P4-C4 · Export Controls → Domestic Substitution¶
Core One-Liner
Regulatory actions have a 6-24 month lag — short-term "obvious losers" may be offset by other factors, and domestic substitution in banned regions becomes a new thesis.
Real Public Cases — Use Part 3 Tools to Analyze 5 Real Events
P4-C4. All based on public data (U.S. Department of Commerce announcements + NVDA earnings + Huawei / SMIC public data).
1. Event: 2022-2026 U.S. AI Chip Export Controls on China — 3 Rounds¶
| Round | Time | Content | Market Reaction at Time (NVDA) |
|---|---|---|---|
| Round 1 | 2022/10 | Ban A100 / H100 sales to China | NVDA -10% (short-term) |
| Round 2 | 2023/10 | Ban H800 (customized throttled version) + AMD MI300 | NVDA -7% (short-term) |
| Round 3 | 2024/12 - 2025 | Restrict investment in China affiliates + ASML EUV continued ban (note: ASML has never sold EUV to China per CNBC 2024; now restrictions extend to certain advanced DUV) | NVDA +3% (already priced) |
→ Each round caused short-term panic, but **NVDA's total revenue from 2022→2025 still grew +400%+**. Why?
2. Four Dimensions of What Actually Happened¶
2.1 NVDA China/HK Revenue Has Steadily Declined¶
| Time | China/HK as % of NVDA Revenue | Source |
|---|---|---|
| 2022 (pre-controls) | ~25% (Data Center segment historical figure) | NVDA pre-2022 IR |
| FY2024 | Data Center China ~14% | NVDA 10-K FY24 |
| FY2025 (10-K) | ~13% company-wide ($17.1B / $130.5B) | NVDA 10-K FY25 |
Note: The pre-2022 "25%" was the Data Center segment figure before export controls; from 2022 onward, ~13% company-wide is the actual figure. Trend: continuing decline → eventual zero.
→ Post-controls revenue impact limited. But...
2.2 NVDA's Other Markets Filled the Gap¶
- Stargate $500B (P4-C3) U.S. capex explosion
- European sovereign AI (Mistral / sovereign cloud)
- UAE / Saudi sovereign funds (aggregate $100B+ committed across sovereign AI initiatives — UAE = Stargate UAE 1GW Abu Dhabi (Phase 1 200MW 2026, OpenAI disclosed no UAE dollar figure) + MGX investments; Saudi = HUMAIN $40B + PIF AI fund; NOT a single UAE Stargate $100B project)
- Japan / Korea AI (Softbank + Kakao)
→ NVDA 2024 revenue +126% YoY, far exceeding the China exposure (~13% FY25 per 10-K).
2.3 China's Domestic Substitution Emerged¶
Huawei Ascend 910B/C (per industry reports, not officially benchmarked): - 2022: First generation not viable - 2023: Ascend 910B close to A100 (per industry reports) - 2024: Ascend 910C — per DeepSeek research (Tom's Hardware 2025/02) ~60% of H100 inference performance (training still lags); process still trails at 7nm/5nm - Customers: Alibaba / ByteDance / Huawei Cloud
SMIC (per industry reports): - 2023: 7nm mass production (despite ASML not selling EUV, using DUV multi-patterning) - 2H23 launch yield <40% → per Semiecosystem / TD Cowen 2025 estimate 7nm yield has ramped to ~60-70%; 5nm still pilot (<20%) — still materially below TSM leading-edge - Revenue +30%+
YMTC / CXMT: - Domestic HBM mass production 2026+ - Domestic DRAM / NAND already mature
2.4 DeepSeek Trained GPT-4 Level Model Using H800 + Algorithm Efficiency (P4-C1)¶
Proves banning chips ≠ banning capability. Algorithm + efficiency breakthroughs can compensate partially.
3. P3-C4 Multi-PM Perspectives¶
| PM | NVDA Impact View | Beneficiary Candidate View |
|---|---|---|
| Value PM | China lost, long-term unrecoverable | NVDA's diversified customer base spreads risk |
| Growth PM | Total revenue rising, temporary loss compensated | SMIC / Huawei rising, but hard to invest publicly |
| Macro PM | Geopolitical escalation, medium-to-long term irreversible | U.S.-China AI dual-track development (split market) |
3 perspectives combined: **NVDA's short-term negative is offset by other markets, long-term China market permanently lost** (must accept).
4. Lesson — Regulatory Lag + Substitution Effect¶
4.1 Regulatory Actions Lag 6-24 Months¶
| Event | Immediate Reaction | Real Impact (12-24 months later) |
|---|---|---|
| 2022/10 Round 1 controls | NVDA -10% | NVDA revenue 2023 +126% |
| 2023/10 Round 2 | NVDA -7% | NVDA revenue 2024 +120% again |
| 2024/12 Round 3 | +3% (already priced) | TBD |
→ Control announcement ≠ immediate impact. Companies have lead time to adjust (Inventory / alternative markets).
4.2 Substitution Effect¶
Banned regions develop domestic alternatives: - Huawei Ascend (replaces H100) - SMIC 7nm (replaces TSM) - YMTC HBM (replaces SK Hynix / Micron China share)
→ Domestic substitution companies rise, but most foreign investors can't invest (Huawei private, SMIC HK listed but illiquid).
4.3 Dual-Track¶
Next 5-10 years, AI industry splits in two: - U.S. + allied markets: NVDA / TSM / ASML / OpenAI - China market: Huawei / SMIC / DeepSeek
No cross-penetration. Investors must pick a side — can't own globally.
5. What Your Thesis Should Learn¶
- Control announcements treat as short-term panic, not long-term thesis break — see if the company can fill the gap
- Banned regions = new window for domestic substitution thesis (but public investment hard)
- Cross-region exposure is risk — NVDA China/HK ~13% FY25 ($17.1B / $130.5B per 10-K), lost or not, total revenue still grew due to diversification
- 2026+ monitor: Semiconductor equipment export controls expansion (ASML / AMAT) / 5nm equipment expansion / cross-border investment restrictions
6. Next → P4-C5¶
You've seen the geopolitical wildcard. Final case: **Buffett used his 5-step framework, why didn't he buy NVDA**.
→ P4-C5 · Buffett's Forensic on Not Buying NVDA — Use P3-C2 Buffett's 5 steps to analyze NVDA item by item.
7. Further Reading (this chapter — US 3-round export controls)¶
All free sources, BIS official rule text / primary 10-K preferred
Primary BIS rule text (original source, all free PDFs):
- BIS · 2022/10/07 Interim Final Rule "Implementation of Additional Export Controls" (Federal Register) — Round 1, advanced computing / supercomputer / semiconductor manufacturing controls (14nm/16nm + advanced computing triggers); A800/H800 are NVDA China-compliant variants made after the rule, blocked by 2023/10 follow-up
- BIS · 2023/10/17 Updated Rules — Round 2, H800/A800 also banned, new performance density / total processing performance metrics
- BIS · 2024/12/02 Final Rule "Implementation of Additional Export Controls" (HBM rule) — Round 3, HBM2/HBM3 export controls + Entity List expansion
- BIS homepage · all current rules — continually updated, read the official text directly
Primary 10-K risk factors (company original disclosures):
- NVIDIA 10-K FY25 (2025/02 SEC filing) — Risk Factors section explicitly quantifies China export controls revenue impact (~$5B/quarter for H20)
- AMD 10-K (SEC filing) — MI300 export risk paragraph
- SMIC annual report (HKEX disclosure) — corresponding disclosure from China foundry perspective
Reuters / primary news (free):
- Reuters: "U.S. tightens chip export controls on China" (2022/10/07) — Round 1 day-of coverage
- Reuters: "Huawei Mate 60 Pro teardown reveals SMIC 7nm chip" (2023/09) — key report on export controls' reverse incubation of SMIC 7nm
- Reuters: "BIS December 2024 HBM rule" coverage (2024/12/02) — Round 3 day-of
CSIS / Stanford / free think-tank reports:
- CSIS · "Choking Off China's Access to the Future of AI" (Gregory Allen, 2022/10) — authoritative read on Round 1 rules (CSIS entirely free)
- CSIS · "Updated October 7 Export Controls" (2023/10) — Round 2 read
- Stanford HAI · "Compute and U.S.-China Tech Competition" series — academic, free
Books (borrow from library):
- Chris Miller "Chip War" (2022) — authoritative narrative on semiconductor geopolitics, best background book for the export controls thread
Wikipedia (background):
- "United States New Export Controls on Advanced Computing and Semiconductors to China" — full 3-round rule timeline + citations
- "SMIC" + "Huawei" — background on the 2 core domestic-substitution companies
Pair with chapter self-check: After reading any 3-4 of the above, you should be able to answer the self-check more confidently — especially on "3 rounds + 6-24 month lag" and on regulatory-text + real-event evidence for "why controls reverse-incubated SMIC/Huawei".