Odds on Open — Lihong Wang: ex-IMC semis quant, AI stack portfolio

Opening

Ex-IMC semiconductor options market maker argues foundation models are plausibly good enough to beat the S&P on a risk-adjusted basis — but only inside a hard-to-build harness — and holds a diversified ~50-name AI/chip stack book at ~2× rather than picking the winner layer. Ethan Kho interviews Lihong Wang (~74m), former discretionary semis options trader at IMC, now founder of Freeport (YC-backed perps exchange). Ground covered: how a prop MM desk makes money on flow and relative vol (NVDA/AMD/AVGO), structured-product hedging from Asian banks, correlation blowups (DeepSeek, July deleveraging), seat-vs-person leverage at top firms, agentic research workflow inspired by desk analysts, and a personal AI-investing thesis across the full supply chain. YouTube auto-captions (asr — proper nouns/numbers marked; "SK Hynix" often "Highex"/"SKHX," "Broadcom" sometimes "Brocom"). Watch.

Key takes

MM edge is counterparty/flow identification, not directional "punt" without flow. After a trade, US options exchanges often reveal who you traded against; IMC reportedly did ~20–40% of volume in some names (asr), so the desk sees who is buying/selling risk. Trade against "stupid" flow, get out of the way of or follow "smart" flow; warehouse benign risk when price is good rather than externalize everything. [asr]

Relative vol between highly correlated semis (e.g. NVDA vs AMD) is a core trade — sell rich V, hedge factor risk with the peer when needed. Historical realization and percentile pricing matter, but flow signals dominate. Asian bank structured-product hedges on long-dated NVDA create predictable ball-selling/buying as spot approaches barriers — data on issuance informs anticipated flow. [asr]

Correlation regimes break hedges: NVDA–AVGO used to trade tightly on shared "AI inference" tape; AVGO's custom-ASIC / Google TPU narrative decorrelated the pair. DeepSeek selloff cited as ~17% NVDA drop (asr — "don't quote me") with other chips also down but not identically — long-one/short-other books can blow up. July deleveraging: desks picked up upside calls when liquidation flow was identified (Intel monster guidance then sold off into deleveraging named as a tell); once a better balance-sheet buyer emerged, names pumped ~20–30% (asr) and upside vol rebid. [asr]

Foundation models are "probably good enough" to beat average retail and, with a well-designed harness, plausibly the S&P on a risk-adjusted basis — harness design is the bottleneck, not raw model IQ. Solo build in a few months: no; with a team: more confident. Pro trading firms with resources already there. Freeport's product thesis: AI agents filtering narrative sources into trader briefings the way desk analysts once shouted/emailed context — sub-second for some alerts vs slower for long research. [asr]

Personal book: ~50 "random" chip/AI names covering the whole stack because value capture layer is hard to forecast; overweight rotates via conversations with semis traders/analysts still at firms (5–6 named as dinner sources, firms unnamed). Hypothetical: if AI generates ~$10T of value in some year (asr), stack bifurcation makes single-layer bets fragile. Current narrative chatter (at recording): SoftBank/Masayoshi Son ("Leo Paul" in ASR) mechanical bid; connectivity; Korean memory (SK Hynix/Samsung) with governance/oligarchy caveat on profit diversion to shareholders. [asr]

"AI is not a bubble" = long-run aggregate compounded returns above normal — explicitly not a path statement. Dot-com analogy: some supply-chain names may be Cisco/Enron-class overvalued; Mag7-class survivors bought at the top still beat index over 20y. More "Julys" expected; fair value can still see 50–70% drawdowns on sentiment. Personal path: bottom-to-peak ~470%, then ~70% drawdown after June 22 mega-move, still ~60% off peak (asr); friends long Anthropic/OpenAI at "responsible" 2–3× up ~4–500% (asr). Optimal long-run market leverage framed ~2.25×; with AI bullish posterior, ~2× on AI stocks as his sizing. [asr]

Information hierarchy for narrative trading: Twitter leads Bloomberg/WSJ by ~1–3 days; conversation-level access at well-connected funds can lead Twitter by days/weeks; Chinese media led CNBC on DeepSeek by multiple days — weekend NVDA puts would have paid on Monday ~17% open (asr). Cites Citrini / SemiAnalysis as paid for research and access; Situational Awareness-type connectedness as edge on BTC-miner→DC contract pumps (~10–30% on announcements, asr). [asr]

Key math

IMC share of some options markets ~20–40% (asr) — flow visibility claim. [asr]

Top-five US options MMs ~80–90% of liquidity provision (asr) — seat concentration. [asr]

DeepSeek: NVDA ~−17% (asr — speaker hedge) — corr/blowup example. [asr]

July aftermath pumps ~20–30% on names + upside calls (asr) — liquidation-to-rebid path. [asr]

Personal: +470% peak → −70% off peak after June 22 → still ~−60% vs peak (asr) — path risk under leverage. [asr]

Friends: Anthropic/OpenAI at 2–3× leverage, up ~4–500% (asr) — peer sizing anecdote. [asr]

His AI-stock leverage ~2×; Kelly-style long-run market optimal ~2.25× (asr) — sizing frame. [asr]

Hypothetical AI value ~$10T in some year (asr) — why diversify across stack. [asr]

Quotes

"You trade against the people who are stupid and you get out of the way or you follow the people who are smart." — Lihong Wang [asr]

"There is a plausible argument that foundation models are good enough to beat the S&P 500 on a risk adjusted basis." — Lihong Wang [asr]

"It is difficult to tell where value will occur, which is why right now I think my personal portfolio is 50 random chip/AI stocks that pretty much covers the entire stack." — Lihong Wang [asr]

"When I say AI is not a bubble I think I mean I do believe that these companies over the long run will generate… compounded returns above normal. It is not a statement about path." — Lihong Wang [asr]

"If you bought puts on Nvidia over the weekend… on Monday open stock down 17%, you would have made like an insane amount of money." — Lihong Wang [asr] (DeepSeek / Chinese media lead)

"I ran up… 470% and then had like a 70% draw down… I'm literally down an entire Ferrari today." — Lihong Wang [asr]

Variant perception

Priced in — Semis options are flow-driven; NVDA/AMD/AVGO are the AI vol complex; DeepSeek and July were violent; Mag7/dot-com survivor framing is familiar.

What's new — Buy-side-usable map of how an IMC-class semis vol seat actually monetizes (counterparty ID, Asian SP hedges, warehouse-vs-externalize); explicit models-beat-S&P conditional on harness; 50-name full-stack book as the rational response to layer uncertainty; 2× Kelly-ish AI leverage with brutal personal drawdown honesty; information lead-times (CN media → Twitter → wires) as tradable microstructure.

Bear case — Harness claim may overstate current agent reliability for live risk; 50-name "random" book is still one-factor AI/semis beta with leverage — July-class events will recur; Korean memory governance risk could trap "bottleneck" longs; Freeport promo may color AI-workflow claims; ASR numbers (±17%, 470%/70%) are soft.

Discount — Talking his book on Freeport (perps, Hyperliquid routing, pre-IPO AI names). Ex-IMC seat nostalgia can inflate flow-edge permanence after leaving. Degenerate personal sizing undercuts the sober Kelly frame. Young show / founder guest — conviction > audited track record.

Positioning

AI capex durability — STRENGTHENS (soft). Not-a-bubble on long-run compounders; more Julys on path; full-stack demand still the bet — aligns with durability of spend even through sentiment air-pockets.

HBM supply binds — STRENGTHENS (soft). Korean memory (SK Hynix/Samsung) named in live overweight chatter alongside connectivity; governance caveat doesn't remove bottleneck framing.

Enterprise agent stall — WEAKENS (soft). Agents as research/briefing workflow already shipping in his product thesis; harness-for-alpha claim implies agents useful before full enterprise F500 production dollars.

Inference margin inversion — NEUTRAL. Stack-layer uncertainty and open vs closed not resolved into lab gross-margin arithmetic; full-stack diversification is the tell that model-layer rents are contested.