US–China Competition

POLISCI 158 · AI & Politics · Session 14 · Monday, November 9, 2026

Simone Paci

Roadmap

  1. Part I: Why a race?
    • Security dilemma -> why states frame AI as a race
    • Two prizes: first to AGI vs diffusion marathon
  2. Part II: Who’s winning?
    • Scoreboard: where US leads, where China leads
    • Winner depends on which race
  3. Part III: The chokepoint and its leaks
    • Compute as chokepoint -> how export controls leak
    • China’s counter-chokepoints

What do Americans think?

Pew Research Center chart: More Americans think China is the leader in developing AI than say this of the U.S., but a third are not sure. U.S. adults: 12% U.S. is more advanced than China, 18% both are equally advanced, 36% China is more advanced than U.S., 33% not sure. Survey of U.S. adults conducted June 22-28, 2026

Ahead at what?

33% not sure

Pew Research Center, survey June 22–28, 2026 (n = 3,488)

Does keeping up with China matter to Americans?

69% keeping up w/ China in AI = important very 40 · somewhat 29

81% safety > staying at forefront of innovation innovation: 14

65% hold both views

2 separate questions · safety item doesn’t mention China

Quinnipiac University Poll, Sep 24–27, 2026 (n = 1,202, ±3.5)

Part I: Why would states race over AI?

Last week the compute had an address. This week it has a flag.

Why did one Chinese model move markets?

−17% ≈$593B one day

DeepSeek-R1 out: Jan 20, 2025

Mon, Jan 27: Nvidia −17% in 1 day, ~$593B market value

“DeepSeek-R1 is AI’s Sputnik moment.” Marc Andreessen (VC)

$589–600B, by data vendor

DeepSeek release note, Jan 20, 2025; Reuters/LSEG, Jan 28, 2025; Fortune, Jan 27, 2025

At what level should we look?

Waltz: 3 images of the causes of war

First image: individuals (leaders, human nature) Second image: the state (internal structure: regime, domestic politics) Third image: the international system (anarchy, distribution of power) this session

Cover of Kenneth N. Waltz, Man, the State, and War, Columbia University Press edition

Waltz, Man, the State, and War (1959; Columbia UP, 2018 ed.)

Why does one side’s defense look like a threat?

A builds capability (for its own security) B reads it as a threat B builds A reads it as a threat anarchy: no authority above states

“Many of the means by which a state tries to increase its security decrease the security of others.” Jervis (1978)

Jervis (1978), World Politics, p. 169, as quoted in Scharre (2021)

Zero-sum or positive-sum?

Realism Liberalism
Relative gains: your gain = my loss Absolute gains: both can gain
Security first · balance of power Trade, research, institutions, interdependence

Which does AI look like?

What if the “race” is itself an idea?

Constructivism: race framing -> state behavior

What counts as threat -> depends on regime

“…not only Chinese anxiety over losing ground … but also concerns about threats to ideological and regime security.” Zeng (2025), abstract, on China’s reaction to ChatGPT

Zeng: security dilemma · constructivist reading = ours

Zeng, British Journal of Politics and International Relations, Nov 5, 2025

Is it really an arms race?

Texas National Security Review article page: Vol 4, Iss 3, Summer 2021, pages 121–132; Debunking the AI Arms Race Theory, by Paul Scharre

“There is no AI arms race.”

But belief in one -> “cut corners on testing” -> “unsafe AI systems”

2021 · military AI · pre-ChatGPT

Scharre, Texas National Security Review, Summer 2021

What if the first to arrive wins?

AGI / superintelligence USChina the leader AI improves AI the gap locks in

Cover of America's AI Action Plan, Winning the Race, The White House, July 2025

“The United States is in a race to achieve global dominance in artificial intelligence.” AI Action Plan, July 2025

“Superintelligence will give a decisive economic and military advantage.” Aschenbrenner, 2024

“A temporary lead could be parlayed into a durable advantage.” Dario Amodei (Anthropic CEO), Jan 2025

Aschenbrenner: self-published · Amodei: AI CEO, stake in chip policy

White House, Jul 2025; Aschenbrenner, Situational Awareness, Jun 2024; Amodei, “On DeepSeek and Export Controls,” Jan 2025

Or is it a marathon?

Power = spreading a general-purpose tech across the economy

decades adoption invention diffusion across industries
  • “an AI diffusion marathon” vs sprint to AGI (Ding 2025)
  • “Diffusion occurs over decades, not years.” (Narayanan and Kapoor 2025)

N&K: generalize from past tech · institute essay, not peer-reviewed

Ding, RAND, Sep 30, 2025, and Technology and the Rise of Great Powers (2024); Narayanan and Kapoor, Knight Institute, Apr 2025

Do leads last?

Leads leak Complexity protects leaders
Lawfare article header: The AI Race Isn't Real, by Yonathan A. Arbel and Matthew Tokson, Tuesday, May 19, 2026
“AI knowledge is leaky.” “There is no finish line.” (Arbel and Tokson 2026) Complexity -> weapons hard to imitate, even w/ stolen designs: China vs US stealth fighters (Gilli and Gilli 2019)

Frontier AI: paper or stealth fighter?

Gilli & Gilli: weapons, not AI · Arbel & Tokson: EUV + stealth leads lasted decades

Arbel and Tokson, Lawfare, May 19, 2026; Gilli and Gilli, International Security, 2019

Is the race itself the danger?

Racing causes risk A lead makes safety affordable
Racing -> “skimping on safety precautions” (Armstrong, Bostrom and Shulman 2016) “Maintaining a healthy lead will be decisive for safety” (Aschenbrenner 2024)
Superintelligence race -> great-power conflict, loss of control, weaker liberal democracy (Katzke and Futerman 2024)

Katzke & Futerman, Aschenbrenner: not peer-reviewed

Armstrong et al., AI & Society, 2016; Katzke and Futerman, arXiv, Dec 2024; Aschenbrenner, Jun 2024

Could the threat of sabotage keep the peace?

Mutual Assured AI Malfunction: bid for AI dominance -> “preventive sabotage by rivals”

  • Covert cyberattacks -> strikes on datacenters

arXiv abstract page: Superintelligence Strategy: Expert Version, by Dan Hendrycks, Eric Schmidt and Alexandr Wang, submitted 7 Mar 2025, revised 14 Apr 2025; abstract introduces Mutual Assured AI Malfunction (MAIM)

Working paper, not peer-reviewed

Hendrycks, Schmidt and Wang, “Superintelligence Strategy,” arXiv, Mar 2025

Part II: Who’s winning, and at what?

If the prize differs, so does the scoreboard.

Who has the compute?

Epoch AI chart: The US accounts for over 70% of new AI compute since 2022; China, about 10%. Stacked area of share of cumulative arrivals in H100-equivalents by country, 2023 Q3 to 2026 Q2

New AI compute, H1 2026: 78% -> US

Installed GPU-cluster perf., May 2025: US ~75% · China 15%

Customs-data estimates (arrivals ≠ use) · cluster data ≈ 10–20% of global compute

Epoch AI, Oct 27, 2026 and Jun 5, 2025

Who has the money?

Private AI investment, 2025

AI Index 2026 chart: global private investment in AI by geographic area, 2013–25, billions of US dollars; 2025: United States 285.88, Europe 20.92, China 12.41. Source: Quid

Private investment only; understates China’s state guidance funds.

Stanford HAI, AI Index 2026

Who builds the best models?

Epoch AI chart: Chinese AI models have lagged the US frontier by 7 months on average since 2023. Step lines of Epoch Capabilities Index for the top US and top Chinese model over time, 2023 to end of 2025

Since 2023: every frontier model American · China ~7 mo behind on avg (4–14)

One benchmark composite (Epoch Capabilities Index) · data to Dec 2025

Epoch AI, Jan 2, 2026

Who makes the chips?

Epoch AI chart: Huawei's 2026 AI compute output is likely less than 4% of Nvidia's. Estimated 2026 production in H100-equivalents: Huawei Ascend 0.9M (0.6M–1.2M), Nvidia 23M (19M–29M), 90% intervals

Huawei 2026 AI chip output: < 4% of Nvidia’s compute

Huawei chips 3–4 yrs behind Nvidia, until ≥ 2030

Epoch estimates · China’s own advanced lithography: likely after 2030

Epoch AI, Sep 2026

Which models do users prefer?

AI Index 2026 chart: performance of top United States vs. Chinese models on the Arena, May 2023 to early 2026; latest: 1,503 top US model, 1,464 top China model

Near parity

  • DeepSeek-R1 briefly = top US model (Feb 2025)
  • US lead, Mar 2026: 2.7%

Arena = crowd votes on model pairs, not benchmarks · same period as slide 20, different verdict

Stanford HAI, AI Index 2026 (Arena data)

Whose open models get downloaded?

Hugging Face chart: Hugging Face downloads from February 2025 to February 2026 by country of model developer, millions: China 663, United States of America 595, user 157, Germany 78, United Kingdom 61, International/Online 51, France 12, Israel 5, Switzerland 3, Indonesia 3

Chinese models: 41% of Hugging Face downloads (Feb 2025–Feb 2026) -> largest share

Qwen, Jan–Jul 2026: ~2.05B downloads · 151,448 derivative models

Misses API + private use · ~half: no clear country · newest large Kimi/Qwen: licence limits

Hugging Face, Mar 17 and Aug 14, 2026

What does “61%” measure?

Viral claim: Chinese models = “61% of total token consumption on OpenRouter”

  • Actually: 61% of top-10 models’ tokens (Ding)
  • All models, Nov 2024–Nov 2025: ~13% of weekly tokens on avg, ≈30% some weeks

Top ten only

OpenRouter LLM leaderboard screenshot, this week: top ten models by tokens, including MiniMax M2.5, Gemini 3 Flash Preview, DeepSeek V3.2, Kimi K2.5, Claude Opus 4.6, Grok 4.1 Fast, Claude Sonnet 4.6, GLM 5, Trinity Large Preview and Claude Sonnet 4.5

Top-10 leaderboard, week of Mar 1, 2026

All models

OpenRouter chart: Token Share by Source Type and Origin, weekly, November 2024 to November 2025; stacked shares of Chinese open-source, rest-of-world open-source and closed models

Token Share by Source Type and Origin, all models, weekly

OpenRouter = small, developer-heavy slice of AI use

Ding, ChinAI, Mar 2, 2026; a16z and OpenRouter, “State of AI,” Dec 2025

Who is automating factories?

China: 59% of world’s new industrial robots, 2025

IFR chart: annual installations of industrial robots, 5 largest markets 2025, thousand units: China 354.2, United States 38.4, Japan 36.2, Republic of Korea 30.2, Germany 24.8

1-year installations, not robot stock

IFR, World Robotics 2026, Sep 24, 2026

Who builds humanoids, and for what?

Chinese vendors: over 95% of ~25,000 humanoids shipped, H1 2026

69% -> research/education, display, govt data-collection centres (83.8% in 2025)

Trackers disagree on total · 69%: global or China only? IDC doesn’t say

IDC, Sep 29, 2026

Who is adding the power?

Solar added in 2025: China 378 GW, 58% of the world

China · 378 GW 58% of the world rest of world

Electricity demand growth, 2025: China +5%, US +3%

ChinaUS +5% (+503 TWh) +3% (+131 TWh)

Solar in GW (DC).

Ember, Global Electricity Review 2026, Apr 21, 2026

Where is it cheaper to build?

Data center cost/MW: ~$6M China vs ~$10M US

Compute per $: US ahead in both modelled scenarios · H200s close raw-performance gap

ChinaTalk chart: construction costs for a 400 MW data center: China $2.4B, U.S. $4.0B, with error bars; 40% cheaper

ChinaTalk chart: 400 MW data center compute efficiency, memory-bandwidth limited, GFLOPS per dollar: U.S. with GB200 NVL72 23.96; China with CM384 18.41; China with H200 13.64; with error bars

One modelled 400 MW build, single source · hardware prices not public

Zakaria, ChinaTalk, Feb 19, 2026

Who patents more?

GenAI patent families, 2014–2023: China 38,210 vs US 6,276

ChinaUS 38,210 6,276

US leads on higher-impact patents

Quantity ≠ quality · by inventor location

WIPO, Jul 3, 2024; Stanford HAI, AI Index 2026

Where do the researchers work?

NeurIPS authors by work location, 2022 → 2025: US 46% → 34% · China 27% → 41%

46%27% 34%41% 20222025 US China

n = 10,280 complete profiles of 25,677

40% subset w/ full career histories, not random · US still net gainer, keeps 89%

Ma and Yang, Carnegie Endowment, Sep 23, 2026

What is Beijing aiming for?

CSET translation of the State Council AI+ Opinions, page 2 excerpt: By 2027, broad and deep integration of AI with six key fields; penetration of new generation smart terminal devices and intelligent agents will exceed 70%; by 2030, adoption will exceed 90%; by 2035, China will have fully entered the new intelligent economy and smart society stage

“AI+” plan, Aug 26, 2025

  • 2027: > 70% adoption, smart devices + AI agents · 2030: > 90%
  • AGI: not mentioned

Also: “enhance the fundamental capabilities of models” · no AGI target

State Council (2025), CSET translation, Sep 24, 2025

What is Washington aiming for?

America's AI Action Plan, Introduction, first paragraph: The United States is in a race to achieve global dominance in artificial intelligence. ... Just like we won the space race, it is imperative that the United States and its allies win this race.

“Just like we won the space race … win this race.” AI Action Plan, July 2025

Sep 29, 2026: executive order renames AI “Super Intelligence” · WH fact sheet: winning “the SI race”

“SI” = new name for AI, existing legal definition, not superintelligence

White House, Jul 2025; Executive Order 14434 and fact sheet, Sep 29, 2026

What does Xi ask for?

CSET translation, page 1 headline: At the 20th Collective Study Session of the CCP Central Committee Politburo, Xi Jinping Stresses: Persist in Being Self-Reliant, Be Strongly Oriented Toward Applications, and Push the Orderly Development of Artificial Intelligence; Xinhua News Agency, Beijing, April 26 [2025]

“Persist in Being Self-Reliant, Be Strongly Oriented Toward Applications…” (title of the readout)

  • Master “high-end microchips and foundational software”
  • Also: “…win superiority in the field of AI”

State-media readout, not transcript · no AGI mention

Xinhua readout of the Politburo study session, Apr 26, 2025; CSET translation

Is Beijing racing to AGI?

Several analysts: no

“Beijing seems not to care about the so-called ‘race to AGI’.” Kyle Chan

“They aren’t trying to build AGI. They’re trying to make a profit.” Selena Xu

Some founders: yes

“Our destination is AGI.” Liang Wenfeng, DeepSeek

Liang, 2024 interview, as quoted by Chan

Chan, High Capacity, Feb 15, 2026; Sheehan and Xu, Your Undivided Attention, Dec 18, 2025

Or is AGI quietly state policy?

“AGI is now written into China’s operating system for modernization.” Matthew Johnson

  • Beijing, 2023: general-AI measures -> “explore new paths to general AI”
  • Shanghai, 2025: open competition, general-AI tasks, embodied intelligence first
  • 15th Five-Year Plan: “exploration of development pathways toward general-purpose AI”

Partly translation: 通用人工智能 = “general-purpose AI” or “AGI”

Johnson, Jamestown China Brief, Oct 2025; Beijing municipal government, May 2023; Shanghai SHEITC, Apr 21, 2025; DigiChina, Mar 17, 2026

Two caveats before we move on

Diffusion strategy ≠ diffusing well

  • Ding: judge states by “capacity to diffuse, or widely adopt, innovations”

China’s approach “is partly driven by necessity” -> US controls on Nvidia chips (Chan)

Ding predates AI+ plan · application here = ours

Ding, Review of International Political Economy, 2024; Chan, Brookings, Mar 9, 2026

Debate: does the US–China rivalry stand in the way of slowing down AI?

Assigned sides: left half = yes · right half = no

Can anyone slow down while the other races?
Should the US slow down even if China does not?
Can the US afford to slow down?
  1. Pairs · 3 min: build your side’s case from 1+ angles
  2. Floor · 8 min: sides alternate, one point at a time
  3. Wrap-up · 1 min

Argue assigned side, not own view · written answers welcome

Part III: Can you stop a rival by cutting off its chips?

The US lever acts on the input where it leads most.

Where are the chokepoints?

SIA and BCG 2024, Exhibit 2: semiconductor industry value-added by activity and region, 2022, percent; for example EDA and core IP USA 68%, logic USA 65%, manufacturing equipment USA 47%, Japan 26%, EU 18%, memory South Korea 60%, wafer fabrication Taiwan 18%, Mainland China 24%

Design and EDA: USEDA & core IP: US 68% · logic design: US 65%
Equipment: US, Japan, EUUS 47% · Japan 26% · EU 18%
EUV lithography: ASML (Netherlands)
Leading-edge fabs: Taiwan, Korea
HBM memory: KoreaHBM revenue, Q2 2026: SK hynix 50% · Samsung 33%

Value added by region, 2022 · one European firm: 87% of lithography market

SIA and BCG, May 2024; Counterpoint Research (HBM), Q2 2026

How does controlling a hub give you power?

hub chokepoint panopticon

Control a central network node

  • Panopticon effect: gather info from flows
  • Chokepoint effect: deny adversaries access

Cases: SWIFT, internet -> chips?

Farrell and Newman, International Security, 2019 (today’s reading)

How did the US tighten the screws?

Oct 7, 2022

advanced computing and chip-equipment rules

Oct 17, 2023

rules tightened (effective Nov 17)

Dec 2, 2024

high-bandwidth memory, equipment, Entity List additions

Jan 13, 2025

AI Diffusion Rule (published Jan 15)

Bureau of Industry and Security; Federal Register

Then why loosen them?

Apr 2025

licence required: Nvidia H20

May 13, 2025

Diffusion Rule rescinded

Aug 2025

reported deal: 15% revenue share, H20 sales to China

Jan 2026

H200 licences case by case · cap 50% of US shipments · US testing · 25% import tariff

Blackwell

still presumption of denial

BIS press release, January 13, 2026: Department of Commerce Revises License Review Policy for Semiconductors Exported to China

15% deal: reported, never a published rule · 25% = tariff on chips entering US for testing

BIS, May 13, 2025 and Jan 13, 2026; White House fact sheet, Jan 14, 2026; Nvidia filings

Do the controls work?

US advantage, AI compute produced in 2026

  • No exports, no smuggling: 21–49×
  • Unrestricted H200 exports: ~1.2× to ~6.7× (volume, supply)

IFP Figure 1, scenarios 1–5 with legend: share of 2026-produced AI compute; no US AI chip exports to China, no smuggling: US 73%, US advantage over China 21x–49x; with smuggling 10x–12x and 9x–11x; conservative H200 exports: 4.0x–6.7x and 3.4x–5.4x

IFP Figure 1, scenarios 6–9: baseline H200 exports, US advantage over China 2.7x–5.1x and 2.1x–3.6x; aggressive H200 and comparable chip exports, 1.9x–3.8x and 1.2x–2.1x

Think-tank model · range depends on export volume + FP4/FP8 assumptions

Institute for Progress (IFP), Dec 7, 2025

A lead in what?

“a hardware lead of around 4 years when it comes to training frontier models”

“essentially no lead when it comes to serving those models to users”

No peer-reviewed causal estimate found: export controls -> Chinese model capability

Erdil: author’s views · our search (Nov 2026): not systematic

Erdil, Epoch AI Gradient Updates, Dec 6, 2024

Leak 1: how many chips get smuggled?

Est. 290,000–1.6M H100-equivalents smuggled to China through 2025 · median 660,000 ≈ ⅓ of China’s compute

  • Operation Gatekeeper (Dec 2025): ≥ $160M of H100s/H200s exported or attempted
  • Indictment (Mar 2026): SE Asian buyer, ~$2.5B of US maker’s servers · ≥ $510M allegedly -> China in 1 month
00.5M1M1.5M 290kmedian 660k1.6M alleged cases: nearly 300k H100e, 90% interval

DOJ press release, Monday, December 8, 2025: U.S. Authorities Shut Down Major China-Linked AI Tech Smuggling Network; Operation Gatekeeper disrupts trafficking network and seizes more than $50 million in advanced GPUs

Indictments + complaints = allegations · only Hsu, Hao Global pleaded guilty

Epoch AI, Apr 29, 2026; DOJ, Dec 8, 2025; SDNY, Mar 19, 2026

Leak 2: why buy chips if you can rent them?

Chinese firms rent Nvidia capacity in SE Asian data centers -> legal under current US law

Remote Access Security Act: passed House -> stuck in Senate committee

Apr 7, 2025

introduced

Apr 9, 2025

committee 51–0

Jan 12, 2026

House 369–22

Jan 13, 2026

referred to Senate Banking

Status as of Nov 7, 2026 · not law

Tan, Carnegie Endowment, May 5, 2026; Congress.gov

Leak 3: can you copy a model by talking to it?

CISA cybersecurity advisory page: China-Based Artificial Intelligence Companies Conducting Industrial-Scale Distillation Campaigns Against U.S. AI Companies; release date September 08, 2026; alert code AA26-251A; executive summary names DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI

Distillation: training on a rival model’s outputs

  • US agencies (Sep 8, 2026): 6 Chinese labs, “likely with Chinese government awareness” -> “billions of tokens” from US models
  • Anthropic (Feb 2026): ~24,000 fraudulent accounts · > 16M exchanges w/ Claude

Govt attribution + company allegations, not court findings · Anthropic’s figures = its own

NSA, CISA and FBI, advisory AA26-251A, Sep 8, 2026; Anthropic, Feb 23, 2026

Leak 4: what does documented espionage look like?

DOJ press release, Friday, January 30, 2026: Former Google Engineer Found Guilty of Economic Espionage and Theft of Confidential AI Technology

Jan 29, 2026: jury convicts ex-Google engineer Linwei Ding: economic espionage + AI trade-secret theft

  • Aug 2026: judge vacates espionage counts · trade-secret convictions stand
  • Sentence: just under 1 yr in prison

One case · judge: insufficient evidence he meant to benefit PRC govt

DOJ, Jan 30, 2026; KQED, Sep 1, 2026

Leak 5: what if you buy before the door shuts?

$16B+ H20 orders

$16B Nvidia H20 orders, 3 Chinese firms (reported, Apr 2025)

Huawei stockpiled chip dies + high-bandwidth memory around Dec 2024 controls

Order figure first reported by The Information.

IFP, Apr 15, 2025

Do controls backfire?

Controls spur the target Controls slow the target
Hit firms “respond disproportionately by investing in research and development” (Clayton et al. 2025) China’s chip-equipment localization “predated any modern effort to impose meaningful export controls” (Allen 2024)
Controls “can at best slow and disrupt Chinese efforts, not stop them” (Allen 2025)

Clayton et al.: NBER working paper.

Clayton, Coppola, Maggiori and Schreger, NBER, Jul 2025; Allen, CSIS, Nov 26, 2024 and Mar 7, 2025

Who pays for the controls?

NY Fed Staff Report 1096, Figure 3: cumulative abnormal returns of affected US suppliers in a window from 10 days before to 20 days after a target's inclusion in BIS lists; Panel A 3-factor model and Panel B 5-factor model, both falling to about minus 3 to minus 4 percent

US suppliers: −$130B market value after Entity List additions, 2010–22

“In China, we have now dropped to zero.” “…already largely backfired.” Jensen Huang, Nvidia, May 2026

Allies: Dutch licence rules widened (Sep 2024, Apr 2025); Nexperia dispute (Sep–Nov 2025)

NY Fed: Entity List additions 2010–22, not AI-chip rules · “zero” = Nvidia’s share of China’s AI-accelerator market

Crosignani et al., NY Fed Staff Report 1096, 2025; Tom’s Hardware, May 3, 2026; Government of the Netherlands

What can Beijing squeeze?

Export curbs: gallium, germanium, antimony -> US (Dec 2024) · rare earths (Apr + Oct 2025)

Antitrust: preliminary finding vs Nvidia (Sep 2025)

Customs: H200s “not permitted to enter China” (Jan 2026) · ~600,000 reportedly waiting in Hong Kong (Nov 5, 2026)

US China chips, tools rare earths, gallium, germanium, market access

MOFCOM website, Announcement 2025 No. 70 of the Ministry of Commerce and General Administration of Customs, dated 2025-11-07, suspending announcements Nos. 55–58 and 61–62 from the date of issue until November 10, 2026

Customs + Hong Kong reports: anonymous sources (Reuters) · unconfirmed by Nvidia or Beijing

MOFCOM announcements; CNBC, Sep 15, 2025; Reuters, Jan 15 and Nov 5, 2026

Does every chokepoint invite a counter-chokepoint?

“Each side’s attempt to become safer makes others feel less secure, driving further escalation.” Farrell, Newman and Tan

  • Cold War: control regimes expand -> “weakening enforcement and encouraging defection” (Lind and Mastanduno)

Who writes the technical standards?

Nov 10
Tomorrow: China’s rare-earth curb suspension expires · wider truce -> Jan 10, 2027

US chip controls China rare-earthcurbs …

Foreign Affairs article header: The Chokepoint Trap: How the Search for Economic Security Will End in Geopolitical Calamity, by Henry Farrell, Abraham Newman and Yeling Tan, November/December 2026, published October 20, 2026

No MOFCOM extension of Nov 10 date, as of Nov 7, 2026

Foreign Affairs, Nov/Dec 2026; Texas National Security Review, Sep 2025; MOFCOM Announcement 2025 No. 70; MOFCOM readout, Sep 28, 2026

What should you take from today?

  1. Security dilemma -> race framing · stakes depend on prize: first to AGI vs diffusion marathon
  2. US leads: compute, chips, money, frontier · China: open models, robots, power -> winner depends on race
  3. Chips = US chokepoint, but leaky: smuggling, remote access, distillation, espionage, stockpiling · China has chokepoints too

Back to the start: ahead at what?

See you next class

Implementation Milestone · Sun, Nov 15, 11:59 p.m. · 800–1,000 words

Wednesday, Nov 11: Agentic AI: From Chatbots to Agents · Yiqing Xu

Next Monday, Nov 16: AI Sovereignty