U.S. keeps the AI frontier and compute chokepoints

## Assumptions I will use - **Thesis:** “The U.S. wins the AI race.” - **Investment horizon:** 18–30 months, with a tactical checkpoint every month and a hard review on **December 31, 2026**, because many live prediction markets resolve around then. - **Risk budget:** model this as a **$100,000 portfolio** with a **$15,000 planned-loss budget**. I’ll give a strict defined-risk version and a higher-delta version. - **Definition of “U.S. wins”:** U.S.-based or U.S.-aligned firms keep the frontier-model lead, U.S./allied semiconductor and cloud infrastructure remain the bottleneck, China remains competitive on cost/open-source diffusion but does not capture the majority of frontier AI rents. - **Constraint:** no assumption that “AI goes up forever.” The market is already paying attention to AI capex, and that matters. --- # Phase A: UNDERSTAND ## 1. Thesis Parser Your thesis decomposes into four sub-theses: 1. **Frontier capability lead:** U.S. labs — Google, Anthropic, OpenAI, Meta, xAI, etc. — stay ahead at the frontier. 2. **Compute chokepoint:** U.S./allied semiconductor designers, foundries, equipment suppliers, memory suppliers, and cloud platforms remain the scarce assets. 3. **Policy moat:** U.S. export controls and allied supply-chain coordination prevent China from cheaply accessing the same frontier compute stack. 4. **Profit capture:** the winners are not just model labs, but the listed infrastructure layer: semis, cloud, data centers, power, networking, memory, and defense/enterprise AI. The current evidence supports a **qualified** version of this thesis: J.P. Morgan says the U.S. still leads at the frontier, while China has narrowed the gap and competes aggressively on cost, openness, and distribution. That means the edge is not “U.S. AI good, China AI bad”; it is “U.S. controls the premium frontier stack, China pressures margins at the low-cost/open-source layer.” (am.jpmorgan.com) ## 2. Scenario Formalizer I’d formalize the scenario tree like this: | Scenario | Probability I’d assign | What happens | Trade implication | |---|---:|---|---| | **A. U.S. decisive win** | 45% | U.S. labs dominate frontier benchmarks; U.S./allied compute remains scarce; China competes but mostly below frontier | Long U.S. semis/cloud/power; short or hedge China AI catch-up | | **B. U.S. wins, but AI monetization disappoints** | 25% | U.S. keeps technical lead but capex/FCF pressure causes drawdowns | Use call spreads, not just spot; hedge bubble risk | | **C. Bipolar / cheap-China world** | 20% | Chinese open-weight models commoditize inference and reduce pricing power | Avoid overpaying for pure model margin; own picks-and-shovels | | **D. China catch-up / U.S. lead breaks** | 10% | China reaches frontier, export controls leak or fail, U.S. valuations compress | Stop-loss, China hedge, defined-risk options | The important point: **the market already prices a lot of U.S. AI leadership.** Polymarket’s “Which company has best AI model end of 2026?” currently prices **Google at about 59%** and shows Anthropic as the main alternative, while U.S. names dominate the visible odds set. (polymarket.com) So the cleanest trade is not simply “buy U.S. tech”; it is **own U.S. AI bottlenecks with convexity and hedge valuation/capex risk**. --- # Phase B: MAP THE BUTTERFLY EFFECT ## 3. Consequence Graph **Thesis shock:** U.S. AI lead widens → frontier models require more high-end compute → hyperscalers continue capex → semis, HBM/memory, networking, power, and data-center infrastructure benefit → U.S. and allied suppliers capture economic rents → China faces higher cost of frontier compute → Chinese model firms compete on efficiency/open-source but have weaker premium economics → U.S. AI stack attracts capital, talent, sovereign partnerships, and enterprise adoption. Second-order effects: - **Semis:** NVDA, AVGO, AMD, MRVL, MU, SMH/SOXX. - **Cloud/model platforms:** MSFT, GOOGL, AMZN, META, ORCL. - **Power grid / electrification:** GEV, CEG, VST, ETN. - **Defense / sovereign AI:** PLTR, defense primes, government cloud suppliers. - **China hedge:** KWEB, FXI, BABA, BIDU, China AI names where available. - **Prediction-market proxies:** U.S. lab benchmark markets, Chinese top-model markets, export-control markets. The capex side is very real: hyperscalers including Alphabet, Amazon, Meta, Microsoft, and Oracle have reportedly raised **$326 billion so far in 2026** through stock and bond issuance to fund AI expansion. (axios.com) Amazon also reported sharply negative trailing free cash flow tied to AI-related property and equipment purchases, with expected 2026 cash capex of **$220 billion**, showing the market’s concern that winning the AI race may still be expensive. (axios.com) ## 4. Instrument Mapper | Consequence | Public-market expression | Derivative / event expression | |---|---|---| | U.S. frontier compute stays scarce | NVDA, AMD, AVGO, MRVL, SMH, SOXX | SMH/NVDA/AVGO call spreads | | HBM/memory demand rises | MU, SK Hynix proxies, DRAM baskets | Hyperliquid MU, SKHX/SKHY, DRAM perps | | Cloud/platforms capture enterprise AI | MSFT, GOOGL, AMZN, META, ORCL | GOOGL/MSFT call spreads | | Data centers strain power grid | GEV, CEG, VST, ETN | Spot basket, not options-heavy | | China remains below frontier | Short/put KWEB, FXI, BABA, BIDU | Polymarket Chinese top-model “No” markets, Hyperliquid China AI shorts | | AI capex bubble risk | Smaller premium, call spreads, cash | Polymarket “AI bubble burst” Yes as hedge | | U.S.–China policy split widens | Long U.S. AI stack / short China stack | Polymarket export-control / U.S.–China AI markets | --- # Phase C: ESTABLISH THE EDGE ## 5. Market Baseline / Pricing Current market baseline is **bullish but nervous**: - NVDA is trading around **$213.66**, with a market cap above **$5.1T** and a P/E around **26.8**. - MSFT is around **$530.14**, market cap around **$3.95T**, P/E around **27.7**. - GOOGL is around **$283.82**, after a large daily decline, with a P/E around **14.8**. - AMD is much more valuation-sensitive, around **$566.42** with a P/E above **145**. - PLTR is also very expensive, around **$196.42** with a P/E around **168**. So: **the edge is not in discovering AI. The edge is in selecting the parts of the U.S. AI stack where the payoff is convex and the risk is bounded.** Prediction markets also show that “U.S. frontier lead” is partly priced. Polymarket’s AI page lists **hundreds of AI markets**, and its visible top 2026 model market currently assigns Google the leading position. (polymarket.com) It also prices an **“AI bubble burst by June 30, 2027”** outcome around **23%**, which is important because “U.S. wins technically” and “equity holders win financially” are not identical. (polymarket.com) ## 6. Historical Analog Finder Best analogs: 1. **Space race / Cold War tech procurement:** government and national-security competition can sustain spend even when near-term ROI is unclear. 2. **Internet buildout:** the infrastructure layer can boom first, but overcapacity and valuation compression can still hurt investors. 3. **Semiconductor strategic cycles:** countries try to localize supply chains, but the existing leaders often keep compounding because know-how, tooling, ecosystems, and capital intensity are hard to replicate. 4. **Energy infrastructure cycles:** power bottlenecks can become investable second-order winners even when the end application is uncertain. Conclusion from analogs: **own chokepoints, not hype. Use spreads, not unlimited downside. Expect violent drawdowns.** ## 7. Response / Beta Estimator Approximate thesis sensitivity: | Asset / instrument | U.S.-wins beta | AI-bubble risk | China-catch-up risk | Comment | |---|---:|---:|---:|---| | NVDA | Very high | High | High | Best pure bottleneck, but crowded | | SMH | High | High | Medium-high | Cleaner diversified semis exposure | | AVGO | High | Medium | Medium | Custom silicon/networking angle | | MSFT | Medium-high | Medium | Medium | Cloud + enterprise AI | | GOOGL | Medium-high | Medium | Medium | Frontier model + cloud; cheaper P/E than many AI names | | AMD | High | High | High | More upside beta, more valuation risk | | GEV / CEG / VST | Medium | Medium | Low-medium | Power/data-center second derivative | | PLTR | High | Very high | Medium | Sovereign/defense AI, but valuation risk | | KWEB / FXI puts | Negative China beta | Hedge | Loses if China rallies | Useful as hedge, not core | | Polymarket AI markets | Direct event beta | Binary | Binary | Good for small, precise views | | Hyperliquid perps | High tactical beta | Liquidation risk | Liquidation risk | Use small isolated margin only | ## 8. Trade Backtester I ran a simple scenario P&L model for a **higher-delta $100k implementation**: - $67k spot U.S. AI basket. - $10k U.S. AI call-spread premium. - $5k China-tech put-spread premium. - $18k cash / T-bills. This is **not a historical price backtest**; it is a scenario backtest based on the thesis shocks above. | Scenario | Spot P&L | U.S. call-spread P&L | China put-spread P&L | Total P&L | Return on $100k | |---|---:|---:|---:|---:|---:| | Decisive U.S. win | +$27,500 | +$15,000 | +$4,000 | **+$46,500** | **+46.5%** | | Muddle | +$6,320 | -$4,000 | -$3,000 | **-$680** | **-0.7%** | | China catches up | -$19,800 | -$10,000 | -$5,000 | **-$34,800** | **-34.8%** | | AI bubble burst | -$28,950 | -$10,000 | +$3,500 | **-$35,450** | **-35.5%** | | Taiwan/geopolitical shock | -$11,140 | -$8,000 | +$3,000 | **-$16,140** | **-16.1%** | Takeaway: the **higher-delta basket has excellent upside but violates the 15% loss budget in adverse scenarios**. Therefore the primary recommendation should be a **defined-risk barbell**, with the higher-delta spot basket used only if you accept drawdowns above 15%. --- # Phase D: BUILD THE TRADE ## 9. Expression Selector Best expression: > **Long U.S. AI bottlenecks through call spreads and selective spot; hedge with China-tech downside and bubble-risk hedges; use prediction/perp markets only as small tactical overlays.** Avoid: - Unhedged all-in NVDA. - Levered SOXL as a long-term holding. - Large perpetual futures positions. - Pure model-lab prediction markets unless the odds are attractive. ## 10. Basket Optimizer ### Recommended strict-risk version: $100k model portfolio This version keeps the **maximum planned loss near $15k**. | Sleeve | Allocation | Instrument | |---|---:|---| | Cash / T-bills | $85,000 | Dry powder and risk control | | U.S. AI call-spread premium | $8,000 | SMH, NVDA, GOOGL, MSFT, AVGO call spreads | | China-tech hedge premium | $3,000 | KWEB / FXI / BABA put spreads | | Polymarket event sleeve | $1,500 | U.S. lab / China AI / AI bubble markets | | Hyperliquid tactical sleeve | $2,500 | Isolated-margin perps only | ### Higher-delta equity basket, if you accept larger drawdowns Reference spot basket using current prices: | Ticker | Role | Target $ | Current ref price | Approx shares | |---|---|---:|---:|---:| | SMH | Diversified semis | $20,000 | $561.67 | 35.6 | | NVDA | Frontier compute | $8,000 | $213.66 | 37.4 | | AVGO | Networking/custom silicon | $8,000 | $345.00 | 23.2 | | MSFT | Cloud/enterprise AI | $7,000 | $530.14 | 13.2 | | GOOGL | Frontier model + cloud | $7,000 | $283.82 | 24.7 | | AMD | Compute challenger | $4,000 | $566.42 | 7.1 | | GEV | Power infrastructure | $4,000 | $887.85 | 4.5 | | CEG | Nuclear/power | $3,000 | $264.58 | 11.3 | | VST | Power/data-center electricity | $3,000 | $140.95 | 21.3 | | PLTR | Defense/sovereign AI | $3,000 | $196.42 | 15.3 | Prices referenced from current market snapshots for SMH, NVDA, AVGO, MSFT, GOOGL, AMD, GEV, CEG, VST, and PLTR. ## 11. Risk & Sizing Rules: 1. **Max single-name spot exposure:** 10% of NAV, except diversified ETFs. 2. **Max options premium at risk:** 10–12% of NAV. 3. **Max prediction-market exposure:** 2% of NAV. 4. **Max Hyperliquid isolated margin:** 2.5–3% of NAV. 5. **Hard thesis stop:** cut risk by 50% if Chinese models hold frontier #1 across multiple major benchmarks for two consecutive months. 6. **Financial stop:** cut risk by 50% if the equity sleeve loses 15% from entry or underperforms QQQ by 12 percentage points. 7. **Capex stop:** reduce cloud/platform exposure if hyperscaler capex keeps rising while backlog/revenue growth fails to accelerate. 8. **No cross-margin perps.** Use isolated margin only. --- # Phase E: ACT ## 12. Trade Proposal / Execution Layer ### Step 1 — Enter the defined-risk U.S. AI call-spread sleeve Use 18–24 month expiries where possible. Suggested premium split: | Underlying | Premium share | Rationale | |---|---:|---| | SMH call spreads | 35% | Diversified U.S./allied semi exposure | | NVDA call spreads | 25% | Highest bottleneck beta | | GOOGL call spreads | 15% | U.S. frontier lab exposure at less extreme P/E | | MSFT call spreads | 15% | Enterprise/cloud monetization | | AVGO call spreads | 10% | Networking/custom silicon | Execution rule: only buy call spreads where the debit is **≤35–40% of spread width**. If spreads are too expensive, wait. ### Step 2 — Add China-tech downside hedge Use put spreads, not naked shorts. | Instrument | Premium share | Rationale | |---|---:|---| | KWEB put spreads | 60% | China internet / AI-app ecosystem hedge | | FXI put spreads | 25% | Broader China large-cap hedge | | BABA or BIDU put spreads | 15% | Chinese AI platform-specific hedge | ### Step 3 — Optional higher-delta spot basket If you want more equity participation, buy the spot basket in **three tranches**: 1. 40% now. 2. 30% on a 5–8% pullback in SMH/QQQ. 3. 30% only after confirmation: U.S. labs retain clear benchmark lead, capex concerns stabilize, and power/data-center bottlenecks remain tight. ### Step 4 — Monthly rebalance - Add to GOOGL/MSFT if frontier odds improve but valuation remains reasonable. - Add to power names if AI capex persists and grid bottlenecks worsen. - Reduce NVDA/AMD/PLTR if valuation expands faster than earnings estimates. --- # Final check: Hyperliquid and Polymarket ## Hyperliquid: what to trade Hyperliquid now has a large set of non-crypto perps, including equity and ETF perps. Third-party trackers show listed markets such as **NVDA, GOOGL, MSFT, META, AMD, MU, ORCL, PLTR, GEV, VST, SMH, SOXL, TSM, ASML, AMAT, AVGO, KWEB/KSTR-style China proxies, CXMT, ZHIPU, MINIMAX, and others**. (usekeel.io) Hyperliquid Guide also notes that these non-crypto perps are contracts, not actual stocks, and funding/market behavior can differ from the underlying equity market. (hyperliquidguide.com) Best Hyperliquid expression: | Trade | Direction | Why | |---|---|---| | NVDA-PERP | Long | Highest U.S. compute bottleneck beta | | SMH-PERP | Long | Cleaner diversified semi exposure | | GOOGL-PERP | Long | U.S. frontier-model proxy | | MSFT-PERP | Long | Cloud/enterprise AI proxy | | AVGO-PERP | Long | Custom silicon/networking | | MU / DRAM / SKHX / SKHY | Long | HBM/memory bottleneck | | GEV / VST | Long | Power bottleneck | | CXMT-PERP | Short | China memory catch-up hedge | | ZHIPU / MINIMAX | Short or avoid | Chinese model-lab proxy; thin/idiosyncratic | | BABA-PERP | Short hedge | China AI/cloud ecosystem hedge | My suggested Hyperliquid tactical package: - **Long:** 40% SMH-PERP, 20% NVDA-PERP, 15% GOOGL-PERP, 10% MSFT-PERP, 10% MU/DRAM, 5% GEV/VST. - **Hedge:** short 20–30% of that notional in CXMT/ZHIPU/MINIMAX/BABA if liquid. - **Leverage:** max **1.5–2.0x**, isolated margin only. - **Stop:** if the perp sleeve loses 35–40% of its isolated margin, close it. - **Do not use SOXL-PERP** unless it is strictly intraday; it is already levered exposure. ## Polymarket: what to trade First, access matters. Polymarket’s international site shows U.S. users a notice that trading is blocked on polymarket.com and directs users to Polymarket US. (polymarket.com) Polymarket also says Polymarket US is operated by QCX LLC d/b/a Polymarket US, a CFTC-regulated Designated Contract Market. (polymarket.com) Also check operational status before trading, because AP reported that Polymarket paused international trading after a third-party security issue on October 29, 2026. (apnews.com) Most relevant Polymarket trades: | Market | Current visible odds | Trade for your thesis | My view | |---|---:|---|---| | **Which company has best AI model end of 2026?** | Google around 59%, Anthropic also visible as major contender | Buy U.S. lab outcomes if combined pricing is below your U.S.-win probability | Good, but U.S. lead is already heavily priced | | **Which company has best AI model end of November/December?** | Google leads visible short-dated markets | Tactical U.S. lab bet | Useful for timing, not core thesis | | **Will a Chinese company have a top ___ AI model by Dec. 31?** | Listed among China AI markets | Buy **No** if the odds imply too much China catch-up | Best direct China-catch-up hedge if liquid | | **US x China agree to pace the AI frontier in 2026?** | Around 5–7% Yes | Buy **No** only if you want a low-return “rivalry continues” bet | Low upside; probably not worth much capital | | **US Government removes public access to a major Chinese AI model in 2026?** | Around 16% Yes | Buy **Yes** as export-control / decoupling convexity | Small lottery ticket | | **AI bubble burst by June 30, 2027?** | Around 23% | Buy small **Yes** as hedge against your equity basket | Good hedge | | **OpenAI announces AGI before 2027?** | Yes around 13% | Buy small **Yes** if you want U.S. breakthrough convexity | Very binary, resolution-risk heavy | | **NVIDIA daily close markets** | Many short-dated markets | Use only tactically around earnings/news | Not a secular AI-race trade | Polymarket’s AI page shows hundreds of active AI markets, including the “best AI model end of 2026” market, AI bubble market, and various lab-ranking markets. (polymarket.com) The China page also shows China-specific AI and geopolitics markets, including the U.S.–China AI pacing agreement and public-access-to-Chinese-model market. (polymarket.com) --- ## Bottom line The cleanest implementation is: > **Long U.S. AI bottlenecks, not broad hype. Use SMH/NVDA/AVGO/MSFT/GOOGL call spreads, add selective power/grid exposure, hedge with KWEB/FXI/BABA put spreads, and use Polymarket/Hyperliquid only as small tactical overlays.** If you want one concise trade package: 1. **Core:** SMH + NVDA + AVGO + MSFT + GOOGL call spreads. 2. **Second derivative:** GEV / CEG / VST spot. 3. **Hedge:** KWEB / FXI put spreads plus small “AI bubble burst” Yes on Polymarket. 4. **Hyperliquid tactical:** long SMH/NVDA/GOOGL/MU perps, hedge with CXMT/ZHIPU/MINIMAX/BABA shorts. 5. **Polymarket tactical:** buy U.S. lab outcomes only when pricing is attractive; buy “No” on Chinese top-model markets where odds overstate China catch-up; buy small “Yes” on export-control escalation and AI-bubble risk as hedges.