This summer, the maximum electricity load of the State Grid will exceed 1.3 billion kilowatts, increasing by about 6% year-on-year. 今夏国家电网最大用电负荷将超13亿千瓦，同比增约6%

These two seemingly unrelated news stories converged today, sketching the raw reality beneath the AI frenzy: On one side stands TSMC Chairman C.C. Wei’s ambitious vision at the shareholder meeting—from generative AI to agentic AI, surging token consumption, and semiconductor demand "fundamentally underpinned," projecting over 30% annual revenue growth in a blaze of expansion. On the other side is the cool, calculated assessment from China’s State Grid: peak power load this summer will exceed 1.3 billion kilowatts, up 6% year-on-year, with 168 critical projects racing to be completed before the summer peak. One cheers the limitless demand for computing power, while the other tallies the finite supply of electricity.

These are no coincidence. Of TSMC’s projected 30% growth, how much is driven by AI chips? These chips, packed into server clusters, each function as a power-devouring monster. Training a single large model may consume electricity equivalent to the annual usage of a small city. As the entire industry toasts the future of agentic AI, the power grid is already sounding alarms about looming physical limits. Wei speaks of business prospects and customer confidence, while the State Grid discusses substation capacities and transmission line loads. One paints a cosmic vision from the cloud, while the other calculates the landing point of every kilowatt on the ground. This gap is the most fantastical—and dangerous—part of the industry.

Trending news items further illustrate this tableau: Intel plans to end NVIDIA’s monopoly? That means more, more powerful AI chips flooding the market. 1.6 billion Windows users storming into the Agent era? Every Agent invoking models in the background continuously drains computing power and electricity. Volcengine raising its MaaS revenue target to 15 billion, with a single product earning a billion monthly—behind these numbers lie countless data centers burning bright through API calls. Tencent, Alibaba, ByteDance battling over Skill stores—the richer the ecosystem, the more vibrant the application layer, the more exponentially the underlying computing resources are demanded. We all seem to suffer from selective blindness: fixated on chips advancing another nanometer or model parameters swelling by trillions, yet pretending not to see the energy bills powering it all.

This dissonance is unsettling. TSMC’s confidence rests on "technological differentiation and a broad customer base"—undeniably a hard strength. But even the toughest chips cannot defy the laws of physics. Power supply is not software; it cannot iterate and scale at algorithmic speed. It depends on power plants, grids, and storage infrastructure, which require lengthy construction cycles and massive investment. As AI evolves in days while power infrastructure improves over years or even decades, conflicts will inevitably intensify. A peak load of 1.3 billion kilowatts exceeds the total power generation of many nations—a number that is itself a heavy shackle.

Ironically, while we advocate AI to enable "intelligence for everything," we gloss over the most fundamental issue of energy sustainability. Amid the arms race for computing power, there is little genuine scrutiny of energy efficiency. The industry competes over whose model is larger and faster, but rarely over whose computing power per watt is lower—as if electricity were free and limitless. The State Grid’s “accelerated construction” may quench immediate thirst, but unless we move away from the crude, computing-power-driven growth model, bottlenecks will only arrive faster and harder. The 30% growth Wei envisions might well be an adrenaline shot borrowed from the future.

Ultimately, this feast may be tripped up by the oldest problem of all: conservation of energy. Information flow demands energy. Every stunning moment created by AI corresponds to a combustion in a power plant or the natural flow of wind and water. As TSMC’s production capacity races against State Grid’s load limits, we might need less fantasy about infinite growth and more respect for the boundaries of reality. Otherwise, when agentic AI can autonomously tap into trillions of resources, the first crisis we face may not be ethical dilemmas—but the embarrassment of rolling blackouts. No matter how powerful the chips, without stable power, they are nothing but elegant shards of silicon.

两条看似无关的新闻，今天撞在了一起，勾勒出AI狂潮下最真实的底色：一边是台积电董事长魏哲家在股东会上描绘的宏伟蓝图——从生成式AI到代理式AI，token消耗暴增，半导体需求“具备根本性支撑”，全年营收增长将超过30%，一派烈火烹油；另一边，是国家电网发布的冷静测算：今夏最大用电负荷将突破13亿千瓦，同比增6%，168项迎峰度夏重点工程正在抢工期。一个在欢呼算力的无限需求，一个在计算电力的有限供给。

这两件事根本不是巧合。台积电增长的30%里，有多少是AI芯片贡献的？这些芯片装进服务器集群，每一块都是一台“吞电巨兽”。训练一个大型模型需要的电力，可能相当于一个小城市的年用电量。当整个产业都在为代理式AI的未来举杯庆贺时，电网已经在为即将到来的物理极限拉响警报。魏哲家谈的是商业前景和客户信心，而国家电网谈的是变电站的容量和输电线路的负荷。一个在云端描绘星辰大海，一个在地上计算每一千瓦的落点。这中间的落差，才是这个行业最魔幻也最危险的部分。

热榜里那些新闻，更是这幅图景的注脚。英特尔要终结英伟达的垄断？那意味着更多、更强大的AI芯片将涌入市场。16亿Windows用户冲进Agent时代？每一个Agent在后台调用模型，都在持续消耗算力和电力。火山引擎把MaaS营收目标提到150亿，单款产品月入10亿——这数字背后是无数API调用背后的数据中心灯火通明。腾讯、阿里、字节混战Skill商店，生态越繁荣，应用层越活跃，对底层算力的调用就越是指数级增长。我们好像集体患上了选择性失明：只盯着芯片制程又前进了一纳米，模型参数又膨胀了几万亿，却假装看不见支持这一切运转的电力账单。

这种割裂感令人不适。台积电的信心建立在“技术差异化与广泛客户群”上，这当然是硬实力。但再硬的芯片，也硬不过物理定律。电力供应不是软件，不能像算法一样快速迭代扩容。它依赖于发电厂、电网、储能设施，这些需要漫长的建设周期和天量的投资。当AI的发展速度以天计，而电力基础设施的完善以年甚至以十年计时，矛盾必然会激化。13亿千瓦的峰值负荷，比许多国家的总发电量还大，这个数字本身就是一副沉甸甸的镣铐。

更讽刺的是，我们一边鼓吹AI要赋能“万物智能”，一边却在最基本的能源可持续性问题上语焉不详。算力军备竞赛的喧嚣中，很少听到对能源效率的真正追问。行业都在比谁的模型更大、跑得更快，但很少有人比谁的单位算力能耗更低。仿佛电力是免费且无限的。国家电网的“加快建设”能解一时之渴，但若不改变当前以堆算力为主导的粗放式发展路径，瓶颈只会来得更快、更猛烈。魏哲家看到的30%增长，何尝不是一剂预支未来的兴奋剂？

最终，这场盛宴可能被最古老的问题绊倒：能量守恒。信息的流动需要能量的支撑。AI创造的每一个惊艳瞬间，背后都对应着发电厂里的一次燃烧，或自然里的一次风起水流。当台积电的产能和国家电网的负荷赛跑时，我们或许该少一些对无限增长的幻想，多一些对现实边界的敬畏。否则，当代理式AI真的可以自主调用万亿资源时，我们首先要面对的，可能不是伦理危机，而是拉闸限电的尴尬。芯片再强，没有稳定的电力，也不过是一堆精致的硅片。