China's Cheap Energy Is Winning the AI Data Center Race — And US Curbs Helped

China has an AI infrastructure advantage that Western analysts have been slow to price in: electricity is cheap, abundant, and becoming more so. While US and European hyperscalers negotiate multi-gigawatt power purchase agreements at prices that have tripled over three years, Chinese data center operators are drawing from a grid where coal, hydroelectric, and nuclear power combine to produce some of the world's lowest industrial electricity rates. The result, as Al Jazeera reported in late May 2026, is that China is building AI infrastructure faster than any country in history — and the energy cost advantage is a structural factor, not a temporary one.

Alongside that, SMIC — China's largest chip manufacturer — reported record revenue for 2025 at $9.3 billion, up 16% year-on-year, with projections exceeding $11 billion in 2026. The company that US export controls were designed to constrain is having its best years ever.

How Cheap Is China's Electricity?

China's industrial electricity price averages roughly $0.05–0.07 per kilowatt-hour in many provinces, compared to $0.07–0.12 in the US (varying significantly by state and utility) and $0.15–0.25 in much of Western Europe. For an AI data center consuming 100 megawatts continuously, that gap translates to tens of millions of dollars in annual operating cost difference.

The structural reasons are multiple. China's coal-heavy generation fleet, while environmentally costly, is domestically fueled and priced in renminbi. Hydroelectric capacity in Sichuan and Yunnan provinces — where several major data center clusters have been built — produces electricity at near-zero marginal cost. The Three Gorges Dam alone generates more electricity annually than all of Australia's installed capacity. Nuclear expansion continues at a pace that no Western country is matching: China has more nuclear reactors under construction than the rest of the world combined.

China's Data Center Buildout: 30% Annual Growth Since 2016

China's data center rack count grew at 30% annually from 2016 to 2023, according to the China Academy of Information and Communications Technology. That growth rate, sustained over seven years, means China's data center capacity has roughly sextupled in less than a decade.

The AI boom has accelerated the buildout further. ByteDance, Alibaba, Baidu, Huawei, and Tencent are all expanding capacity aggressively. The combination of cheap energy, government support (data centers are designated strategic infrastructure), and domestic chip suppliers means China's AI infrastructure can grow without the bottlenecks that have plagued Western expansion — specifically the power interconnection queues and permitting timelines that add years to US data center deployment.

SMIC's Record Revenue: The Export Control Paradox

SMIC posted 2025 revenue of $9.3 billion, a 16% increase from 2024, and is projecting revenue above $11 billion in 2026. This is the company that US export controls have tried, since 2020, to prevent from accessing advanced manufacturing equipment.

The paradox is not complicated once you map the incentives. US export controls on advanced chips and equipment pushed Chinese companies to buy whatever domestic alternatives existed, even at inferior specifications. That demand validated SMIC's business case, funded capacity expansion, and gave the company the revenue base to invest in R&D. Chinese fabless designers who previously used TSMC for leading-edge work shifted programs to SMIC for mid-range nodes — not because SMIC matched TSMC's capabilities, but because the alternative (relying on a foreign foundry that could be cut off by US policy) was strategically unacceptable.

The result: SMIC in 2025 is a significantly larger and more capable company than SMIC in 2019, when export controls began tightening. Revenue has roughly tripled. The specific advanced nodes it cannot manufacture remain blocked, but the company's foundational business — 28nm, 40nm, and older process nodes that cover the vast majority of chips by volume — is booming.

Chinese chip firms more broadly reported record revenues in 2025, driven by the same dynamic. US restrictions on Nvidia GPU exports forced Chinese AI companies to design around domestic alternatives, which funded those domestic alternatives' development cycles.

The Huawei Signal

Huawei's Tau Scaling Law announcement on May 25, 2026 (covered previously on this blog) is the next data point in this trajectory. A design-level innovation that claims 55% higher transistor density at existing process nodes — if validated — means that the performance gap between sanctioned and unsanctioned chip manufacturing is being attacked from the architecture side rather than the process side. SMIC's manufacturing capacity, combined with Huawei's design innovation, is the combination that US policy was designed to prevent from closing the gap.

For the technical details of Huawei's approach, see Huawei Tau Scaling Law: China's Moore's Law Alternative Without EUV. For how Beijing is weaponizing materials exports against US allies, see China Bans Dual-Use Tech Exports to Japan Military Over Taiwan Remarks.

What This Means for the AI Race

Three implications for anyone building AI infrastructure or tracking the competitive landscape:

Energy cost is becoming a primary AI infrastructure variable. Training large models requires massive compute sustained over weeks. At scale, the electricity cost of a training run is a significant fraction of total compute cost. China's energy advantage is not just a data center cost story — it is an AI training economics story.

US export controls are working in a narrow band. They have successfully denied China access to the very leading-edge nodes (sub-5nm) required for the most advanced GPU architectures. They have not prevented China from building a large, capable, and growing domestic semiconductor industry that serves the majority of chip demand by volume.

Chinese AI companies are not compute-starved in the way 2022 assumptions suggested. The combination of older Nvidia GPUs (stockpiled before restrictions tightened), domestically produced Huawei Ascend chips, SMIC-manufactured processors, and cheap energy means Chinese AI labs can run large-scale training and inference at competitive cost. The gap is real but narrower than the export control narrative implies.

Key Takeaways

• China's industrial electricity price: $0.05–0.07/kWh — significantly below US ($0.07–0.12) and Europe ($0.15–0.25)
• SMIC 2025 revenue: $9.3 billion — up 16% YoY, projected $11B+ in 2026 — record despite US export controls
• China data center racks: 30% annual growth since 2016 — capacity roughly 6x in less than a decade
• Export control paradox: US restrictions on advanced chips redirected Chinese demand to domestic suppliers, funding SMIC's expansion
• For developers: energy cost is now a primary AI training economics variable — Chinese labs have a structural cost advantage in sustained large-model training
• What to watch: Huawei Ascend performance benchmarks against Nvidia H100 equivalents; SMIC 2026 Q1 earnings; US policy response to SMIC revenue growth

Sources

• China's cheap energy as secret weapon in AI race — Al Jazeera, May 28 2026
• Chinese chip firms hit record revenue driven by AI boom and US curbs — CNBC
• AI spending nears $800 billion, Semicon China 2026 — Digitimes