGV correlation normalizing; Korea/EWY vol and margin-call headlines quieting.

Bull acceleration / “next breakthrough”

Assumptions: Adoption upshift or model breakthrough; compute scarce; capex re-accelerates; lab IPOs digest; ROIC visibility improves enough to re-rate hyperscalers without killing optionality spend. Narrative vol may spike — retail clock returns. Even bull path may stair-step (June peaks often don’t rip straight to highs).

Bear / “Pillar crack” (not base; “lot of pain if rolls over”)

Assumptions: (1) adoption crack, and/or (2) capex stop, and/or (3) forward compute prices signal glut (Elon/neocloud path). Amplifiers: yields blow out, funding shut, or 2021-like confluence (retail + leverage + cyclical peak in secular clothes). Peak-on-peak cyclicals see severe pain; short-selling skill valuable (Keller). Tells: ARR deceleration; cancelled projects; compute price collapse; retail still crowded; HF leverage high into the turn.

Left-tail process / “forced to delever into opportunity”

Another July-like 3σ (Asia margin calls, HF degross) without pillar crack → books with optionality buy liquidations; maxed books become supply. Watch EWY/SOXX realized vol under calm VIX; “uninvestable” rhetoric; extreme shorting-day prints.


COMPANY / ASSET WATCHLIST

No ratings. No buy/sell from this desk. Items are research map nodes from the source. Keller’s portfolio comments are process color, not disclosed holdings to copy.

SOXX / SMH / semiconductor & semi equipment complex

IGV / application software vs cyber-infra

MongoDB / Elastic / CDN–Fastly class (usage infra)

Nvidia / Micron (as cited “not extrapolating”)

Mag7 / hyperscalers (Amazon, Google, Microsoft, Meta; Apple services as growth color)

CoreWeave / Nebius / neoclouds + price of compute

Anthropic / OpenAI / lab complex (+ SpaceX as funding/IPO analog)

EWY / Korea leverage node

ARK / 2021 analog set (negative template)

Data-center starts / politics complex

Capelight process markers (meta-watch)


DILIGENCE QUESTIONS & RESEARCH AGENDA

  1. P1 — Live Q3 sector tape: Replicate host’s software/services +20% / hardware flat / semis −7.5% from Q3 start through interview “yesterday” and update through memo date (8 Sep 2026). External check needed.
  2. P1 — SOXX vs IGV correlation: Estimate rolling correlation through July delever and Q3; test “−1” as rhetorical vs statistical.
  3. P1 — Three pillars dashboard: Lab ARR / open-source adoption; hyperscaler + lab capex; forward compute price term structure — weekly.
  4. P1 — Anthropic S1: Actual filing date/metrics vs late-Aug “next week” expectation; compare to leaked ARR.
  5. P1 — Nvidia print: Tape said not yet reported — capture guide tone vs “not extrapolating” frame. External check needed.
  6. P2 — Cyber/infra multiples: Which large names were ~25–30× revenue at interview; subsequent compression/expansion.
  7. P2 — Compute supply map: Validate ~9–10 scaled providers list (Oracle, SpaceX, labs direct, CoreWeave, Nebius, etc.) and share/capacity.
  8. P2 — ~$500B YTD raise claim: Equity+debt tally for labs + SpaceX + Google — primary-source rebuild.
  9. P2 — EWY/SOXX realized vol: Confirm ~73 / ~55 one-month and ~185 peak; relate to VIX surface.
  10. P2 — HF pairing residual: Software short vs AI long positioning proxies (prime broker / 13F lag / borrow).
  11. P3 — DC politics vs starts: Moratorium headlines vs MW under construction post-midterms path.
  12. P3 — LLM echo-chamber test: Track narrative half-life of FinTwit consensus theses vs price — behavioral inefficiency claim.
  13. P3 — Meta capacity signaling: Any actual sublease volumes vs messaging-only; implication for “bull AI / bear Meta” consistency.
  14. P3 — Powered-shell optionality: Capex that is shell/power vs chips — accounting visibility in hyperscaler disclosures.

RISK ANALYSIS

Thesis risks
- Rotation ≠ regime. Q3 divergence can reverse quickly; treating certainty premium as permanent overfits a short window.
- Baker-via-Keller. Multiple-dispersion point is second-hand; verify against Baker primary if used in LP memo.
- “Not worried” funding view. Debt-funded buildout risk acknowledged then downplayed — bond blowout is the explicit caveat.
- CrowdStrike / semis aside. Spoken comparative “real AI winners” line is conversational, not a sized thesis — do not elevate to desk view.
- Meta underwrite refusal. Keller won’t underwrite Zuckerberg binary — any Meta expression needs separate work.
- Open-ended growth short-danger can strand bears and trap late longs if pillars crack late.

Timing risks
- Information vacuum between lab leaks and IPOs → whipsaw.
- June enthusiasm overhang vs possible breakthrough acceleration — two-sided near-term.
- Fed hike / bond freakout clocks vs multi-year S-curve clock.
- July-like liquidations “every several months” claim — optionality must be pre-funded before the event.

Execution / data risks
- ASR-only — Capelight, SOXX, IGV, CoreWeave, Nebius, Cliff Asness, EWY, ARK, etc. locked from context; LP-facing quotes need audio verify.
- Spoken figures (sector %, 25–30×, $500B, vol 73/55, Amazon 10×, Google FCF path) are guest/host estimates, not audited.
- Holdings, nets, gross unknown — process description ≠ position book.
- Publish date 31 Aug 2026; memo 8 Sep 2026 — week of staleness on “next week S1,” Nvidia, live levels.
- Channel: Other People's Money / Monetary Matters — entertainment/education framing; not advice.

External / regime risks
- Adoption S-curve crack (pillar 1).
- Capex stop or funding shut (pillar 2).
- Compute glut / price collapse (pillar 3), possibly accelerated by large new supply (Elon/neoclouds).
- Yields “really blowing out.”
- Cyber disruption by labs falsifying certainty-premium winners.
- Another Archegos/ARK-style technical top without waiting for 2022-style macro — Keller says ingredients can appear even if fundamentals differ.
- Companion desk sources (Baker primary interviews, Hunt repression, Cboe vol) must stay attributed if combined later.


Desk copy. Source-disciplined. Not a trade recommendation. Erica · 8 Sep 2026 (America/Toronto). PRIMARY: zaI0QdiAmjs (Monetary Matters × Jeff Keller / Capelight).

Desk copy · not a trade recommendation · Erica · 8 Sep 2026 · HIGH batch