State of the corpus — 2026-05-20

Bottom line up front

This week's flow is overwhelmingly a single trade with multiple expressions: AI physical infrastructure scarcity is being re-priced across memory, optics, power, HPC landlords and silicon, with NBIS's Q1 (revenue +684% YoY, raised >4GW power guide) and CEG/GEV/VST/EOSE/GLP-ramp prints validating that AI capex is converting into signed megawatts and contracted compute. The dominant story is not "buy NVDA" — it's "own the bottleneck Nvidia depends on," and that is increasingly the power layer and the optical/memory chokepoints. Incremental capital today is best deployed in the *uncrowded reinforcing legs* — InP substrates (AXTI), power conversion semis (NVTS/ON), and proven HPC landlords (NBIS) — rather than the crowded GPU/treasury-flywheel trades. The single largest framework risk is that c13's breadth/yield warning becomes the governor that forces a capex multiple reset across c1–c5 simultaneously.

Reinforcing complexes

• The AI physical-bottleneck stack = c1 (memory) + c2 (optics) + c3 (power) + c4 (HPC landlords) + c5 (silicon). All five depend on the same upstream: hyperscaler AI capex through 2027-2028. @semicon_eng1↗, @StockSavvyShay↗, @TheValueist↗ and @ParadisLabs↗ cross-author all five, which is why they move together. The least crowded expression is AXTI (InP substrates) — a single chokepoint that benefits whether the bottleneck resolves through optics or memory, with far less author concentration than MU or COHR.
• Power + landlords joint dependency = c3 + c4. Data center landlords (NBIS, IREN, CRWV) need contracted megawatts; utilities (CEG, VST) and equipment (GEV, ETN, VRT) need signed AI load. They reinforce, but the uncrowded expression is VRT/ETN — institutional-quality, contracted backlog, far less single-author pump risk than OKLO/SMR or FCEL.
• Crypto rails + HPC landlords = c7 + c4 through dual-purpose miners (CIFR, IREN, WULF). @cantonmeow↗ and @Sam_Badawi↗ explicitly bridge these. Uncrowded expression: CIFR over MSTR — same power/rail thesis, without the financing-spread fragility @dampedspring↗ flags.
• Korean memory + Korean foundry = c1 reinforcement that 005930/000660/EWY captures with less U.S.-author crowding than MU.

Contradictions

• c5 custom silicon vs. c5 GPU scarcity: AVGO/AMD hyperscaler custom-silicon wins (c5) directly cap the "Nvidia compounds forever" version of c5. Both can't be true at the magnitudes claimed. AVGO/AMD side has more credible-author backing (@TheTranscript, @TheValueist↗, BethKindig); pure NVDA-maxi case is louder but more retail-concentrated.
• c6 AI helps SaaS vs. c6 AI destroys SaaS: DDOG/FIG/NOW (AI expands TAM) vs. WIX/ADBE/CRM (AI compresses moats). Cannot both be true — application layer either captures or loses AI value. Bull side has cleaner earnings evidence (FIG beat-and-raise, DDOG observability); bear side has @KeithMcCullough↗'s disclosed shorts and @GerberKawasaki↗ on Salesforce. Bulls have more credible-author backing right now, but the resolution is name-by-name.
• c3/c4 require sustained capex vs. c13 says capex is at melt-up extension: @TheAroraReport↗, @dampedspring↗ and @Barchart↗ warn this is dot-com-like; the capex bulls need yields and breadth to behave. @dampedspring↗ (HIGH cred) is the sharpest skeptic — and he's right that c3+c4 cannot survive a yield shock.
• c12 oil shock vs. c14 consumer turnarounds vs. c13 small-cap rotation: If Hormuz premium sticks, c14 (COST/SBUX/HD) and c13 (IWM rotation) both break. @JohnDoss1↗ explicitly trades USO as the hedge against AI longs — the bearish read on the whole AI complex.
• c7 MSTR flywheel vs. direct BTC: @dampedspring↗ and @rcwhalen↗ (both HIGH cred) argue MSTR/STRC financing is subprime-like vs. @saylor↗/@BTCtreasuries↗ treasury flywheel. Bears have higher credibility per author.

Highest-conviction trades (the 3 best ideas)

1. NBIS — long — neo-cloud HPC landlord. Q1 blew out (revenue +684% YoY, raised >4GW power guide, $7-9B 2026 ARR target), multiple PT raises (BofA $205, Citi $287), Nvidia 13F validation. The cleanest multi-source confirmation in the corpus and the trade that the per-cluster reports collectively rate highest. Best voice: @StockSavvyShay↗ ("multi-year AI compute utility").
2. VRT + ETN + GEV basket — long — institutional-quality AI power equipment. GEV raised 2026 guide on data-center orders; @wallstengine↗ and @ripster47↗ confirm. Far less crowded than OKLO/SMR/FCEL and survives even if pre-revenue nuclear disappoints. Best voice: @munster_gene↗ ("AI energy exposure has not been fully rewarded").
3. AXTI — long — InP substrate chokepoint for AI optical. @aleabitoreddit↗ (HIGH cred CPO bull) frames it as upstream of the optical reprice; survives both copper-extension and CPO-ramp paths. Less crowded than LITE/COHR but with the same structural tailwind. Best voice: @aleabitoreddit↗ ("chokepoint thesis").

Pair / spread ideas

• Long NBIS / short or underweight CRWV — same theme, but NBIS has clean multi-source earnings proof while @RealJimChanos↗ (HIGH cred) has a credible CRWV profitability short on debt-funded growth.
• Long IGV / short HD into earnings week — captures c6 software resilience while expressing the c14 consumer-stress short that @JesseCohenInv (HIGH cred) explicitly laid out as "Buy NVDA, sell HD this week."
• Long XLE+USO / short SMH or IWM — @JohnDoss1↗'s explicit hedge: if Hormuz premium sticks, AI longs and small caps both lose. Costs little if oil fades.
• Long NVO / short HIMS — c10 cleanest expression. NVO has Wegovy pill data and broader author breadth; HIMS earnings miss + crowded promotional voices (@TheLongInvest↗, @MisterInversor↗, both pump-flagged).
• Long RKLB / underweight ASTS — both space, but RKLB has clean operating proof while ASTS upside rests on three concentrated medium-cred voices (@SpacBobby↗, @AorakiTrading↗, @CKCapitalxx↗).

What's MISSING (negative-space analysis)

• No coherent Fed/rates framework despite TLT, yields and small-cap rate sensitivity being cited as the #1 risk across c13. The corpus has no NEWS_BREAK cluster on FOMC despite the obvious calendar relevance.
• No AI capex financing cycle framework — c4 landlords are raising convertibles at 1% (IREN $3B), utilities need ratebase financing, hyperscalers are levering balance sheets. No author connects these into one credit-cycle frame. @BlueJay87476298↗ hints at it but it's not a cluster.
• No copper/grid/construction bottleneck despite massive c3 power-buildout signals. Where is the FREEPORT/SCCO/electrical-construction theme? Conspicuously absent — likely an alpha-leaking gap.
• No coherent consumer-credit deterioration cluster — NU credit costs, AXP, MELI provisioning, AFRM all appear in fragments across c11 and c14 but never aggregate. Given the c12 oil shock + c14 weak consumer evidence, this is a missing risk lens.
• No China-geopolitics framework despite TSM, H200 export, Apple/China exposure threading through c1/c5/c11/c14. Treated stock-by-stock, never as a single thesis.
• No cybersecurity-as-systemic-risk theme — cyber appears only as investable software (PANW/CRWD-style), never as the risk to the AI infrastructure stack itself.

Crowded vs uncrowded

Crowded (lots of authors, all bullish, late): MU, NBIS upside calls, IREN bull case, OKLO/SMR squeeze, ASTS post-earnings recovery, HIMS rescue thesis, MSTR/STRC treasury flywheel, ZETA, FCEL, PURR/HYPE. These have visible single-author concentration and pump flags in their author briefs.

Uncrowded but credible: AXTI, VRT/ETN, CIFR over MSTR, FIG (post-print but underowned), LQDA/INSM (launch-execution, low retail crowd), Korean memory via 005930/000660 (HBM exposure without U.S. MU crowding), YPF (EM real-asset, distinct from crowded LatAm fintech), F Energy (Morgan Stanley $10B opportunity, just emerging).

Risks to the entire framework

• Yield shock / Fed surprise hike: invalidates c3+c4+c5 simultaneously by forcing an AI-capex multiple reset; @TheAroraReport↗ and @dampedspring↗ already flag dot-com parallels. This is the single highest-impact risk because it kills the reinforcing complex in one move.
• Hormuz de-escalation + oil spike combined: ambiguous but dangerous — @kshitizkapoor_↗ shows de-escalation dumps oil, while persistent disruption inflates yields and crushes c14/c13. Either path damages part of the book.
• Hyperscaler capex moderation signal (one bad guide from MSFT/META/GOOGL): would simultaneously crack c1/c2/c3/c4/c5. NVDA earnings on 2026-05-20 is the immediate test.