As the US and China compete for AI dominance, energy is emerging as the critical battleground, and China’s more integrated approach to power, land, and permitting may deliver a durable structural edge for domestic AI buildouts. While chip procurement and export controls dominate the headlines, the near-term speed limit on AI capacity is increasingly set by megawatts and grid access rather than accelerators and model architectures alone. BloombergNEF projects that US data centers will account for roughly 20% of the nation’s electricity consumption by 2035, up from 5.9% today, a figure that underscores how profoundly the AI buildout is reshaping energy demand in the world’s two leading AI economies.
Why power is now AI’s binding constraint
AI’s energy appetite is compounding as models scale, inference proliferates across products, and training cycles compress. Global data center electricity demand is projected to roughly double by 2030, from around 485 TWh in 2025 to approximately 950 TWh, driven in part by AI workloads, according to the International Energy Agency’s latest outlook, underscoring a widening gap between power-hungry build plans and grid readiness in key markets. Even with aggressive efficiency gains in power usage effectiveness and new cooling techniques, the numerator, total compute, is growing faster than the denominator.
China is not immune to this tension. As EastFrontier reported in June, operators face a “green energy paradox”: political pressure to green AI infrastructure collides with the reality that the cheapest, firmest power remains thermal in many regions, while renewable integration, curtailment, and transmission remain uneven. The result is a strategic question: who can secure, shape, and schedule power at scale, at predictable prices, while staying within emissions targets?
Beijing’s integrated playbook: compute parks wired to generation and grid
China’s approach stacks industrial policy levers across the energy and compute domains to assemble turn-key “energy-compute” clusters. The approach builds on the “East Data, West Computing” pattern, placing compute in resource-rich interior provinces and moving bits over fiber while moving electrons across provinces via ultra-high-voltage (UHV) lines. Under the 15th Five-Year Plan, provinces have signaled tighter coupling of data center siting with access to dedicated renewable portfolios, long-term power purchase agreements, and flexible thermal capacity to firm intermittent output.
This is not merely about building more solar or wind. It is about orchestration: aligning grid upgrades, load siting, inter-provincial transmission, and permitting on a synchronized schedule so compute parks receive both capacity and priority interconnection. Such central coordination can compress timelines significantly. Instead of developers securing land, then grid studies, then power purchase agreements in sequence, provincial authorities package land, power quotas, water, and fiber into single-window deals, a bundling approach that is harder in fragmented markets but aligns with China’s broader industrial policy style. As EastFrontier analyzed in its coverage of the 15th Five-Year Plan’s AI-energy provisions, this integration is codified at the policy level, not just practiced informally.
The US contrast: grid bottlenecks and permitting drag
In the US, the AI buildout is increasingly entangled with grid interconnection queues, local permitting disputes, and uneven transmission planning. Developers are gravitating toward sites with immediate firm power, often near existing gas plants or in regions with surplus capacity, while pursuing long-dated power purchase agreements that may not fully match the hour-by-hour profile of AI loads. Ambitious nuclear timelines and advanced geothermal pilots could help later in the decade, but near-term additions risk skewing toward gas unless transmission expansion accelerates and storage scales.
This fragmentation creates a different cost curve. Where China can socialize the cost of UHV lines and co-plan data center clusters with provincial energy roadmaps, US operators face bespoke negotiations across utilities, regulators, and communities. The result: longer lead times, higher soft costs, and more variance in delivered prices per megawatt. That delta compounds when translated into cost-per-token for inference or cost-per-training-day for foundation models.
Energy and chips as converging strategic variables
Energy and chips are converging strategic variables. If China can provision large, predictable blocks of low-cost power, it can sustain high utilization across domestic accelerators and mitigate some performance gaps with scale and scheduling. This link is already visible in procurement behavior: chip roadmaps are being co-optimized with power and cooling footprints, and operators are adopting energy-aware schedulers that align training bursts with renewable availability to cut marginal costs.
For a look at how Beijing is trying to close the silicon gap while power planning advances in parallel, see EastFrontier’s deep dive Inside China’s all-out push to catch up with American AI chips. The interplay is two-way: a stronger domestic chip supply chain increases the return on each megawatt of provisioned power, while cheaper, cleaner electricity improves the economic case for more accelerators. The AI race will increasingly be won in ministries of energy and grid control rooms as much as in research labs and chip fabs, and on that front, Beijing is moving to design the system end to end.