How much energy do data centers and artificial intelligence use?

Few — if any — technologies have been adopted as quickly as artificial intelligence (AI).

很少有（如果有的话）技术能像人工智能 (AI) 那样迅速被采用。

Since that computation runs on electricity, discussions around AI often return to what this means for energy demand.

由于计算需要电力来运行，因此围绕人工智能的讨论通常会回到这对能源需求意味着什么。

These concerns tend to take three forms.

这些担忧往往表现为三种形式。

One is environmental: growing energy demand for artificial intelligence will drive increases in carbon emissions, and make it harder to decarbonize.

一是环境方面：人工智能对能源需求的不断增长将推动碳排放量的增加，并使脱碳变得更加困难。

Another is the impact on local communities: high electricity demand could strain local supplies and push up energy prices.

另一个是对当地社区的影响：高电力需求可能会导致当地供应紧张并推高能源价格。

The third is that AI’s energy consumption could be the technology’s bottleneck; for those who want to see it expand, this could be a key limiting factor.

第三，人工智能的能源消耗可能成为技术瓶颈；对于那些希望看到它扩张的人来说，这可能是一个关键的限制因素。

We can look at this question on two levels.

我们可以从两个层面来看待这个问题。

The first is the total amount of electricity AI uses.

首先是人工智能的用电量总量。

The second is about individual impact, or how much electricity each query consumes.

第二个是关于个人影响，或者每个查询消耗多少电量。

Before digging into the data, it’s worth clarifying what is included in AI energy consumption.

在深入研究数据之前，有必要先澄清一下人工智能能耗包含哪些内容。

It’s the electricity consumed for both training and running the models (called “inference”).

这是训练和运行模型（称为“推理”）所消耗的电力。

Tech companies rarely publish data on how much energy is consumed when training their models, but based on the estimates we do have, it’s likely that energy demand is dominated by inference, not training.

科技公司很少公布有关训练模型时消耗多少能源的数据，但根据我们现有的估计，能源需求很可能由推理而非训练主导。

1

1

The estimates we’ll look at include the electricity used specifically for the servers, plus additional energy used for cooling, lighting, and other things needed to keep the data centers running.

我们将考虑的估算包括专门用于服务器的电力，以及用于冷却、照明和保持数据中心运行所需的其他设施的额外能源。

They don’t include the energy used to power the device — a laptop, desktop, or phone — that someone is using to access AI.

它们不包括用于为访问人工智能的设备（笔记本电脑、台式机或手机）供电的能源。

Importantly — and this matters when comparing to other sources — they do not include demand from cryptocurrency mining.

重要的是——与其他来源相比这一点很重要——它们不包括加密货币挖矿的需求。

Let’s start with the big picture.

让我们从大局开始。

How much electricity did data centers consume last year?

去年数据中心消耗了多少电力？

According to the International Energy Agency (IEA), around 485 terawatt-hours (TWh).

根据国际能源署 (IEA) 的数据，约为 485 太瓦时 (TWh)。

2 That’s equivalent to the annual electricity generation of Germany.

2 这相当于德国的年发电量。

And for context, that’s around 1.5% of the world’s electricity generation .

就背景而言，这大约占世界发电量的 1.5%。

3

3

Now, data centers are more than just AI: they’re facilities that contain the servers and IT infrastructure behind all of our digital services.

现在，数据中心不仅仅是人工智能：它们是包含我们所有数字服务背后的服务器和 IT 基础设施的设施。

That’s everything from email and Internet browsing to Netflix streaming, Google Maps, online banking, and messaging friends.

从电子邮件和互联网浏览到 Netflix 流媒体、谷歌地图、网上银行和给朋友发短信，一切应有尽有。

AI data centers are dedicated facilities for running AI models.

人工智能数据中心是运行人工智能模型的专用设施。

The distinction between the two is not always clean-cut, but the chart below shows an estimated breakdown.

两者之间的区别并不总是清晰的，但下图显示了估计的细分。

Non-AI data centers consumed two-thirds of the total, and AI-focused ones, the remaining third.

非人工智能数据中心消耗了总量的三分之二，专注于人工智能的数据中心消耗了剩下的三分之一。

Based on this data from the International Energy Agency, I estimate that AI consumed around 0.5% of the world’s electricity in 2025.

根据国际能源署的数据，我估计 2025 年人工智能将消耗全球约 0.5% 的电力。

4

4

Electricity currently accounts for around one-fifth of the world’s primary energy.

目前，电力约占世界一次能源的五分之一。

5 Hence, AI likely consumed around 0.1% of total primary energy in 2025.

5 因此，到 2025 年，人工智能可能消耗一次能源总量的 0.1% 左右。

Given how quickly demand for AI has been increasing, it might be surprising that AI-focused data centers consume less, on aggregate, than non-AI ones.

考虑到对人工智能的需求增长得如此之快，以人工智能为中心的数据中心的总体消耗量比非人工智能数据中心的消耗量要少，这可能会令人惊讶。

Projections expect this gap to close quickly.

预计这一差距将很快缩小。

In the chart, I’ve also included the IEA’s base-case projection of data center demand in 2030.

在图表中，我还纳入了 IEA 对 2030 年数据中心需求的基本情况预测。

These projections are highly uncertain, and some have argued that the IEA is among the most conservative in its assessment of AI demand growth.

这些预测具有高度不确定性，一些人认为 IEA 对人工智能需求增长的评估是最保守的。

Even so, most of the growth in data center demand will come from AI-focused facilities.

即便如此，数据中心需求的大部分增长将来自以人工智能为中心的设施。

In this scenario, data centers grow to 3% of global electricity in 2030, and AI centers then use about the same amount as non-AI ones.

在这种情况下，到 2030 年，数据中心将占全球电力的 3%，而人工智能中心的用电量将与非人工智能中心大致相同。

Above, we looked at estimates of global electricity demand for data centers from the International Energy Agency, one of the most widely used sources on energy data.

上面，我们研究了国际能源署对全球数据中心电力需求的估计，这是使用最广泛的能源数据来源之一。

However, other sources also publish estimates of data center demand.

然而，其他来源也发布了数据中心需求的估计。

Another of these comes from S&P Global, which is used and reported by the Energy Institute.

其中另一个来自 S&P Global，由能源研究所使用和报告。

Its latest Statistical Review of World Energy included estimates of data center demand for the first time.

其最新的《世界能源统计回顾》首次包含了对数据中心需求的估计。

In the chart below, I’ve shown both sources from 2020 to 2025.

在下图中，我显示了 2020 年至 2025 年的两个来源。

This means we’re comparing historical estimates only, not projections, where differences between sources are far larger.

这意味着我们仅比较历史估计，而不是预测，因为来源之间的差异要大得多。

As you can see, there are large differences between them.

正如您所看到的，它们之间存在很大差异。

S&P’s figures for 2025 are around 60% higher, but have been consistently higher for the entire time series.

标准普尔 2025 年的数据大约高出 60%，但在整个时间序列中一直处于较高水平。

Neither organization provides detailed and transparent breakdowns of their methodology, so it’s hard to give a definitive, complete explanation for the differences.

这两个组织都没有提供其方法的详细和透明的细分，因此很难对差异给出明确、完整的解释。

But one major one is the inclusion of electricity for cryptocurrency mining.

但其中一个主要问题是为加密货币挖矿提供电力。

This is included in the EI/S&P data, but is not in the IEA data.

这包含在 EI/S&P 数据中，但不包含在 IEA 数据中。

Electricity consumption for crypto mining is likely to be in the range of 150 to 200 terawatt-hours (TWh), which would account for almost two-thirds of the difference.

加密货币挖矿的电力消耗可能在 150 至 200 太瓦时 (TWh) 之间，这几乎占了差异的三分之二。

Even taking this into account, there would still be a gap.

即使考虑到这一点，差距仍然存在。

This is likely explained by the differences in methodological approach.

这可能是由于方法论方法的差异造成的。

S&P models power demand based on bottom-up data on installed data center capacity.

标准普尔根据已安装数据中心容量的自下而上数据对电力需求进行建模。

That relies on assumptions about how often this capacity is used, so the results are sensitive to them.

这依赖于对该容量的使用频率的假设，因此结果对它们很敏感。

The IEA attempts to estimate the electricity consumption of IT equipment, cooling, and infrastructure demand in data centers more directly.

IEA 试图更直接地估算数据中心 IT 设备、冷却和基础设施需求的电力消耗。

But again, this comes with significant uncertainty.

但同样，这也带来了很大的不确定性。

These estimates are not directly comparable, and measure different things, but it’s useful to keep this comparison in mind if you’re faced with quite different numbers for global power demand from data centers.

这些估计值不可直接比较，并且衡量的是不同的事物，但如果您面临数据中心的全球电力需求的截然不同的数字，则记住这种比较很有用。

1.5% of the world’s electricity might not seem like much.

世界电力的 1.5% 看起来似乎并不算多。

But in some places, that share is far higher.

但在某些地方，这一比例要高得多。

This is really the key challenge with growing data center demand: it’s geographically concentrated, meaning the world’s demand is served by a small number of electricity grids.

这确实是数据中心需求不断增长的关键挑战：它在地理上集中，这意味着世界的需求是由少数电网满足的。

In the chart, you can see the share of electricity used for data centers in different regions.

在图表中，您可以看到不同地区数据中心的用电量比例。

5% of electricity in the United States is used to power data centers.

美国 5% 的电力用于为数据中心供电。

For AI-focused ones specifically, it’s probably around 2%.

特别是对于专注于人工智能的企业来说，这个比例可能在 2% 左右。

But in fact, this demand is even more locally concentrated.

但事实上，这种需求更加集中于本地。

Beneath Europe’s 1.6% figure, we have Ireland, where data centers account for more than 20% of electricity consumption.

在欧洲 1.6% 的数字之下，爱尔兰的数据中心占电力消耗的 20% 以上。

Beneath the 5% US figure, there are states where data centers make up more than 10% of demand.

在美国 5% 的数字之下，有些州的数据中心占需求的比例超过 10%。

6 In states such as Virginia, it’s more than one-quarter.

6 在弗吉尼亚州等州，这一比例超过四分之一。

This, combined with the rapid pace of AI growth, could put pressure on local grids, even if total electricity demand is not overwhelming for the world as a whole.

这与人工智能的快速发展相结合，可能会给当地电网带来压力，即使总电力需求对整个世界来说并不是压倒性的。

I know many people who are conscious of their own use of large language models (LLMs) and AI tools for environmental reasons.

我认识很多人，他们意识到自己出于环境原因而使用大型语言模型 (LLM) 和人工智能工具。

Some abstain or use them as little as possible.

有些人放弃或尽可能少地使用它们。

Others continue to do so, but feel guilty for it.

其他人继续这样做，但为此感到内疚。

How much energy do our individual queries consume?

我们的各个查询消耗了多少能量？

If someone asks an LLM — like ChatGPT, Gemini, or Claude — a question, how much additional electricity demand do they generate?

如果有人问法学硕士（例如 ChatGPT、Gemini 或 Claude）一个问题，他们会产生多少额外的电力需求？

Again, accurate and up-to-date figures on this are hard to find because most technology companies have not released detailed analyses of the energy consumption of their models.

同样，很难找到这方面的准确和最新数据，因为大多数科技公司尚未发布对其模型能耗的详细分析。

One of the first companies to do so was Google; in 2025, it released energy estimates for its Gemini model.

谷歌是最早这样做的公司之一。 2025年，它发布了Gemini模型的能源估算。

It estimated the median text-based query (basically asking Gemini a text question) consumed around 0.24 watt-hours (Wh) of electricity.

据估计，基于文本的查询（基本上是向 Gemini 询问文本问题）消耗的电量约为 0.24 瓦时 (Wh)。

For context, that’s the amount of energy a microwave would consume in less than one second, or a television for ten seconds.

就上下文而言，这是微波炉在不到一秒内消耗的能量，或者是电视在十秒内消耗的能量。

The CEO of OpenAI, Sam Altman, also previously wrote that an “average query” on ChatGPT consumed around 0.34 Wh (but without a detailed breakdown of where this number comes from).

OpenAI 的首席执行官 Sam Altman 此前也曾写道，ChatGPT 上的一次“平均查询”消耗了大约 0.34 Wh（但没有详细说明这个数字的来源）。

Epoch AI also provided its own independent estimate of around 0.3 Wh for a “typical” ChatGPT query.

Epoch AI 还为“典型”ChatGPT 查询提供了自己的独立估计，约为 0.3 Wh。

So a number of sources tend to converge on a similar figure.

因此，许多来源往往会集中在一个相似的数字上。

However, most queries are simple, so the median query is probably small in both size and complexity.

然而，大多数查询都很简单，因此中值查询的大小和复杂性可能都很小。

For longer queries or requests that rely on AI agents or reasoning, the footprint is higher.

对于依赖人工智能代理或推理的较长查询或请求，占用空间更高。

Epoch AI estimated that a long query (7,500 words of input) could consume about 2.5 Wh, and a very long one (75,000 words) could consume 40 Wh.

Epoch AI 估计，一个长查询（输入 7,500 个单词）可能消耗约 2.5 Wh，而一个很长的查询（75,000 个单词）可能消耗 40 Wh。

Those are far higher than the simple text query.

这些远远高于简单的文本查询。

There are fewer estimates for other types of requests.

对其他类型请求的估计较少。

In its 2026 report, the IEA provides estimates of GPU electricity consumption for agentic tasks with reasoning.

IEA 在其 2026 年报告中通过推理提供了代理任务的 GPU 电力消耗估计值。

7 A standard request to an AI agent — such as Claude — is estimated to consume around 1.1 Wh.

7 对 AI 代理（例如 Claude）的标准请求预计消耗约 1.1 Wh 的电量。

An agentic request with reasoning, around 50 Wh.

带有推理的代理请求，大约 50 Wh。

This is similar to the estimates for maximum-length text queries.

这类似于最大长度文本查询的估计。

However, all of these are still relatively small compared to the average person’s daily electricity consumption, especially in high-income countries.

然而，与普通人的日常用电量相比，所有这些仍然相对较小，特别是在高收入国家。

In the European Union, average electricity consumption per person is around 17,000 Wh per day.

在欧盟，人均每天用电量约为 17,000 瓦时。

8 That’s equivalent to around 6,800 long-input queries (which consume 2.5 Wh each).

8 这相当于大约 6,800 个长输入查询（每个查询消耗 2.5 瓦时）。

Or 425 of the maximum-input queries.

或者最大输入查询的 425 个。

The average person in the US consumes approximately twice as much electricity as the EU average, so the contribution of AI queries to someone’s footprint there is about half the size.

美国人均消耗的电力大约是欧盟平均水平的两倍，因此人工智能查询对某人足迹的贡献约为欧盟平均水平的一半。

9

9

In the chart below, I’ve provided some comparisons to give a sense of how this energy consumption compares to other products or activities.

在下图中，我提供了一些比较，以了解这种能源消耗与其他产品或活动的比较。

10

10

We send two regular newsletters so you can stay up to date on our work and receive curated highlights from across Our World in Data.

我们定期发送两份时事通讯，以便您及时了解我们的工作并接收来自“我们的数据世界”的精选亮点。

In this article, I’ve focused on historical estimates of data center and AI demand.

在本文中，我重点关注数据中心和人工智能需求的历史估计。

Getting reasonable recent estimates of how much electricity these are already consuming was already difficult.

最近对这些设备已经消耗了多少电量进行合理的估计已经很困难了。

Predicting how this will change in the future is even more difficult.

预测未来情况将如何变化更加困难。

There is a wide range of estimates for future demand, and the differences between them tend to grow the further into the future you go.

对未来需求的估计有很多种，而且随着未来的发展，它们之间的差异往往会越来越大。

This divergence is particularly clear after 2030, since much of the medium-term demand pre-2030 relies on infrastructure being built today.

这种差异在 2030 年之后尤其明显，因为 2030 年之前的中期需求大部分依赖于目前正在建设的基础设施。

Future demand will depend on a range of factors.

未来的需求将取决于一系列因素。

Not just how user demand grows, but also how the efficiency of chips and other hardware grows, too.

不仅用户需求如何增长，芯片和其他硬件的效率也如何增长。

The IEA, for example, has four future scenarios : one projects far higher user demand growth, and another sees much faster efficiency gains.

例如，国际能源署有四种未来情景：一种预计用户需求增长要高得多，另一种则预计效率提升要快得多。

While we don’t know how much AI energy demand will grow in the future, the key points we learn from the data today are likely to hold true.

虽然我们不知道未来人工智能能源需求将增长多少，但我们今天从数据中了解到的要点可能是正确的。

For an individual worried about their own use of AI, the energy use of individual chatbot queries is small.

对于担心自己使用人工智能的个人来说，单个聊天机器人查询的能源消耗很小。

For the world as a whole, data centers make up a relatively small share of total electricity consumption, and that won’t change in the coming years.

对于整个世界来说，数据中心在总电力消耗中所占的份额相对较小，并且这一点在未来几年不会改变。

The issue is that supply is so geographically concentrated: whether grids can meet this surge in demand while keeping emissions and local prices under control is the real test.

问题在于供应在地理上如此集中：电网能否满足需求的激增，同时控制排放和当地价格才是真正的考验。

A final point that’s worth keeping in mind is that the figures we’ve looked at measure electricity consumption, not carbon emissions.

值得记住的最后一点是，我们研究的数据衡量的是电力消耗，而不是碳排放。

For those worried about the climate impacts of data centers and AI, how that electricity is generated arguably matters more than how much it uses.

对于那些担心数据中心和人工智能对气候影响的人来说，电力的产生方式可以说比用电量更重要。

A query served by a data center on a coal-dominated grid will have a much larger impact than one running on a renewables- or nuclear-heavy one.

由以煤炭为主的电网上的数据中心提供的查询将比在可再生能源或核能为主的电网上运行的查询产生更大的影响。

How the growth of AI affects emissions depends not just on its energy efficiency and demand, but also on how clean the grids are that supply it.

人工智能的增长如何影响排放不仅取决于其能源效率和需求，还取决于为其供电的电网的清洁程度。

Many thanks to Max Roser, Esteban Ortiz-Ospina, and Edouard Mathieu for comments and feedback on this article.

非常感谢 Max Roser、Esteban Ortiz-Ospina 和 Edouard Mathieu 对本文的评论和反馈。

AI already has a large impact on our world.

人工智能已经对我们的世界产生了巨大的影响。

Explore research and data to understand the trajectory of this important technology.

探索研究和数据以了解这项重要技术的发展轨迹。

Explore data on where our electricity comes from and how this is changing.

探索有关我们的电力来源及其变化的数据。

Fossil fuels are the dirtiest and most dangerous energy sources, while nuclear and modern renewable energy sources are vastly safer and cleaner.

化石燃料是最肮脏和最危险的能源，而核能和现代可再生能源则更加安全和清洁。

Epoch AI estimates that training Grok 4 consumed around 0.31 terawatt-hours (TWh) of electricity.

Epoch AI 估计训练 Grok 4 消耗了约 0.31 太瓦时 (TWh) 的电力。

As we’ll see later, total demand for AI in 2025 was around 155 TWh.

正如我们稍后将看到的，2025 年人工智能的总需求约为 155 TWh。

So, training Grok 4 — a fairly large model — was around 0.2% of the total.

因此，训练 Grok 4（一个相当大的模型）约占总数的 0.2%。

Ember estimates that in 2025, the world generated around 31,800 TWh of electricity.

Ember 估计，到 2025 年，全球发电量约为 31,800 太瓦时。

The IEA estimates that AI-focused data centers consumed 155 TWh in 2025.

IEA 估计，到 2025 年，以人工智能为中心的数据中心将消耗 155 太瓦时的电量。

That’s 0.49% of global electricity.

相当于全球电力的 0.49%。

Ember estimates that in 2025, the world generated around 31,800 TWh of electricity.

Ember 估计，到 2025 年，全球发电量约为 31,800 太瓦时。

Primary energy use in 2025 was 167,000 TWh .

2025 年一次能源使用量为 167,000 太瓦时。

Electricity generation was 31,800 TWh .

发电量为 31,800 太瓦时。

[31,800 / 167,000 * 100 = 19%].

[31,800 / 167,000 * 100 = 19%]。

As the IEA puts it: “nearly half of data center capacity in the United States is in five regional clusters.”

正如 IEA 所说：“美国近一半的数据中心容量分布在五个区域集群中。”

This data comes from Ember Energy.

该数据来自Ember Energy。

The average per person in the EU is around 6,200 kWh per year , which is 17 kWh per day.

欧盟人均每年用电约为 6,200 千瓦时，即每天 17 千瓦时。

The estimates for AI queries come from Epoch AI .

AI 查询的估计来自 Epoch AI。

The estimates for other activities are described at: https://hannahritchie.github.io/energy-use-comparisons

其他活动的估算如下：https://hannahritchie.github.io/energy-use-comparisons

Our articles and data visualizations rely on work from many different people and organizations.

我们的文章和数据可视化依赖于许多不同的人和组织的工作。

When citing this article, please also cite the underlying data sources.

引用本文时，请同时注明底层数据来源。

This article can be cited as:

这篇文章可以引用为：

We are a live site and we keep our work maintained.

我们是一个活跃的网站，我们会保持我们的工作。

So our articles are often changed, including the data.

所以我们的文章经常改，包括数据。

You can browse previous versions of this work here.

您可以在此处浏览此作品的先前版本。