Will China overtake the US in the AI race?

Slides for the YouTube video

Electricity generation since 2000

Source:Redrawn from: Institute for Progress' "How Should the US Prepare for Increasingly Automated AI R&D?", Aug 2026.

New electricity addition in 2025

Source:Ember, Global Electricity Review 2026; IEA, Global Energy Review 2026.

Electricity usage breakdown, 2025

Source:China: National Energy Administration. US: EIA. Data centres: IEA, same method for both countries. "Services" is China's tertiary industry and the EIA's commercial sector; "Other" is agriculture (China) and transport (US).

Data centres: electricity used in 2025

TWh, all data centres (not just AI)

Source:IEA, Key Questions on Energy and AI (2026).

Data centres: capacity in 2026

Gigawatts in operation, with China's pipeline

Source:SemiAnalysis, via the Financial Times (Oct 2026). The paler red is China's 50 GW under construction or announced; the FT gives no US pipeline figure.

US data centres: projected electricity use

Source:US: Lawrence Berkeley National Laboratory, 2025 Update (June 2026). China: IEA. Dashed lines are projections.

Electricity usage per person

Source:Ember. "At home" is an estimate: residential use divided by population.

Waiting times to build power infrastructure

Source:LBNL, Queued Up 2026; IEA (2025); Wood Mackenzie; TBEA; GE Vernova. China's transformer figure is a rough estimate. Building a data centre: Financial Times (Oct 2026), 12-18 months in Inner Mongolia against 18-24 in the US.

China's "Eastern Data, Western Computing" plan

Computing hubs

94% of people live east of this line ~1,600 km Inner MongoliaInner Mongolia: Ulanqab89 data centres, 15 GW committedNingxiaNingxiaGansuGansuGuizhouGuizhouChengdu-ChongqingChengdu-ChongqingBeijing-Tianjin-HebeiBeijing-Tianjin-HebeiYangtze River DeltaYangtze River DeltaGreater Bay AreaGreater Bay Area
western hubs: cheap powereastern hubs: the users

Rack usage in data centres

Source:ASPI Strategist (May 2026); hub list from China's state planner; Ulanqab: Financial Times (Oct 2026); map: Natural Earth. The 20-30% is from May; the FT found the western hubs filling by October.

Ulanqab, Inner Mongolia

Where it is

BeijingUlanqab89 data centres, built or planned

Cost of electricity (US cents per kWh)

05105.0Ulanqab~8.5China industrialaverage~11Beijing~9US industrialaverage

Time to build a data centre

06121824 months12-18 monthsUlanqab18-24 monthsUnited States

Capacity

010203024 GWAll of China today15 GWUlanqab (committed)

Wind power (GW)

0100200117 GWInner Mongolia~30 GWUnited Kingdom

Average temperature (C)

010204.3Ulanqab~13Beijing~15Guiyang~17Shanghai~23Guangzhou
Source:Financial Times, reporting from Ulanqab (Oct 2026); China total: SemiAnalysis. Prices converted at about 7.1 yuan per dollar; US price and city temperatures are approximate averages. UK wind is derived from the FT's "nearly four times". Map: Natural Earth.

Usable AI compute

Millions of H100-equivalents

Source:ChinaTalk (April 2026). The US bar is my estimate from Epoch AI data.

AI spending, 2026

Source:Goldman Sachs estimates, via Al Jazeera (Sept 2026). US: Google, Amazon, Microsoft, Meta, Oracle. China: Alibaba, Tencent, Baidu, ByteDance.

AI chips made in 2026

AI chips made in 2026 (millions)

Compute (millions of H100-equivalents)

Source:Epoch AI estimates (Sept 2026). One H100-equivalent = the compute of one Nvidia H100.

AI compute made per year

Source:Epoch AI, Data on AI chip sales. Dashed = Epoch's forecast for 2026.

Speed of each company's best AI chip, relative to Nvidia's H100

Source:Epoch AI. Raw speed of each chip relative to Nvidia's H100.

When SMIC and TSMC reach each chip generation

Source:TSMC announcements; TechInsights teardowns of SMIC-made chips. TSMC is Taiwanese.

EUV lithography machines sold

Source:ASML annual reports (units sold, summed); ASML via Reuters (Dec 2025).

EUV, a serious bottleneck

1996200020042008201220162020202420282032ASML1997 programme starts2006 first prototypes2010 pre-production machine2013 first production machine2019 chips in products2024 next generation (High-NA) at Intel2026 light source hits 1,000 W in a testChina2019 programme starts2025 prototype makes light, no wafer2028 government target2030 first working chip
Source:ASML, "Making EUV: from lab to fab"; China: Reuters, 17 Dec 2025. China's start year is inferred from Reuters' "six-year initiative"; 2028 is a target, 2030 the project's own estimate. The red dot on ASML's line: UBS (Sept 2026) puts China's EUV at about where ASML stood in 2004.

China's usable AI compute, by source

Source:ChinaTalk (April 2026). Median estimates with wide ranges.

AI chips in China: share of new sales and of total compute, by maker

New AI chips sold in China each year (% of sales)

All the AI compute China can use (%)

Source:Sales: Bernstein via Tom's Hardware. Compute: ChinaTalk. *Everyone but Nvidia: almost all Chinese vendors.

Share of the world DRAM memory market, by country

Share of the world market for DRAM memory chips, by the maker's home country (% of revenue)

Source:TrendForce quarterly DRAM revenue rankings; CXMT's listing prospectus. US = Micron, China = CXMT.

Memory chips vs logic (AI) chips

Memory Logic (AI chips) Layout The same cell, repeated Irregular: every part is different A defect A spare row takes over. The chip still ships. One defect kills the whole chip. How it improves Stack more layers Chinese tools can do this Shrink the features Needs EUV machines China today About 10% of the world market 4-5 years behind, low yield
Source:Standard semiconductor reasoning. Shares: TrendForce, Counterpoint.

Model capability: US frontier vs models available in China

Source:Redrawn from IFP's chart (Epoch Capabilities Index). Dashed = IFP's projection.

Three races

1. Frontier modelUS, by about 8 months
2. DiffusionContested
3. Base modelsChina, on open weights

Open models: downloads, derivatives and share by country of origin

Hugging Face downloads, Jan-Jul 2026 (millions)

Models built on each family (thousands)

Share of downloads, by where the model is from, Feb 2025 - Feb 2026

Source:Hugging Face, State of Open Models (Aug 2026) and State of Open Source, Spring 2026. Pie: shares worked out from Hugging Face's published download counts.

Where elite AI researchers work

Where elite AI researchers work (% of all of them)

Source:Carnegie Endowment, AI Talent v3.0 (2022 and 2025): authors at one top AI conference. Dashed: 2019, from the tracker's first edition (MacroPolo).

Where elite AI researchers trained and where they work, 2022 vs 2025

Out of every 100 elite AI researchers

2022 26 trained in China, work in China 20 trained in China, work abroad 20 trained in the US 34 trained elsewhere 2025 39 trained in China, work in China 18 trained in China, work abroad 13 trained in the US 30 trained elsewhere
Source:Carnegie Endowment, AI Talent v3.0. "Trained in" = country of first degree. Counts are rounded from its percentages.

RAND: seven strategies for the United States

StrategyThe US wins when...
ADominanceIt gets there first and suppresses rivals
BCo-DevelopmentIt governs superintelligence jointly with rivals
CPreparednessInformal coordination, resilient to accidents
DMoratoriumEveryone agrees to halt above a threshold
EDeterrenceIt prevents dangerous AI by threat and force
FContinuity of SocietySelf-sustaining settlements survive
GAccelerationMaximum development, markets sort it out
Source:RAND PE-A5105-1 (Sept 2026), Table 1, condensed.

RAND's decision tree for choosing a strategy

Source:Redrawn from RAND PE-A5105-1, Figure 1.

How it looks from Beijing

"Fear-mongering, confrontation and vicious competition will only hamper efforts toward sound global AI governance, which serves no one's interest."
Guo Jiakun, China's Foreign Ministry, 14 September 2026 - asked about calls by Dario Amodei, Sam Altman and Elon Musk to slow frontier AI
Source:China's Foreign Ministry, regular press conference transcript, 14 Sept 2026; reported by Rest of World.

US-China steps on AI safety, March 2024 to November 2026

2024 2025 2026 Sep 2026 October November The last three months are drawn larger March 2024 Scientists from both countries sign shared red lines, in Beijing May 2024 Companies from both countries sign safety commitments, in Seoul July 2025 China: AI governance belongs at the UN July 2026 1,300+ AI lab employees sign the "pacing" letter 12 September Amodei's essay calls for AI labs to slow down 14 September China rejects Amodei's call to slow down 23-25 September Trump-Xi summit: an AI dialogue and an incident channel 28 September Sun Chenghao: restraint "has to become observable" November First meeting of the dialogue
ChinaUnited Statesboth
Source:CNBC, Fortune (summit); IDAIS Beijing statement; IFP; Sun Chenghao.

What collaboration could look like: a ladder, cheapest first

1Incident hotlineAgreed, September 2026
2Shared red lines, common testsThe same safety tests on both sides
3Observable restraintSay what was slowed, and why
4Verification of computeKnow where the big training runs happen
5Joint developmentA "CERN for AI"
Source:IFP; RAND; Sun Chenghao. The ladder is my arrangement.
"Before restraint can become reciprocal, it has to become observable."
Sun Chenghao, Tsinghua University, 28 September 2026
Source:Sun Chenghao, "Why China doubts US calls to pause AI" (Sept 2026).

Where things stand

Frontier modelUS, by about 8 months
Diffusion through the economycontested
Installed baseChina, on open weights
TalentChina: more elite researchers now work there
Energydepends on the horizon
ChipsUS, resting on allies
China aheadUS ahead?  contested, or it depends

Thank you

Watch the videosalvozappa.comsalvozappasalvozappa
Watch the videosalvozappa.comsalvozappasalvozappa