Alibaba Cloud discusses China's AI infrastructure power constraint, analyzing total supply, grid flexibility, costs, and power-compute coordination, warning per-unit energy costs will rise long-term.

In recent years, whenever AI comes up in conversation, you can't escape one word: electricity. Investors worry about power rationing, media outlets run stories on "electricity shortages," and even local investment promotion departments put "green electricity quotas" front and center. Everyone assumes the same thing: China's AI Achilles' heel is "insufficient power supply."

That consensus was directly challenged by someone who manages data centers. Wang Zhaoyang, General Manager of Alibaba Cloud's Global Data Centers, put it bluntly in an interview: China doesn't lack total electricity capacity; what's missing is the coordination to align electricity and computing power. As someone at Alibaba Cloud who oversees both the computing load side and their own data centers, this assessment carries his business perspective—but his direction of argument is worth hearing.

My own judgment goes one step further: whether total capacity is sufficient is an old question; the real hard constraint is that the energy bill per unit of computing power will only grow heavier over the long term.

**First ledger: Generation exceeds US, Europe, Japan combined by 24%, but "local" bottlenecks will lock us up**

On total capacity, Wang Zhaoyang mentioned this observation: China's electricity generation already exceeds the combined total of the United States, Europe, and Japan. The direction is sound. According to the National Bureau of Statistics, China generated 10,575.25 billion kilowatt-hours in 2025 (approximately 10.58 trillion), up 4.8% year-on-year. The combined output of the US, Europe, and Japan totals roughly 8.5 trillion—this denominator uses industry standards (Wind/BP series) and isn't perfectly comparable to China's statistical methodology, but the magnitude is clear: we generate 24% more than their combined total.

China clearly has sufficient total capacity. But the problem lies in "concentration."

Data centers aren't scattered across the map. Wherever there's computing demand, cheap electricity, and available land, facilities cluster—Ulanqab has a mean annual temperature of just 4.3°C, green electricity comprises 67% of its supply, and network latency to Beijing compresses to 2.1 milliseconds, so data centers naturally concentrate there. But once concentrated, local contradictions follow.

Wang Zhaoyang made a piercing observation: once local contradictions reach a large enough scale, they become "irreplaceable." Meaning: once a hub's electricity, land, and transmission corridors fill with computing load, you can't simply relocate that entire workload elsewhere and start fresh—like those few perpetually congested transit hubs during holiday travel rushes, where national transportation capacity is adequate, but Beijing West Station can't be moved, and once it breaks down the entire network seizes up. This is a structural problem, not a total capacity problem.

This follows the same logic as coal transport from Xinjiang and the Da-Qin Railway—the country doesn't lack coal, the bottleneck is the corridor from Xinjiang to coastal power plants; the country doesn't lack electricity either, the bottleneck is those few local nodes where computing power concentrates. Fundamentally, it's the same condition across different scenarios: total capacity is adequate, but corridors are constrained; surplus total capacity paired with structural misalignment.

Adequate total capacity doesn't mean we're clear.

**Second ledger: Flexibility—Data centers aren't rigid loads, but flexibility comes in degrees**

The second ledger defies conventional wisdom.

The industry commonly assumes data centers are "rigid loads": they can't go offline and must reserve capacity based on peak demand, so they're treated as fixed, requiring additional capacity charges for brief peak spikes—pure waste.

Wang Zhaoyang's assessment inverts that: data centers are 100% flexible. They can pair with energy storage and backup power sources; computing tasks during shortage periods can be delayed or even rerouted to other facilities.

But that "100%" needs an asterisk—and policy has already provided one. Point Seven of the Action Plan states: "Implement classified management of computing facilities according to task type." In plain terms: training tasks and offline batch processing can be delayed or shifted when electricity is tight, offering genuine high flexibility; but online inference is constrained by latency—no matter how cheap night electricity is, you can't send a conversational query 2,000 kilometers away to compute. Flexibility isn't a single setting; it's incremental: the more "real-time" the task, the less flexibility available.

An important caveat here—this is my observation, not his direct words. A cloud service provider naturally wants electricity pricing to reflect actual fluctuations and pay less for peak capacity, so "adjustable loads" carries inherent cost-saving motivation when it comes from his perspective. But setting aside the framing, the technical adjustability is real—the difficulty lies elsewhere.

"Flexibility" doesn't mean "adjust at will." Data centers are critical infrastructure; flexibility carries zero margin for risk. Backup power and energy storage carry genuine safety and operational costs; if something fails, the loss isn't an electricity bill—it's computing output and reputation. The real challenge isn't "can we adjust," but "how do we adjust with zero risk."

An analogy: ordinary factories can shift staffing elastically, but an operating room can't move to night shift just because electricity is cheaper at night. A data center is that non-stop operating room—flexibility is its capability, zero risk is its bottom line, and threading both together is the real skill.

Data centers aren't rigid loads; they're 100% flexible. But flexibility carries zero tolerance for risk—that's where the difficulty lies.

Flexibility is a key that unlocks the savings-on-electricity-bills lock, but it can't unlock the lock that solves the root problem.

**Third ledger: Costs—Electricity is only 10%, but the energy bill per unit computing power only grows heavier**

The third ledger is most prone to misleading people.

You often hear this claim: electricity costs represent only about 10% of total AI costs, so energy isn't a concern. Wang Zhaoyang broke down the categories: if you bundle IT hardware like servers and networking into the calculation, electricity is indeed diluted to roughly 10%; but if you exclude expensive chips and servers and count only infrastructure, electricity's share reaches roughly 60%. Both accounting methods are valid, but neither answers the real question: which direction is this share heading?

This is a calculation I worked out myself, not from the interview.

Over time, two curves diverge: computing power and chips follow technology's trajectory, and technology's characteristic is continuous cost reduction—the cost curve per unit of computing power trends downward; energy upgrade is inelastic, even rising, so the energy cost curve per unit of computing power trends upward. One down, one up—like scissors opening—this is what I call the "scissor gap": the energy cost share per unit of computing power will only climb over the long term.

I call this "energy sediment"—technology makes computing cheaper and cheaper, yet every inference, every training run burns electricity that weighs increasingly heavily on the ledger.

Of course, this calculation's premise (technology continuously reducing costs, energy inelastic and rising) only holds if neither is misread; if either breaks down, the scissor gap's slope changes—so it should be continuously validated, not treated as settled.

Some push back: isn't chip efficiency also a technology curve? PUE declines year-on-year, liquid cooling spreads, ASICs drive unit energy consumption downward—why should energy "sediment"? That's half right. Efficiency improvements compress "the electricity per unit computing," but AI's appetite grows faster—per public reports, token consumption and inference call volumes roughly doubled over the past two years. The denominator shrinks, but the numerator surges even faster. When I say "energy sediments," I don't mean "energy consumption can't decline"—I mean the decline speed can't keep pace with rising demand.

CAICT data offers supporting evidence (cited in Finance & Economics, January 2026): 2024 China data center electricity use was 166 billion kilowatt-hours, representing 1.68% of national total electricity consumption. At moderate growth rates, by 2030 that share reaches roughly 3%—about a doubling. But to be clear: the CAICT figure is "electricity share of national total"—supporting evidence from the total-volume side; my point about "cost-structure-side share climbing" is a different thing; direct evidence needs elsewhere—they're not the same column.

Doubling your computing power output while cutting your electricity bill per unit sounds good until you realize you're running twice the workload and burning twice the fuel.

After the cost ledger comes one more.

**Fourth ledger: Coordination—Storage capacity alone won't solve the root problem**

The final ledger wraps this up.

Many think: just add more storage capacity at data centers and we're done? Wang Zhaoyang shakes his head: storage doesn't solve the root problem. From a cloud provider's perspective, he naturally wants to pay less for peak capacity reserves—so "storage isn't the root solution" carries his interest perspective too, but the logic holds: storage is backup power, a planned-economy mindset, where utilities reserve capacity and you pay for peak; but computing demand is market-driven, hot one day, cold the next.

The real solution is finding dynamic equilibrium between generation, grid, and load sides, calculating globally optimal costs, and using AI for scheduling and operations. In short, it's not buying a few more batteries—it's organizing electricity and computing power into a self-adjusting network.

Policy is actually moving in that direction. Wang Zhaoyang traces this arc: before August 2025, policy focused on renewable energy integration; afterward, AI industry support and bidirectional empowerment began appearing. The logic translated into paper is the Action Plan for Promoting Bidirectional Empowerment of Artificial Intelligence and Energy, jointly issued April 8, 2026, by four departments—the National Development and Reform Commission, National Energy Administration, Ministry of Industry and Information Technology, and National Data Administration (National Energy Development Science Technology [2026] No. 34)—which institutionalizes coordination. This document isn't just slogans: it specifies establishing energy security systems by early 2027 and achieving world-leading clean energy supply capacity by 2030; operationally, it encourages computing facilities to configure network-type storage, undertake direct green electricity connections, participate in electricity and ancillary service markets in multiple forms, and even encourages computing facilities to apply for REITs in the infrastructure sector. Energy integration comes first, computing coordination second; integration is only the elementary stage.

The tension lies here: the power industry wants stable, planned supply; computing companies face market-fluctuating demand. One side demands stability, the other demands flexibility, and the process of finding "middle ground" is the industry's bottleneck.

Storage capacity alone won't solve the root problem. We need dynamic equilibrium and globally optimal costs across generation, grid, and load sides.

I'm watching three signals for this.

First: whether national data center electricity consumption as a share of national total follows the trajectory CAICT described—1.68% in 2024, reaching roughly 3% by 2030. If this line climbs, it only indicates "absolute energy consumption is accelerating"; to verify "the bill grows heavier," we need to wait for public sequences on cost-structure-side share—which is why I said earlier "direct evidence needs looking elsewhere."

Second: whether China's generation relative to US-Europe-Japan combined stays above 1.2x (roughly 1.24x in 2025, per the interview and Wind data). Once total capacity advantage slips, local contradictions sharpen.

Third: after that April 2026 Action Plan, whether integration of source-grid-load-storage and AI-plus-energy pilots actually accelerate into deployment, not stay buried in documents.

These three signals tell you the industry's wind direction better than the old question "Does China lack electricity?"

So returning to that opening consensus: AI's bottleneck isn't "insufficient electricity." China has abundant power; what's stuck is electricity and computing power not woven into a dynamically balanced network; more critical still is that the energy bill per unit of computing power only grows heavier. How well this network is woven becomes the next compass to watch. Watching that direction means looking at data, not narratives.