Erica · PM Research MemoAll memos

PM RESEARCH MEMO

Title: Silicon Valley’s Starving for Compute — 2026 Shortage Real; Jon Fades 2027 Scarcity Narrative
Author / source: Jon Y (Asianometry)
Source title: Silicon Valley’s Starving for Compute
Source URL: https://www.youtube.com/watch?v=x8QP9oXgahA
Video ID: x8QP9oXgahA
Channel: Asianometry (@asianometry)
Published: Sunday 13 Sep 2026 ~7:00 PM ET (yt timestamp 1789340433; publish stamp EDT)
Duration: 20:54 (1254s)
Memo date: Monday, 14 September 2026 (America/Toronto)
Transcript: /workspace/youtube-transcripts/x8QP9oXgahA.md · Brief: /workspace/youtube-transcripts/x8QP9oXgahA_brief.md · Optional JSON: /workspace/youtube-transcripts/x8QP9oXgahA.json
Caption source: YouTube automatic English ASR only (en-orig json3 via yt-dlp). No manual captions. Names, tickers, and numbers are provisional — see ASR locks below. Timestamps are cue timestamps from the retrieved timedtext.
Source type: Solo “vibes” narration after Hot Chips 2026 (Bay Area) + Semicon Taiwan trip. Channel description is promo-only (Patreon / X / Stratechery Plus newsletter / LinkedIn). This memo contains no trade recommendations from this desk.
Product: Compute-scarcity / AI-infra / packaging research map and hypotheses. Not advice. Not a recommendation to buy or sell any security.
Source discipline: Primary sources are this transcript and its brief only. Companion Jordi / Gavin Baker / Dan Niles / Jeff Keller (and related) memos on this desk are different sources — may be listed only as External / other desk companions; do not silently import numbers from those memos into this underwriting. Speaker-stated figures (incl. SemiAnalysis / Dylan Patel restatements) are used as spoken and attributed; they are not independently verified (External check needed).

How to read this document: Restatements of the talk are Source. Interpretive links and underwriting judgments are Inference. Any fact not spoken in the transcript is External check needed. Language about glut risk, rent baths, packaging winners, or frontier-lab concentration is source expression / research hypothesis, not an Erica / desk recommendation.

ASR name locks (from brief):

ASR heard Likely / note
Asianry Patreon Asianometry Patreon
silicon photonix / Photonix / for tonics silicon photonics
SKH Highix SK Hynix
AI Infest chips AI inference chips / “AI chips are so hot”
Matt X / Posetron MatX / Positron
Jeff Smith of Edge Etched (founder/Q&A crash)
Taipe's Nang / Tynx / Tinx Taipei Nangang / TaiNEX halls
showg girls showgirls
dive / retical die / reticle
Open Claw product/meme — confirm branding (OpenClaw?)
Dan Nishbal Dan Nistor? / SemiAnalysis neocloud coiner — confirm
Nebus Nebius
Dashpod / Dashesh Dwarkesh podcast (Dylan Patel appearance)
SpaceX hacks AI xAI
Grockbot GrokBot / Grok-related Cursor product
safe super intelligence Safe Superintelligence (SSI)
flapping airplanes unclear lab name — ASR mush
knee-jerk / Lenin “force of nature”? or intentional — treat carefully
gawatt / gawatts gigawatt(s)
Enthropic / Open AAI Anthropic / OpenAI
15 / 45–55 / 30–40 gawatt GW path = estimates Jon cites, not audited capacity
$50bn / $100bn per GW Dylan Patel model — not consensus street
$1.25bn/mo · 300 MW SpaceX Anthropic deal anecdote — confirm vs filings/press
Anthropics S1 Anthropic S-1
Navier Stokes Navier–Stokes Millennium claim — contested

Calendar / context (Source / brief): Year 4 of AI boom; Hot Chips 2026 then Semicon Taiwan; plug for Semicon West / Test Vision Symposium talk Oct 15, 2026 ~12:45pm. Themes: long-running agents / agent swarms, RSI (recursive self-improvement; not a risk-condition acronym here), Hugging Face hack (mentioned, not covered), epic fight for compute. Desk angle: compute scarcity / hyperscaler demand — AI + SMH. Views/likes/comments at pull (~108k / ~2.7k / ~413) are metadata only.

Thesis orientation (VERIFY against Source): Compute shortage is real in 2026 (spot gone; neoclouds; OpenAI/Anthropic ~5 GW each), but Jon fades “2 more years of scarcity” — 2027 capacity (~45–55 GW) + $50–100bn/GW revenue math don’t clear → H2’27 glut / rent-bath risk; chipmax packaging + frontier race stay hot.


EXECUTIVE SUMMARY


SOURCE-ACCURATE SUMMARY

Chapter-ordered with cue timestamps. Short quotes only where they carry the claim (≤20 words).

Frame / trip / year-four boom (00:02–01:24)

  1. (00:02–00:25) Hot Chips 2026 Bay conversations, then flight to Semicon Taiwan; first US trip in a year. Year four of AI boom; “substantial new energy.”

  2. (00:25–00:49) Drivers named: long-running AI agents and agent swarms; pursuit of RSI (recursive self-improvement — “does not refer to a risk condition”); cybersecurity and Hugging Face hack (won’t cover); “epic fight for compute.” Another vibes video on Silicon Valley and Taiwan.

  3. (00:49–01:24) Patreon plug (ASR: Asianry → Asianometry). Returning SF for Semicon West 2026 / Test Vision Symposium talk Oct 15, 2026 ~12:45pm — likely about test. Invite to watch and say hello.

Hot Chips vibes: packed house, EDA, photonics, memory (01:24–03:03)

  1. (01:24–01:58) Third year at Hot Chips. First year: Stanford auditorium warm, jetlagged sleep, back ~“60% full at most.” Today: packed; step over people for seats; “way more finance people this year too.”

  2. (01:58–02:30) Theme: forthcoming invasion of AI into EDA (points to prior video). Silicon photonics still many people present but less prominently featured than last year’s four back-to-back photonics talks.

  3. (02:30–03:03) Favorite: Micron, Samsung, SK Hynix memory session — Sunday 9am yet packed. “500% or whatever stock booms will do that for you.” Last year photonics cluster; this year memory; next year “alpha, I reckon.”

Jalapeno + custom-chip startups (03:03–04:44)

  1. (03:03–03:27) Main event: OpenAI new AI chip Jalapeno — “short and sweet,” “spicy food for thought.” Jon walks back last year’s take that the window for new AI chips closed except Google’s TPU: “I take that back now. I was wrong.” AI chips “so hot right now.”

  2. (03:27–04:02) Outside: cotery of startups — Etched, MatX, Positron, Fractile. Etched splash without presentation; “Jeff Smith of Edge” (ASR → Etched) crashes every Q&A with smart-yet-dissatisfying questions; grab-bag hats. Hardware companies: chips and racks/systems — messy lab work; impressed with execution amid talent scarcity.

  3. (04:02–04:44) Survival filter under compute shortage: Nathan Bedford Forrest paraphrase — get there first with the most. Execution in constrained supply: HBM access (and whether you should); clean tape-out; recovery if not. Then jump to Taipei Nangang / TaiNEX for Semicon Taiwan.

Semicon Taiwan as TSMC RFP; chipmax packaging (04:44–08:01)

  1. (04:44–05:56) Attending since 2022 — then chill (COVID restrictions). 2026: mayhem; TaiNEX 1 and 2 rooms filled; hype/showmanship max; showgirls + bunny-suit/fake-wafer attendees; “dream fab” simulation sold out.

  2. (05:56–06:31) Surprise for first-timers: TSMC and ASML have no booth. ASML reason unclear (Jon sees recruiting ads). TSMC dominates anyway — teams walk floor investigating vendors; info sessions headlined by TSMC execs laying out problems. Realization: “Semicon Taiwan is TSMC’s request for proposal to the whole semiconductor industry.”

  3. (06:31–07:05) Technology line: go bigger. AI chips strain limits — notably reticle limits on die size. Emerging tech breaks past limits or scales ecosystem — i.e. advanced packaging. TSMC slide: “battleship size chips spanning up to 14 reticles.”

  4. (07:05–08:01) Panel-level packaging was 2024 Semicon theme (Jon prior video). Interposers usually silicon; AI systems so large industry wants square glass panel interposers. Two years ago = talk/slideshows; warping/cracks still “believe it when I see the chips,” but floor tooling (handler robots) already accommodates square panels. Chipmaxing: one of the most far-ranging transitions “since the adoption of 300 mm wafers.” Jon is “all for it.”

Agentic demand, spot vanish, neoclouds, VC GPU hunt (08:01–09:53)

  1. (08:01–08:36) Last Bay trip Sep 2025 — different time. Long-running agentic AI took hold, triggered in part by Anthropic Opus 4.5 (~Nov 2025) and Open Claw. Feb 2026: prices rise; scramble to lock compute, led by Anthropic — big deals with SpaceX, Google, etc.

  2. (08:36–09:20) Trip time: spot market “completely vanished.” Great for neoclouds — term coined by SemiAnalysis Dan Nishbal (ASR; brief: Nistor?) Apr 2024 to replace a more profane term; pure-play resellers of AI train/infer compute; examples CoreWeave and Nebius. Believe shortage lasts “at least one to two years” with “complete certainty.” Several unrelated people trying to start neoclouds.

  3. (09:20–09:53) Startup risk: no product until mid-2027 for lack of GPUs — bad investor story. VCs’ new “help”: hunt GPUs — anecdotes of chasing ~1,000 GPUs in random places including East Europe.

Big-two feast; Google frontier fade; xAI as #3 (09:53–11:33)

  1. (09:53–10:13) Contrast: Anthropic and OpenAI “are feasting.” Dylan on latest Dashpod (Dwarkesh): two giants have about 5 GW of compute each (phrasing: “about 5 GW of compute” for the two — brief locks ~5 GW each). Embarrassment of riches; anecdotal inefficiency.

  2. (10:13–10:58) Question Jon poses: will Google return to the frontier? Broad agreement: no — explanations “squishy.” Selling lots of compute to Anthropic ≈ saying they don’t believe it either. Turnaround from Gemini 3 “OpenAI was cooked” moment. Talent loss + culture not conducive to frontier AI — leave to future historians.

  3. (10:58–11:33) If not Google, who joins Anthropic + OpenAI at US frontier (China set aside)? Consensus: “SpaceX hacks AI” → xAI. Elon doesn’t like losing; spent more buying Cursor than building rocket technology (ASR); Meta-style recruiting; Cursor guys behind recent GrokBot — “pretty impressive,” though Jon hasn’t used it.

Neolabs, ROI fog, SF work paradox (11:33–15:53)

  1. (11:33–12:40) Neolabs: pre-revenue research startups to explore intriguing AI ideas; elite researcher founders; little biz plan beyond discovering a new structure. Famous: Safe Superintelligence raised a billion, no product/model yet. Also: Recursive Intelligence (name collision), Humans&, Core Automation, “flapping airplanes” (ASR). Labs with product: Sakana (Japan), Thinking Machines, Reflection. Domain: Physical Intelligence, Periodic Labs, Generalist AI (materials / world models / robotics).

  2. (12:40–14:23) Knee-jerk: asset class “dumb.” Counter: historically universities / gov labs / Bell Labs path broken — funding cuts, talent raided; AI first major tech in ~50 years without government involvement; scaling-era research closed/secret. So Nvidia (compute+cash) + VCs fund asymmetric bets; OpenAI/Anthropic locked into transformer-centric paradigm. Failure likely but funding often in compute; buyout if gold-mine architecture. Better than Nvidia buybacks — Apple jab. “Low-key love these companies” — Silicon Valley awesome (ASR “force of Lenin” — treat carefully).

  3. (14:23–15:32) Token ROI revisited: “remains nebulous… Nobody has measured it. We’re all running on anecdotes.” Coding barriers falling — seen: two young analysts + Claude coding a dashboard (vs old armies of reporting analysts). Spreadsheets/presentations: less noticeable impact so far. OpenAI blog: agents accelerated research but pegged like an intern; far from AI researcher; lacks human “research taste” — maybe enough scale.

  4. (15:32–15:53) SF paradox: everyone works harder than ever. “Maybe rather than automating away work, AI just gives us more work to do.”

2027 glut math — core of the fade (15:53–19:14)

  1. (15:53–16:15) Everyone saying shortage for next 2 years — Jon wonders if wrong. Math “not super sophisticated”: “there’s too much supply hitting in 2027.”

  2. (16:15–16:50) Estimate: EOY 2026 usable AI DC capacity ~15 GW; OpenAI and Anthropic ~5 GW each of that. Supply chain responds → surge to ~45–55 GW EOY 2027; ~30–40 GW incremental. OpenAI + Anthropic share ~30 GW of that; assume half training → 15 GW inference between the big two.

  3. (16:50–17:24) Each GW of inference modeled ~$50bn revenue today — from Dylan Patel on Dwarkesh; also projectable from Anthropic’s $1.25bn/month deal for SpaceX’s 300 MW. Dylan: smarter models could push to $100bn per GW. Apply → $750bn–$1.5tn combined 2027 revenue for both. “Brother, that is a lot.”

  4. (17:24–18:07) Most agree big-two ~$150–200bn this year (maybe lower). Anthropic S-1 “literally any second this or next week” for color. Jon can see $400–500bn next year — not 750bn–1.5tn. Framing: two Walmarts high / two Toyotas low — “strains credibility.” Even if models cure cancer / find superconductors tomorrow, bets revenue number is wrong — far lower than $100bn or even $50bn per GW, “probably a fraction.”

  5. (18:07–18:40) Infra lens: say 40 GW total hit in 2027; current rents ~$20m per MW (~3-year payback) → those 40 GW need $800bn paid to infra providers (internal+external) just for 2027. Doesn’t think they get there → “rental prices must fall and someone somewhere probably takes the bath.”

  6. (18:40–19:14) Boil-down: supply chain brings lots of compute in 2027, most in H2, tips shortage → glut. Not saying AI bubble popping — “soon be awash in compute.” 2026 = agentic demand era with late-2025 supply; supply catches up and probably overshoots. Soft: next phase (e.g. “Persistent AI perhaps”) may swell demand again.

Thanks / Navier–Stokes / anti-AI politics (19:14–20:53)

  1. (19:14–19:47) Thanks interviewees. Prior Hot Chips vibes video (“Silicon Valley doing hard things again”) mentioned prediction AI solves a Millennium problem within a year. As of writing: advanced internal OpenAI model produced proof for Navier–Stokes existence and smoothness — friend “got it right on the dot.”

  2. (19:47–20:20) Achievement marred by plagiarism and data theft accusations. Groundswell of negativity → rising anti-AI sentiment among ordinary people: hate data centers; hate billionaires; think it steals “real” human knowledge; don’t think society needs it.

  3. (20:20–20:53) Sentiment growing; bad feeling it “can win elections.” SF operates in a bubble / microclimate — people inside should watch feelings outside; “metastasize with scary speed.” Dark joke on AI extinction fears. Close: subscribe / Patreon.


SYSTEMS MAP / VALUE CHAIN ANALYSIS

Layer 0 — Demand shock (2025→2026). Opus 4.5 + Open Claw → agentic / long-running workloads → Feb 2026 lock-up scramble → Anthropic multi-party deals (SpaceX, Google, …) → spot GPU market vanishes. Demand side still anecdote-heavy on ROI (intern-class agents; nebulous token ROI).

Layer 1 — Frontier compute concentrators. OpenAI + Anthropic as ~5 GW each “feast” nodes; inefficiency anecdotes. Google as compute seller to Anthropic (frontier skepticism). xAI / Cursor / GrokBot as consensus US #3 path. China explicitly set aside in this talk.

Layer 2 — Neocloud / reseller layer. CoreWeave, Nebius, and would-be founders; SemiAnalysis “neocloud” taxonomy (Apr 2024). Pricing power while spot is gone; narrative certainty of 1–2 years more shortage — the consensus Jon fades.

Layer 3 — Constrained startup / VC layer. Mid-2027 product slips for lack of GPUs; VCs redeploy as GPU hunters (1k-GPU East Europe scraps). Survival filter for custom-chip startups: HBM + tape-out + rack/system execution (Etched / MatX / Positron / Fractile; OpenAI Jalapeno).

Layer 4 — Capacity build path (Jon’s critical path). Usable AI DC: ~15 GW EOY’26 → ~45–55 GW EOY’27 (+30–40). Big-two take ~30 GW of increment; ~15 GW inference share. Bottleneck timing: most new supply H2’27 → tip risk.

Layer 5 — Unit economics (the fade). Inference rev/GW: Dylan $50bn today → $100bn if smarter; Anthropic–SpaceX $1.25bn/mo for 300 MW as cross-check. Implied big-two 2027 rev $750bn–$1.5tn vs Jon’s stretch $400–500bn and street-ish $150–200bn this year. Infra: ~$20m/MW rent × ~40 GW → ~$800bn 2027 landlord bill. Falsifier of scarcity-into-2028 narrative: either $/GW or MW rents (or both) fail to clear.

Layer 6 — Packaging / chipmax (structural tightness independent of rent). Reticle limits → multi-reticle “battleship” (up to ~14) → advanced packaging → glass panel interposers + handler tooling. TSMC as silent RFP owner of Semicon Taiwan; ASML absent-booth curiosity. Memory/HBM (Micron / Samsung / SK Hynix Hot Chips fire) as co-bottleneck with packaging.

Layer 7 — Neolab / R&D optionality. Nvidia cash + VC compute grants → SSI / Sakana / Thinking Machines / Reflection / domain labs. Asymmetric architecture bets outside transformer lock-in; buyout payoff path. Parallel to — not substitute for — GW rental math.

Layer 8 — Political / social thermostat. Anti-AI sentiment (data centers, billionaires, IP theft narrative, “don’t need it”) → election risk; SF bubble. Navier–Stokes claim + plagiarism accusations as cultural flashpoint. Can cut supply (permitting/power) or demand (regulation/tax/liability) orthogonal to SemiAnalysis GW paths.

Bottlenecks (Source): spot GPUs; HBM; tape-out quality; reticle/die size → packaging; glass panel warpage/cracks; talent for systems companies; Anthropic/OpenAI utilization efficiency.
Bottlenecks (Inference): whether 45–55 GW is physical power + interconnect + cooling or slideware; whether $50–100bn/GW ever existed outside a few peak deals; whether H2’27 supply is financeable if rents already soften in H1; whether anti-AI politics hits before glut math.

Companion External / other desk (names only — no number merge): prior Jordi / Gavin / Niles / Keller compute-scarcity or hyperscaler tapes may discuss related themes; this underwriting uses Asianometry + brief only.


SECOND AND THIRD-ORDER EFFECTS

Chain A — Spot vanish → neocloud certainty → 2027 overbuild (Jon’s main chain)

  1. First order (Source): Spot market gone; neoclouds certain of 1–2 more years of shortage; new neocloud formation wave.
  2. Second order (Inference): Capex and PPAs underwritten on scarcity-duration assumptions that Jon rejects; landlords and leveraged neoclouds most exposed to H2’27 rent reset.
  3. Third order (Inference): Soft glut without “AI bubble pop” — utilization and $/MW take the bath first; equity narratives lag rental indices. Watch: spot/neocloud listed rates, contract tenor, secondary GPU listings — not just GW announcements. Falsifier: demand re-acceleration (“persistent AI”) absorbs H2’27 supply before rents crack.

Chain B — $50–100bn/GW model fails → frontier ROI gap → landlord vs lab split

  1. First order (Source): Dylan $/GW + SpaceX 300 MW anecdote imply $750bn–$1.5tn big-two 2027 rev; Jon says no — fraction of $50bn/GW more likely.
  2. Second order (Source/Inference): Even cancer-curing models don’t rescue the revenue-per-GW assumption; Anthropic S-1 is near-term color.
  3. Third order (Inference): If labs monetize below model but keep training race, training share of GW stays sticky while inference ASP softens — mixed signal for GPU OEMs vs pure rental plays. Packaging/HBM can remain tight (training die size) while cloud rental softens. Research split, not a rec.

Chain C — Chipmax / glass panel → multi-year tooling cycle

  1. First order (Source): TSMC RFP for bigger-than-reticle / battleship / panel / glass; floor tooling already moving.
  2. Second order (Inference): Capex cycle for handlers, glass, panel lines can run through a GPU-rent bath — different duration than neocloud leases.
  3. Third order (Inference): If glass yields disappoint, reticle-limit constraint extends scarcity for frontier dies even as commodity GPU racks glut — barbell tightness. Falsifier: rapid glass yield success + H2’27 rack flood simultaneously clears both layers.

Chain D — Custom silicon reopen + HBM filter → startup attrition

  1. First order (Source): Jalapeno + Etched/MatX/Positron/Fractile; Jon walks back “window closed.”
  2. Second order (Source): Survival = HBM access + clean tape-out + systems execution under shortage.
  3. Third order (Inference): 2026–27 attrition concentrates winners; survivors may be buyout targets for hyperscalers/labs if architecture hits (same logic as neolabs). Glut year can kill late tape-outs that miss the scarcity window. Watch: HBM allocation anecdotes, tape-out news, rack demos — not valuation chatter.

Chain E — Anti-AI politics → permitting / election risk

  1. First order (Source): Rising anti-AI sentiment; can win elections; SF bubble warning; Navier–Stokes plagiarism cloud.
  2. Second order (Inference): Local data-center opposition and national politics can slow supply (ironically extending scarcity) or tax/regulate demand.
  3. Third order (Inference): Political path can invalidate both bull “2-year scarcity” and bear “clean H2’27 glut” if supply is blocked unevenly by jurisdiction. Underwatched vs SemiAnalysis GW slides. Horizon: election cycles / state permitting, not just 2027 capacity tables.

SCENARIO FRAMEWORK

Probabilities below are analyst inference for research prioritization, not forecasts Jon assigned, and not desk allocations.

Bull (Inference ~20–25%) — Scarcity extends; $/GW holds; chipmax tight

Path: H2’27 capacity slips (power, interconnect, packaging, glass yield); agentic + “persistent AI” demand absorbs what does arrive; Anthropic S-1 prints revenue trajectory closer to Jon’s $400–500bn stretch or better; rents near $20m/MW hold; OpenAI/Anthropic/xAI keep feasting; custom silicon survivors clear HBM.

Market expression (research hypotheses, not recs): Neocloud / GPU rental narratives stay firm; memory/HBM and advanced packaging stay scarcity-priced; frontier-lab private marks supported. Not a desk long.

What would need to be true: Primary capacity data show 45–55 GW misses; rental indices stay elevated; S-1 / lab disclosures support high $/GW.

Base (Inference ~45–55%) — Jon’s map: 2026 starve → H2’27 soft glut / rent bath; packaging stays structurally interesting

Path: 2026 remains hard for startups (spot gone); neocloud certainty proves too confident; incremental GW arrives mostly H2’27; big-two revenue clears below $750bn–$1.5tn (Jon: maybe $400–500bn stretch, not the Dylan high end); rents fall; someone (levered landlord / late neocloud / overbuilt colo) takes a bath; not full AI-bubble pop; chipmax/panel tooling cycle continues; soft landing if next demand wave ignites later.

Market expression: Differentiate rental/neocloud duration risk from packaging/HBM/tooling and from frontier-lab private assets; watch Anthropic S-1 for which $/GW regime is live.

What would need to be true: Capacity additions roughly as cited; inference monetization lags $50–100bn/GW; utilization/rent data soften before equity narratives do.

Bear (Inference ~20–30%) — Harder landing: glut + ROI fog + politics

Path: Supply overshoots and token ROI stays nebulous (intern-class agents); big-two revenue closer to or below street-ish run-rates without stretch; rents crack hard; inefficiency at 5 GW nodes becomes forced digests; anti-AI politics wins meaningful elections / permitting freezes unevenly; Navier–Stokes plagiarism narrative fuels broader legitimacy crisis; custom-chip startups miss tape-outs into a glut.

Market expression: Broad AI-infra risk-off that does not spare packaging names if capex budgets freeze; neolab funding drought. Not a desk short.

What would need to be true: Joint failure of $/GW and political license and H2’27 supply hit; S-1 / disclosures disappoint vs even conservative paths.

Invalidation markers for Jon’s fade (Inference): (i) EOY’27 usable capacity clearly <<45 GW; (ii) $/GW evidence sustains near $50bn+ on primary deals; (iii) rents hold ~$20m/MW through 2027; (iv) “persistent AI” demand visibly books H2’27 supply in advance; (v) Anthropic S-1 forces upward revision of big-two trajectory toward the Dylan high case.


COMPANY/ASSET WATCHLIST

No buy/sell. No target weights. Monitoring list only — names appear because the source discussed them.

Cluster Names / tickers as spoken Why on the list (Source) Diligence hook
Frontier labs OpenAI (Jalapeno; ~5 GW; Navier–Stokes claim), Anthropic (~5 GW; Opus 4.5; SpaceX/Google deals; S-1 imminent) Feast nodes; revenue path S-1; deal filings; utilization anecdotes
Hyperscaler / cloud Google (TPU; sells compute to Anthropic; frontier skepticism), SpaceX (deal counterparty; xAI adjacency) Frontier map; compute allocation Primary deal terms; Gemini trajectory
Contender #3 xAI; Cursor; GrokBot Consensus US #3 path Product quality; talent/comp
Neoclouds CoreWeave; Nebius; unnamed would-be founders Scarcity-duration consensus Jon fades Spot/list rates; lease tenor; leverage
Custom silicon Etched; MatX; Positron; Fractile; OpenAI Jalapeno Window reopened; HBM/tape-out filter HBM supply; tape-out; rack demos
Foundry / equipment TSMC (RFP owner; battleship ≤14 reticles); ASML (no booth curiosity) Chipmax / packaging roadmap Packaging capacity; glass panel yields
Memory Micron; Samsung; SK Hynix Hot Chips packed memory fire HBM allocation vs GPU rack glut
Neolabs / labs SSI ($1bn, no product); Sakana; Thinking Machines; Reflection; Physical Intelligence; Periodic Labs; Generalist AI; Recursive Intelligence / Humans& / Core Automation / “flapping airplanes” Asymmetric R&D; Nvidia+VC funding logic Product milestones; compute burn
Commentary / data Dylan Patel / SemiAnalysis; Dwarkesh (“Dashpod”); Dan Nistor (neocloud coinage) GW path; $/GW; taxonomy Primary podcast/filings vs ASR
Ecosystem cash Nvidia (compute+cash funder of neolabs; buyback jab foil), Apple (buyback foil) Funding asymmetric bets Capex vs return-of-capital optics
Shows / calendar Hot Chips 2026; Semicon Taiwan / TaiNEX; Semicon West / Test Vision Oct 15, 2026 ~12:45pm Vibes / packaging RFP / Jon talk Floor tooling photos; session agendas
Political / social Anti-AI sentiment; data-center opposition; election risk; Navier–Stokes plagiarism narrative Tail risk to supply or demand Polling; permitting; primary proof claims

Explicitly not a recommendation list. Appearance ≠ endorsement. Do not merge names/numbers from Jordi / Gavin / Niles / Keller companions into sizing.


DILIGENCE QUESTIONS & RESEARCH AGENDA

  1. Capacity path: Primary source for ~15 GW EOY’26 and ~45–55 GW EOY’27 usable AI DC — SemiAnalysis note, Dylan podcast timestamp, or other? Audited vs estimated? Power vs IT load definition? (External check needed.)

  2. Big-two GW: Confirm OpenAI + Anthropic ~5 GW each as of Dylan’s Dashpod appearance — date of claim vs Jon’s trip; training vs inference split.

  3. Anthropic–SpaceX deal: Verify $1.25bn/month for 300 MW — press, filing, or podcast-only? Implied $/MW and vs $50bn/GW math consistency.

  4. Anthropic S-1: Timing (“this or next week” as of recording) and disclosed revenue / compute / commitments — first color event for Jon’s $150–200bn / $400–500bn frames.

  5. Rental indices: Independent series for ~$20m per MW and 3-year payback claim — which markets (US regions), powered shell vs turnkey, contract length?

  6. Spot market: Evidence that spot “completely vanished” at trip time — which venues (Vast, cloud spot, brokers)? Any reappearance by memo date?

  7. Neocloud coinage: Confirm Dan Nistor (vs ASR Nishbal) Apr 2024 SemiAnalysis piece; CoreWeave/Nebius utilization and guidance language on shortage duration.

  8. Packaging / glass: TSMC public slides on ≤14-reticle battleships; panel/glass supplier tooling at Semicon Taiwan 2026 — yield/warpage status beyond Jon’s “believe it when I see the chips.”

  9. Jalapeno + startups: OpenAI Hot Chips Jalapeno disclosures; Etched/MatX/Positron/Fractile funding, HBM access, tape-out status — survival filter evidence.

  10. Open Claw / Opus 4.5: Confirm branding and Nov 2025 trigger narrative for agentic demand wave.

  11. Navier–Stokes claim: Primary OpenAI/internal claim vs plagiarism/theft accusations — treat as contested; not underwritable science alpha.

  12. Companion hygiene: List Jordi / Gavin / Niles / Keller scarcity memos as External / other desk only; do not import their GW, rent, or FCF figures into this file’s underwriting.


RISK ANALYSIS

Source / process risks

Market / thesis risks

Desk product risks

Risk map summary (Inference)

Risk Likelihood (to H2’27) Severity if hit Mitigant / watch
H2’27 rent bath (Jon base) Medium–High (his map) High for levered landlords/neoclouds Primary rental indices; lease resets
Capacity slip keeps scarcity Medium High for fade thesis GW build tracking; power interconnect
$/GW holds near Dylan Low–Medium High for fade thesis S-1; disclosed deals
Persistent AI demand surge Medium Softens glut Workload telemetry; lab capex guides
Glass/panel yield fail Medium Extends frontier-die scarcity TSMC/supplier yield disclosures
Anti-AI politics Medium Asymmetric / jurisdiction-dependent Permitting; polls; legislation
ASR/number error High (process) Medium for memo quality External checks in diligence list

APPENDIX — Unverifiable / provisional numbers (flag for External check)

All speaker-stated or brief-restated; not verified by this desk from primary data in this memo:

Claim As spoken / brief Status
Publish stamp Sun Sep 13, 2026 ~7:00 PM ET (ts 1789340433) Metadata from pull
Duration 20:54 (1254s) Metadata
Views / likes / comments ~108,389 / ~2,682 / ~413 Pull-time only
Prior Hot Chips fill ~60% full Anecdote
Memory “500%” stock booms Soft / rhetorical External
TSMC battleship reticles Up to ~14 Slide claim — External
Opus 4.5 timing ~Nov 2025 External
Spot market “Completely vanished” at trip Anecdote — External
Neocloud coinage Dan Nishbal/Nistor Apr 2024 External
Shortage duration (neocloud view) ≥1–2 years Consensus claim Jon fades
GPU hunt scraps ~1,000 GPUs; East Europe Anecdote
OpenAI + Anthropic compute ~5 GW each Dylan via Jon — External
EOY 2026 usable AI DC ~15 GW Estimate — External
EOY 2027 usable AI DC ~45–55 GW Estimate — External
Incremental 2027 ~30–40 GW Arithmetic from above
Big-two share of increment ~30 GW Estimate — External
Inference share (half) ~15 GW between big two Assumption
Rev per GW inference ~$50bn; up to ~$100bn Dylan model — External
Anthropic–SpaceX $1.25bn/mo for 300 MW Anecdote — External
Implied big-two 2027 rev $750bn–$1.5tn Arithmetic — incredible per Jon
Street-ish this year ~$150–200bn big-two “Most people agree” — External
Jon stretch next year ~$400–500bn Opinion
Rent ~$20m per MW; ~3y payback External
Infra bill 2027 ~$800bn on ~40 GW Arithmetic
SSI raise $1bn; no product yet External
Semicon West talk Oct 15, 2026 ~12:45pm Calendar confirm
Navier–Stokes proof Internal OpenAI model claim Contested — External

CLOSING (for packet routing)

Stance one-liner (Source expression, not desk advice): Shortage is real in 2026 (spot gone; neoclouds; OAI/Anthropic ~5 GW feast), but fade the “2 more years of scarcity” consensus — 2027 capacity (~45–55 GW) plus $50–100bn/GW revenue math don’t clear → H2’27 glut / rent-bath risk; stay constructive on chipmax packaging and the frontier race; watch Anthropic S-1 and anti-AI politics as tails.

This memo: transcript + brief only · ASR caveats locked · companions External without number merge · no buy/sell · markdown file only · not published to here.now.


End of memo.

Desk copy · not a trade recommendation · Erica · 14 Sep 2026