AI RACE (PART 2): CHINA NEVER PLANNED POWER FOR AI - IT WAS ACCIDENTAL AND HAS NOW HIT CONSTRAINTS
In the AI Race based on Authority (Intelligence) and Scale (Deployment), China wins, having reached Quadrant III defined as having achieved civilisational deployment (See AI Race - Part I). China is challenged in frontier breakthrough models, but it beats the rest in coverage with "always on" systems and running everywhere. It has arrived at the stage for possible high payoffs that the rest are struggling to reach.
Many point to China's trump card in the race - their massive power capacity. Even folks like hedge fund manager Michael Burry, CEO of Scion Asset Management point to the "one chart that beats them all" (chart showing China's energy capacity is twice that of EU and US combined).
China did not plan power supremacy for AI dominance. AI is an opportunistic beneficiary, not the cause.All countries build electricity to meet demand. China builds electricity to eliminate vulnerability. AI emerged inside a power system built for regime security, industrial scale, and sanction resilience. Here's a timeline of China's power development that makes this point obvious.
- 1980s-90s: Chronic power shortages. Electricity was a regime-critical infrastructure.
- 1993: China became a net oil importer. Faced energy shock.
- 2001-07: WTO entry led to rapid industrialisation and energy demand surged. China started massive coal build and national grid formation.
- 2008-12: The global financial crisis changed China's development model permanently. China went on a build-build-build rampage. Over-capacity is no longer a bug - it is a feature. Electricity becomes an employment sink, a demand stabiliser and a strategic reserve. Coal, hydro, nuclear plant approvals surged. Grid expansion accelerated regardless of short-term demand. When the 2011 Fukushima nuclear plant accident happened, all countries paused, China redesigned the nuclear programme, upgraded safety features, and continued.
- 2013-2017: Xi Jinping reframed energy as a national security domain. The objective was to reduce vulnerability to maritime blockade and reduce fuel import shocks. The idea was to make electricity sovereign, controllable, domestic. In 2013 the Belt & Road Initiative included energy corridors, 2014 the "Energy Production and Consumption Revolution" policy, 2015 China had their indigenous Hualong One nuclear designs, UHV network scales nationwide, Hydro and coal generation in the West shipped power to the Eastern parts of the country. China is the only country to deploy UHV at national scale - it can move massive inland power securely to coastal industry at continental scale, creating surplus and resilience.
-2018-2020: Trade war and tech sanctions clarified the threat environment to the Chinese. They now see electrification provides insulation from oil and gas coercion. The power system is to be designed for resilience under sanctions. Electrification of industry and transport accelerated. Renewables scaled on top of coal and nuclear. China deliberately overbuilt with redundancy.
- 2020-2022: China chooses supply certainty over efficiency. The strategic logic: - "blackouts cause unrest" is doctrinal; political cost of excess capacity is preferable over cost of shortage. This led to the view power infra as macro stabiliser,. The leadership doubled down on power generation to prevent shortages, Coal plant approvals surged. China's grid hardened, catering for redundancy, reserve margins and regional self-sufficiency.
- 2023-present day: AI arrives into an existing power surplus. When AI arrived, China already has:
* world's largest power generation plants.
* world's largest grid capable of taking on excessive reserves.
* chronic overcapacity in multiple provinces.
Bottom line: The West builds electricity to meet demand. China builds electricity to eliminate vulnerability. This is state-engineered surplus driven by security ideology, fully prepared to accept the inefficiency of excessive surplus. It is basically industrial war logic. This is not Wall Street style ROI discipline, not tech exuberance. AI came at the right moment and data centres plug into ready-made pre-existing slack.
What is China's installed power, potential generation capacity and actual generation in 2025.
Installed power, or nameplate power, is the maximum electrical power a generator or plant is designed to produce under ideal, specified conditions. It is a theoretical power measured by GW (gigawatts).
Potential generation capacity is the maximum it can actually generate. In reality, all generators never run at 100% capacity for various reasons and constraints. Hydro-plants normally run at 35-60%, nuclear plants at 85-95%, coal at 50-65%, wind at 25-55% and solar at 10-25% capacity.
China's total installed capacity or nameplate by 2025 is about 3,590 GW.
To convert power capacity to energy ---- multiply GW by 8,760 (hours per year)
Thus 3,590 GW = 31,448,400 GWh or 31,448 TWh theoretical maximum. (Should be higher since we do not known the GW of 'Others'.)
Actual production is based on demand. In 2025, total consumption was 10,087 TWh.
Potential generation is the highest level it can actually achieve. We want to know the potential to see how much energy China can ramp up if it wanted to. China's energy policy is renewables first and coal last. All the others are maxed out, but coal is producing at 47% capacity below its potential due to surplus generation. Let's compute for coal based on average potential in the industry -- 57%. As for Hydro, generation was affected by drought so the 38% was on the low side. Let's compute based on industry average of 47%. Based on these assumptions, China's potential generation capacity is 11,396 TWh.
Currently, China has a slack of 11,396 less 10,087 = 1,309 TWh. (By comparison, US slack is 500 TWh)
Let's ignore future demands from competing sectors (industry, residential, military) and see how much this 1,309 TWh slack can support new AI deployments.
AI life cycle has 2 phases -- training and inference. The energy demand is different at each phase.
Training Phase: The model is built and learns patterns and relationships by processing vast amounts of historical data. Energy usage is in bursts when the model runs. Think of training as - learning.
Inference Phase: The model is deployed into an operational environment. When a user makes a querry, the model takes in new, live data and uses its learnt parameters to draw conclusions (making an inference) quickly and efficiently. Energy usage is continuous once the model is deployed. Think of inference as - using.
To know what is the energy consumption of AI we need to look at the demand side and the supply side.
The demand side is the inference and is affected by various factors such as the specifics of the AI cluster, the penetration of the model, per capita usage, etc. China's population is almost 5 times more than the US and one instinctively think its energy usage for AI is higher. That is not the case. China has more AI clusters than US but demand side energy use is lower because Chinese AI deployment is mostly in industry. US AI is used globally.
For our purpose here, let's not get side-tracked into the complexities of demand side energy use. Let's just compute the supply side. An AI cluster comprises of computer hardware, network systems, storage systems, orchestration and management software, and power and cooling systems. These clusters vary in capacities and are energy intensive best measured by its Nameplate capacity which is the maximum designed power draw of the facility.
China has 250 AI clusters with average nameplate capacity of 0.1-0.2 GW. Let's assume the average is 0.2 GW and the Utilisation Capacity Factor is 90%.
So the energy consumption per cluster is 0.2 x 8,760 x 90% = 1,578 GWh = 1.576 TWh.
The number of additional AI clusters that the slack can support is 1,309 TWh ÷ 1.576 TWh = 850 clusters.
Many look at China's power capacity and surplus situation and the massive energy needs of AI and say the Chinese have the AI race in their hands.
But is that the reality? Not so fast.
All electricity systems need to retain reserves for peak demand and safety margins. The reserves energy is not generated, but the volume is committed and ready for dispatch anytime. There is a carry cost for reserves. When you pay for your electricity bills, the cost of the reserves is embedded in the invoice. The higher the level of reserves, the more you pay.
Direct data on minimum (baseload) demand is not explicitly reported in official 2025 statistics, but it can be estimated from load curve characteristics which shows a peak of 1,500 GW, the minimum baseload is approximately 975 GW. Annual Baseload in TWh = 975 GW × 8.76 (8,760 hours / 1,000 for TWh conversion) = 8,541 TWh.
The required minimum PRM (Planning Reserve Margin) is 15%, based on national reliability standards to cover peak load plus contingencies like plant outages or demand spikes.
Required reserves = peak 1,500GW x 15% = 225 GW
Translated to TWh = 225 GW x 8,760 = 1971 TWh
What it means is of the slack of 1,309 TWh (149 GW), an amount of 1,971 TWh (225 GW) is withheld for reserves. We end up with a negative 662 TWh. China's electricity system is in effect running with insufficient safety margin of reserves.
For decades, China has prioritized overbuilding power infrastructure as a cornerstone of national energy security. This stems from post-1978 reforms under Deng Xiaoping, emphasizing science and technology for self-reliance, and evolved through five-year plans that encouraged surplus to buffer against shortages, geopolitical risks (e.g., fossil fuel imports), and demand spikes. By the 2010s, this manifested in massive coal capacity (peaking at ~1,200 GW) and interprovincial grids to redistribute surplus from resource-rich areas.
The 2017 "New Generation Artificial Intelligence Development Plan" marked a national priority shift, aiming for global AI leadership by 2030 and accelerating commercial/large-scale deployments (e.g., via Baidu, Huawei). The power surplus already built-up for strategic energy security reasons was foundational accidental-fit for energy-intensive AI. The success and rapid deployment of AI precipitated a surge in energy demand, particularly 2023-2025, which has eroded this surplus.
But wait, there is more. And this is something most analysts miss.
For China, decarbonisation is a constraint managed, not a goal pursued. The real objective is energy sovereigntyElectrify everything! China's national energy and industrial strategy increasingly emphasises electrification of final energy use as a core pillar of its energy-security and industrial policy. We see this push in transport (electric rails, EVs, drones), industry, buildings and more. EV adoption is pushed with massive state subsidies and infra build. EV is both electrification policy and tech race. In industry we see electrification in high-temperature heat pumps, electric heating, electric boilers, etc. In buildings we see clean heating, electric end-use technologies. Electrify everything is not just slogans - it is embedded in official guidance documents. But this is not about going green, it is about energy sovereignty.
China's strategy for dual target of net carbon neutrality by 2060 and energy sovereignty is to use VRE (Variable Renewables Energy) first and develop hydro, nuclear and coal to meet baseload. Coal is the preferred reserves because it is domestic and inventoriable, thus creating tension with overall targets.
With VRE (solar and wind), there is a problem. It cannot provide continuous electricity because there may be cloudy and rainy days and night, and there may be a drop in wind velocity. New storage technologies provide some backup for solar but it is not 24 hours. What this means is renewables cannot be on its own. Coal, hydro or nuclear are still needed to provide reserves for intermittent renewables. The more renewables there are in the generation mix, the more reserves are required.
In 2025, renewables form 33% of China's energy mix. The potential capacity is 2,858 TWh. The question is, how much VRE reserves are required? There are complex computations and IEA guidelines which we avoid getting into. The ballpark figure is about 200-400 TWh reserves is required. For our purpose let's take the ballpark mid-point figure of 300 TWh.
So when computing actual surplus energy capacity of China currently, we take the slack less the PRM reserves less NRE reserves:
Reserves required = 1,971 (for PRM) + 300 (for VRE) = 2,271 TWh
Slack is 1,309 TWh.
Actual surplus = 1,309-2,271 = - 962 TWh.
Currently, far from having abundant energy surplus to power the AI Race, China actually has negative or insufficient slack to even meet its reserves requirements.
Not having sufficient reserves exposes China to risks of blackout and power outages, energy shortfalls and reliability issues, cascading system failures, grid instability from renewables, increased vulnerabilities to extremes, and economic and societal losses. The saving grace in the short term is the economic slowdown the country is currently facing will see a decrease in energy demand.
So how much does deployed AI drain the surplus?
China currently has 250 AI clusters with average nameplate of 0.2 GW making a total footprint of 50 GW. Assuming 90% Utilisation Capacity, the energy needed is :
Supply side : 50 GW x 90% x 8,760 = 394,200 Gwh = 394 TWh.
On the demand side, ie the inference, industry analysts indicate it can add extra effective load by 10-25%. Since China's AI deployment is mostly in the industry side, let's use the lower 10% additional load.
Demand side : 394 x 10% = 39 TWh
Thus the total energy draw from AI is 394 TWh + 39 TWh = 433 TWh.
China didn't plan for intensive energy demands of AI. For decades it has been building substantial surplus power for security reasons. Then AI came along and the power surplus supported the development and deployment nicely. By 2025 AI deployment reduced the surplus by 433 TWh. In addition, Xi Jinping's 2017 policy to build up renewables further bites into the surplus for VRE reserves. By 2025 VTE reserves further reduced surplus by 300 TWh. Power constraints emerged in 2024-2025, causing some AI projects to be delayed due to grid limits.
At this juncture, like all countries in the AI race, China is actually facing power constraints. There is no way China can cross the threshold into the Holy Grail of AI which is the Quadrant IV of authority and scale unless it builds more massive power .
Anyone that shows the "one chart that beats them all" to explain China's power capacity alone wins the AI race for the country does not believe in the adage that the devil is in the details.
This platform has withdrawn it's subscriber widget. If you like blogs like this and wish to know whenever there is a new post, click the button to my FB and follow me there. I usually intro my new blogs there. Thanks.