The AI Race:
China vs. United States
A Decathlon, Not a Sprint
Artificial intelligence is no longer merely a technology competition — it is the central theatre of 21st-century geopolitical, economic and military rivalry. The contest between the United States and China now spans frontier model development, semiconductor sovereignty, capital allocation, and the physical infrastructure of power generation and grid capacity.
The United States retains a commanding lead in frontier model capability, private capital mobilisation, compute infrastructure and allied technology partnerships. China has erased much of the performance gap in AI model benchmarks, dominates in research volume and patent filings, leads global robotics adoption, and has constructed a decisive structural advantage in electricity generation capacity — the increasingly critical physical input to the AI race.
For investors, the divergence between US capital market dominance and China's infrastructure/industrial AI adoption story is the defining tension. The Emit Nexus thesis — the convergence of AI infrastructure capex and energy transition — is validated on both sides of the Pacific, though the mechanism, risk profile and investment vehicle differ substantially.
Technology Landscape
Model Capability
The US-China model performance gap has effectively closed. In early 2023, OpenAI held a commanding lead; by early 2025, DeepSeek R1 briefly matched ChatGPT, earning the label of a "Sputnik moment" for Chinese AI. As of April 2026, Anthropic leads overall model rankings, closely trailed by xAI, Google and OpenAI. Chinese frontier labs — DeepSeek and Alibaba — lag only modestly. The competitive landscape is now defined not by categorical capability gaps but by marginal performance differences, cost efficiency and real-world deployment depth.
The US leads in closed-source frontier models and monetisation, while China has prioritised open-source diffusion — making Chinese models disproportionately attractive to Global South operators seeking low-cost AI infrastructure independent of Western access controls.
Research Output & Patents
China now leads the world in raw AI research volume. In 2023, Chinese institutions produced 23.2% of global AI publications and 22.6% of all citations. China also accounts for 69.7% of all AI patent grants globally as of 2023 — a staggering lead that reflects deliberate state-coordinated IP accumulation. The United States retains leadership in research influence: US institutions have contributed the most top-100-cited AI publications over the past three years.
| Category | United States | China |
|---|---|---|
| AI Publications | 2nd globally (~28,400/yr) | 1st globally — 23.2% of world output |
| Citation Impact | Leads top-100 cited globally | Growing but lags in citation quality |
| AI Patent Grants | Significant but secondary | 69.7% of global AI patent grants (2023) |
| Elite AI Institutions | ~60% of world's premier AI institutions | ~10% of elite AI institutions |
| Elite AI Researchers | ~2/3 of world's elite researchers | ~10% affiliated; ~50% of top-20% born in China |
Semiconductor & Compute Infrastructure
Compute remains the single most contested dimension of the AI race. The US controls access to the world's most advanced AI chips through Nvidia, AMD and the allied semiconductor ecosystem spanning Taiwan (TSMC), the Netherlands (ASML) and South Korea (SK Hynix). Export controls imposed in 2022–2023 meaningfully constrained Chinese labs' ability to build frontier-scale supercomputers.
However, the Trump administration's January 2026 decision to permit Nvidia H200 exports to China has partially reversed this advantage. Without exports, US compute capacity in 2026 would exceed China's by more than ten times; with aggressive H200 exports, that lead could narrow sharply. China concurrently builds domestic chip capacity through Cambricon, Moore Threads and Huawei's Ascend ecosystem — constrained by manufacturing bottlenecks at SMIC (limited to ~7nm equivalent) rather than by lack of demand or capital.
| Category | United States | China |
|---|---|---|
| Leading AI Chips | Nvidia H100/H200/Blackwell, AMD MI-series | Huawei Ascend, Cambricon, Moore Threads (~7nm max) |
| Export Controls | Blackwell restricted; H200 allowed w/ tariffs | Constrained — workarounds via smuggling & domestic build |
| Data Centre Count | ~5,427 (>10× any other country) | Significant, rapidly growing |
| Data Centre Power (2030E) | 430 TWh (+144% from 2025) | 397 TWh (+255% from 2025) |
| Compute Capacity Ratio | >10× China under strict controls | Narrowing — potentially single digits with H200 exports |
Robotics & Industrial AI
Perhaps the most underappreciated dimension of China's AI advantage is its industrial deployment velocity. Chinese manufacturing enterprise AI penetration reached 47.5% in 2025, up from 9.6% in 2024 — a nearly five-fold increase in twelve months. MIIT's lighthouse factory programme reports AI penetration exceeding 70% of business scenarios at designated sites, with over 6,000 vertical-domain industrial models deployed.
Critically, 38.2% of 2025 Chinese AI venture capital was directed at robotics, not consumer applications — signalling a deliberate strategic pivot toward physical-world AI integration. This industrial AI compounding effect is the US's most significant asymmetric risk.
Capital Markets
Venture Capital & Private Investment
The US private capital ecosystem for AI is operating at an unprecedented scale. In 2025, global AI venture capital totalled USD 258.7 billion — representing 61% of all VC investment globally, double its 2022 share. US firms captured approximately 75% of this figure (USD 194 billion), reflecting the concentration of frontier AI development, LLM commercialisation and infrastructure scaling within American borders.
Q1 2026 shattered records: US-based AI companies raised USD 250 billion, representing 83% of global venture funding in the quarter. The second-largest market, China, raised USD 16.1 billion in Q1 2026 — significant in absolute terms but representing a fraction of the US figure. The concentration in mega-deals is striking: rounds exceeding USD 1 billion now represent roughly half of total AI investment value globally.
Public Markets & Valuations
The valuation asymmetry between US and Chinese technology companies is extreme. As of January 2026, the lowest-valued Magnificent Seven member — Tesla — carries a market cap of USD 1.46 trillion. The combined market capitalisation of China's BATX group (Baidu, Alibaba, Tencent, Xiaomi) totals USD 1.22 trillion — less than Tesla alone, 30% of Apple's value and 27% of Nvidia's.
However, BATX companies grew their market capitalisation approximately 3.5 times faster than the Magnificent Seven from Q3 2024 to Q3 2025. Hong Kong has emerged as the world's largest listing destination for Chinese AI companies in 2025-2026, with foundation model firms Z.ai and MiniMax debuting on the HKEX in January 2026 at valuations exceeding USD 6 billion each.
| Category | United States | China |
|---|---|---|
| Market Cap (Top 7) | Magnificent 7: collectively >USD 12T | BATX combined: ~USD 1.22T |
| Lowest Benchmark | Tesla: ~USD 1.46T | BATX combined < Tesla alone |
| Relative Growth (Q3 24–25) | Strong but measured | 3.5× faster market cap growth vs Mag-7 |
| Primary Listing Venue | NYSE / NASDAQ | Hong Kong (HKEX) for new AI IPOs |
| State Market Support | None direct | RRR cuts, swap programmes, relending facilities |
| Valuation Risk | Elevated P/E multiples; capex correction risk | Discount to US; upside from stimulus & re-rating |
Investment Thesis Implications (Emit Nexus)
The capital markets dimension of the AI race directly reinforces the Emit Nexus thesis. On the US side, the data centre infrastructure capex cycle — spanning compute (Nvidia, AMD, Broadcom), power infrastructure (Eaton, Vertiv, GE Vernova, Quanta Services) and power generation (Constellation Energy, Vistra, NRG) — represents the dominant equity investment opportunity. The scale of capital flowing into these verticals is self-reinforcing: hyperscaler capex commitments underwrite infrastructure earnings visibility for 5–10 years.
On the China side, the robotics and industrial AI deployment story — underpinned by cheap energy, state-directed capital and manufacturing scale — argues for selective exposure to Chinese technology platforms with genuine AI integration advantage. The BATX re-rating trajectory is compelling on a relative value basis, though geopolitical risk premia remain the appropriate discount.
Energy Security
Power as a Strategic Asset
Energy availability has supplanted chip access as the binding constraint on AI development velocity. A single AI-related computational task can consume up to 1,000 times more electricity than a traditional web search. Global data centre electricity consumption reached approximately 415 TWh in 2024 — 1.5% of total global electricity use, growing at 12% CAGR since 2017. By 2026, data centres may approach 1,050 TWh, placing them fifth among global energy consumers.
"Everywhere we went, people treated energy availability as a given," noted tech researcher Rui Ma after touring China's AI hubs in 2025. In the US, surging AI demand is colliding with a fragile power grid — the kind of extreme bottleneck that Goldman Sachs warns could severely choke industry growth. In China, it is considered a solved problem.
China's Energy Dominance
China's electricity infrastructure advantage in the AI race is its most underappreciated strategic asset. Between 2010 and 2024, China's power production increased by more than the rest of the world combined. In 2024 alone, China added 543 gigawatts of new power capacity — more than the total accumulated by the US in its entire history. China's national reserve margin has never fallen below 80-100%, meaning it consistently maintains roughly twice the capacity it actually needs.
By 2030, China is projected to have approximately 400 gigawatts of spare power capacity — triple the expected power demand of the global data centre fleet at that time. China installed 357 GW of new wind and solar capacity in the first half of 2025 alone — more than India's entire installed power capacity. Renewables now account for 60% of China's total installed power capacity.
US Energy Vulnerabilities
The US power grid faces a structural mismatch between AI data centre deployment pace and grid infrastructure capacity — most of it built between the 1950s and 1970s. Approximately 70% of the US grid is approaching end of life. High-voltage transformers carry manufacturing lead times of up to four years and costs six times their pre-2022 levels, creating a structural bottleneck in the grid upgrade cycle that cannot be resolved by capital alone.
Natural gas remains the most viable near-term bridging fuel. Nuclear — both fleet life extensions and new SMR deployment — represents the cleanest long-term solution but involves 5–10 year development timelines. In the interim, the "speed to power" imperative is driving AI infrastructure investment toward power-rich geographies: Louisiana, Alberta, UAE and other regions with grid headroom.
| Metric | United States | China |
|---|---|---|
| Power Added (2024) | Modest incremental additions | 543 GW — more than US total history |
| Reserve Margin | Fragile — grid at limits in AI hubs | Consistently 80–100% surplus nationwide |
| Spare Capacity (2030E) | Insufficient for projected AI demand | ~400 GW spare — 3× global DC fleet need |
| Renewables Share | Growing; hampered by permitting delays | 60% of installed capacity; 357 GW added H1 2025 |
| Data Centre Power (2026E) | ~260 TWh (~6% of US electricity) | ~6% share but from much larger supply base |
| Energy Strategy | Market-driven; reactive to demand | Deliberate overbuilding — energy as strategic asset |
Power Grid Infrastructure
Grid Architectures: A Fundamental Divergence
The structural divergence between Chinese and US power grid architectures reflects decades of different policy choices. China built its grid top-down, with centralised state planning enabling rapid overbuilding of both generation and transmission. The result is a system with structural surplus at every layer — generation, transmission and distribution — that is well suited to absorbing the concentrated, volatile load profiles of AI data centre clusters.
The US grid evolved bottom-up through a mix of regulated utilities, independent power producers and deregulated markets. This model drove efficiency and cost competition over decades but systematically under-invested in reserve capacity. The transition from a stable demand profile (2000–2020) to AI-driven hypergrowth has exposed the fragility of just-in-time grid planning.
Data Centre Demand: The New Urban Load
A modern hyperscale AI data centre consumes 500 MW to over 1 GW of power — comparable to a mid-sized city. Global data centre power demand is projected to reach 1,596 TWh by 2035 — a 255% increase from 2025 levels. The US is expected to retain the largest share at 430 TWh (+144% from 2025), while China closes the gap rapidly at 397 TWh (+255%). Together, the US and China account for approximately 80% of projected data centre power demand growth through 2030.
| Metric | United States | China |
|---|---|---|
| Grid Age & Condition | ~70% approaching end of useful life | Modern, centrally planned, recently overbuilt |
| Reserve Margin | Critically low in AI hub markets | 80–100% surplus nationally |
| New Capacity (2024) | Incremental renewables + some gas | 543 GW total — unprecedented |
| Interconnection Backlog | Severe — years-long queue in many regions | Minimal — centralised approval, fast deployment |
| AI Hub Strategy | Power constraint forcing geographic diversification | "Eastern Data, Western Computing" — planned distribution |
| Transformer Lead Times | Up to 4 years; 6× cost inflation | Domestically manufactured; shorter lead times |
| Nuclear Contribution | Fleet life extensions; SMR early-stage | Aggressive new build + thorium R&D |
Grid Investment Outlook
The US near-term (2026–2030) power investment cycle is characterised by pragmatic "speed to power" solutions: accelerated solar-plus-battery deployments, gas turbine procurement, nuclear life extensions and behind-the-meter generation by hyperscalers. McKinsey projects USD 6.7 trillion in global data centre investment between 2025 and 2030 — of which power infrastructure represents a material and growing component.
For investors, the US grid constraint story is bullish for a specific set of assets: existing thermal generation, power infrastructure manufacturers (transformers, switchgear, HV cable), and IPPs with secured interconnection positions. The private-to-public arbitrage remains active: private transactions clearing at 6–9× EV/EBITDA while public IPPs trade above 11×, creating room for accretive consolidation.
Strategic Assessment
The Decathlon Framework
The AI competition between the US and China is not a sprint toward a single finish line. As Foreign Affairs has framed it, the two nations are competing in a "decathlon" — multiple simultaneous races with different leaders in different disciplines. The US leads in frontier model capability, capital markets depth, allied semiconductor partnerships and commercial AI monetisation. China leads in research volume, patent filings, energy infrastructure, robotics deployment, industrial AI adoption velocity and open-source model diffusion.
The US maintains more powerful AI models, more capital, and an estimated 5,427 data centres (more than 10× any other country). China leads in AI research publications, patents, and robotics. The nations have entered what Stanford's Colin Kahl calls "an asymmetric form of AI bipolarity."
Key Risk Factors
- US compute advantage erosion: The Trump administration's H200 export relaxation could narrow the compute gap from >10× to single digits within 24 months.
- US grid constraint: Power unavailability is now a commercial barrier to AI infrastructure deployment — not merely a theoretical risk. This directly constrains US AI capability scaling.
- China industrial AI compounding: The near-5× increase in enterprise AI penetration in Chinese manufacturing (9.6% → 47.5% in 2025) is a leading indicator of structural productivity divergence that may be difficult to reverse.
- Geopolitical escalation: Sustained Taiwan Strait tension and ongoing Middle East conflict represent tail risks that could rapidly restructure global semiconductor supply chains.
- US talent vulnerability: 68% of Chinese AI PhD graduates migrate to US institutions — a structural advantage. Visa restrictions or academic policy changes could reverse this brain drain.
- China capital markets maturity: Hong Kong HKEX's emergence as the primary listing venue for Chinese AI firms reduces the information advantage Western investors previously held.
Portfolio Positioning (Emit Nexus)
| Theme | US Equity Exposure | China / EM Exposure |
|---|---|---|
| AI Compute | NVDA, AMD, AVGO, MRVL | Selective HKEX-listed AI hardware |
| Data Centre Infrastructure | VRT, ETN, EQIX, PWR, DELL | Chinese industrial AI integrators |
| Power Generation | CEG, VST, GEV, NRG | State-owned power utilities (selective) |
| Grid Infrastructure | ETN, PWR, GEV, AES | Chinese transformer / grid OEMs |
| Nuclear | CEG, CCJ (uranium) | China's aggressive new-build pipeline |
| Robotics / Industrial AI | Limited pure-play exposure | CATL ecosystem, Midea, DJI adjacents |
Conclusion
The AI race between the United States and China is entering its most consequential phase. The performance gap at the model frontier has effectively closed; the semiconductor advantage is eroding under policy relaxation; and the energy infrastructure gap — long underappreciated — is emerging as the race's defining physical constraint.
The United States retains decisive advantages in private capital mobilisation, allied technology partnerships, commercial monetisation and talent attraction. China retains decisive advantages in energy infrastructure, industrial AI deployment velocity, research volume and open-source model diffusion. Neither advantage is insurmountable for the other side, but both are durable over a 3–5 year horizon.
For investors applying the Emit Nexus framework, the most important conclusion is this: the AI infrastructure capex supercycle is real, durable and physically constrained by energy. Power — its generation, transmission and security of supply — has become as strategic as the semiconductor itself.
Both nations understand this. The race to power AI is, in the end, a race to power. Emit Capital's positioning at the intersection of AI infrastructure and energy transition reflects this conviction. The companies that solve the power constraint — through nuclear, gas, grid modernisation or efficiency innovation — will be among the defining businesses of the next decade on both sides of the Pacific.
Disclaimer: This research paper has been prepared by Emit Capital Asset Management Pty Ltd (AFSL 551084, ABN 57 652 326 227) for informational purposes only. It does not constitute financial advice and is intended solely for wholesale investors as defined under the Corporations Act 2001 (Cth). Past performance is not indicative of future results. Data sourced from Stanford HAI 2026 AI Index, OECD AI VC Report 2026, Brookings Institution, Crunchbase, CSIS, Foreign Affairs, Fortune and MIT Technology Review. April 2026.