US vs China: The AI Race
The US leads on chips deployed. China leads on the power capacity to deploy more. The race is over which constraint binds first.
Power: The Binding Constraint
China is adding 8–9× more power capacity per year than the US.
In 2024, China added 429 GW of new generation capacity vs 51 GW in the US. In 2025, China added ~540 GW vs ~60 GW. The compute race depends on the power race — and the US is losing the power race.
Total generation 2024: US ~4,300 TWh, China ~10,400 TWh (China 2.4×)
Transformer lead time: US 143 weeks, China 48 weeks (China 3× faster)
Goldman projects China will have ~400 GW spare grid capacity by 2030 — roughly 3× global data center power demand
Capacity additions include all sources (renewables, coal, gas, nuclear, hydro). US figures from EIA. China figures from National Energy Administration via Ember/CEF. Not all new capacity serves data centers, but the asymmetry in expansion rate is what binds future AI compute growth.
China National Energy Administration (NEA) annual reports
US Energy Information Administration (EIA) capacity additions data
Ember Energy China Energy Transition Review 2025
IEA “Energy and AI” Report (April 2025)
Wood Mackenzie “Powering China’s data centres” (July 2025)
Goldman Sachs China power forecasts
OpenAI Lehane letter to White House (October 2025)
But the US has built 7–8× more compute against that constraint
850,000 H100-equivalents (US) vs 110,000 (China) per Epoch AI, March 2025. The chip lead is real — the question is whether the US can keep deploying it as power becomes the bottleneck.
Source: Epoch AI GPU Clusters dataset, CC-BY 4.0. Covers ~10–20% of global cluster performance. China-side data rounded to one significant figure per Epoch’s methodology.
Capability: The Gap Is Closing
The performance gap between top US and Chinese models is now 2.7 percentage points.
Down from 17.5–31.6 percentage points in early 2023
Source: Stanford AI Index 2026 Report, Chapter 1; Arena leaderboard (lmarena.ai), March 2026. Quarterly values interpolated from annual index figures.
The Capex Paradox
$285.9B for a 2.7-point lead.
US private AI investment in 2025 was 23.1× larger than China’s $12.4B — for a model performance lead of just 2.7 percentage points.
Private AI investment, 2025
Model performance gap (Arena Elo, Q1 2026)
Note: Private investment likely understates China’s total. Government guidance funds have deployed ~$184B into AI since 2000, plus a new $138B state VC fund announced in 2025.
Source: Stanford AI Index 2026 Report, Chapter 4 (Economy/Investment).
Measured AI productivity gains: 0.5%–3% by 2030. Realized so far: ~0%.
Despite $1.6T+ in cumulative AI investment since 2013, BLS-measured productivity gains attributable to AI remain statistically indistinguishable from zero.
$1.6T invested since 2013 → ~0% measured TFP gain.
Productivity gains from AI are contested. Estimates vary by methodology. Realized gains lag investment by years. This panel shows the range of credible estimates against current measurement.
Sources: Bick, Blandin & Deming, NBER Working Paper (2024, updated 2026); Goldman Sachs Top of Mind (2024–2026); McKinsey Global Institute, Economic Potential of Generative AI (2023); BLS Multifactor Productivity series; Brynjolfsson et al., Stanford Digital Economy Lab.
Forecaster: where the race goes
Conditional projections, not forecasts. Adjust the levers; see the trajectory.
Current trends continue. All sliders at empirical defaults.
Investment
Inputs
Progress
Conditional projections, not forecasts. See methodology for model structure, parameters, and limitations.
Methodology
1. Power
China capacity additions from National Energy Administration (NEA) annual reports via Ember Energy China Energy Transition Review 2025 and China Electricity Forum (CEF). US capacity additions from EIA Electric Power Monthly. Total generation from IEA World Energy Outlook and “Energy and AI” report (April 2025). Transformer lead times from Wood Mackenzie “Powering China’s data centres” (July 2025). Grid spare capacity projection from Goldman Sachs China power forecasts. Capacity additions include all generation sources — not all new capacity serves data centers.
- China National Energy Administration (NEA) annual reports — accessed June 2025, public domain
- US Energy Information Administration (EIA) Electric Power Monthly — eia.gov — accessed June 2025, public domain
- Ember Energy China Energy Transition Review 2025 — ember-energy.org — accessed June 2025, CC-BY-4.0
- IEA “Energy and AI” Report, April 2025 — iea.org — accessed May 2025, IEA terms
- Wood Mackenzie “Powering China’s data centres”, July 2025 — accessed July 2025, proprietary
- Goldman Sachs China power forecasts — accessed May 2025, proprietary
- OpenAI Lehane letter to White House, October 2025 — public record
2. Compute
Epoch AI GPU Clusters dataset (CC-BY 4.0), accessed March 2025. Covers ~10–20% of global cluster performance; China-side figures rounded to one significant figure per Epoch methodology. Classified government compute, smaller clusters, and non-AI company workloads are not included. The 7–8× figure is indicative, not precise.
- Epoch AI GPU Clusters dataset — epochai.org — accessed March 2025, CC-BY-4.0
3. Capability
Stanford AI Index 2026 Report (Chapter 1) and Arena leaderboard (lmarena.ai), March 2026. Arena Elo reflects human preference on open-ended tasks — it does not capture specialized domains, multimodal performance, or agentic capability. Quarterly values interpolated from annual index figures.
- Stanford AI Index 2026 Report, Chapter 1 — aiindex.stanford.edu — accessed April 2026, CC-BY-ND-4.0
- Arena leaderboard — lmarena.ai — accessed March 2026, open data
4. Capex Paradox
Stanford AI Index 2026, Chapter 4. Figures are disclosed private capital only (Quid / NetBase Quid). Chinese government guidance funds (~$184B deployed since 2000) and the $138B state VC fund announced in 2025 are excluded. The 23.1× ratio shrinks substantially if those are included.
- Stanford AI Index 2026 Report, Chapter 4 (Economy/Investment) — aiindex.stanford.edu — accessed April 2026, CC-BY-ND-4.0
5. Productivity
Projections are linear extrapolations from published range estimates: BCG/McKinsey (3.0pp by 2030), Goldman Sachs Top of Mind (1.75pp), and Bick, Blandin & Deming NBER WP 2024/2026 (1.0pp). The realized data point is from BLS Multifactor Productivity series; the AI-attributable portion is estimated, not directly measured, and is within the statistical noise band. The $1.6T investment figure aggregates CB Insights and Stanford AI Index annual investment data from 2013–2025.
- Bick, Blandin & Deming, NBER Working Paper (2024, updated 2026) — nber.org — accessed May 2026, NBER terms
- Goldman Sachs Top of Mind (2024–2026) — accessed April 2026, proprietary
- McKinsey Global Institute, “Economic Potential of Generative AI” (2023) — mckinsey.com/mgi — accessed June 2023, proprietary
- BLS Multifactor Productivity series — bls.gov/mfp — accessed May 2026, public domain
- Brynjolfsson et al., Stanford Digital Economy Lab — digitaleconomy.stanford.edu — accessed May 2026, academic
Dashboard version: Phase 1 (May 2026). All data points verified against primary sources. This dashboard will be updated as new data becomes available. Methodology questions or corrections: @jakeprokopets.