DeepSeek对数据中心和能源的潜在影响
Potential datacenter and energy impacts of DeepSeek
Analysts - Kelly Morgan, Dan Thompson, Johan Vermij
Publication date: Wednesday, February 12 2025
AI的发展需求，推动了数据中心基础设施的大规模建设需求，全球从业者也一直来认为能源是未来限制数据中心发展的限制因素，但近期DeepSeek的出现，人们开始质疑以往的认知，大模型的资本泡沫似乎要被捅破。近期国内外全网不断发出质疑声，对未来能源需求及建设规模的预期，有转变的趋势。本文简要分析了DeepSeek对数据中心的建设规模规模及公用设施能源公司的影响。
DeepSeek是一家中国AI实验室，自2023年年中以来一直在开发代码生成和大语言模型。该公司的大模型V3于2024年12月发布，但1月份推出的推理模型（R1）引起了开发人员和投资者的注意。该模型是开源的，在类似模型中表现良好，但据报道，它只使用了2,048个NVIDIA H800 GPU芯片进行训练（与OpenAI的GPT4相比，GPT4在2022年使用了大约20,000个A100 GPU进行训练），因此比其他基础模型的成本要低得多。使用更少、更低功率的GPU芯片意味着可能需要更少的能源，并且质疑是否还需要按照之前模型的思路规划许多数据中心，这可能也会减少对电力公司能源的需求。
DeepSeek is a Chinese AI lab that has been developing code-generation and large language models since mid-2023. The firm's large language model V3 was released in December 2024, but a reasoning model launched in January (R1) caught the attention of developers and investors alike. The model is open source and performs well against similar models, but was reportedly trained using only 2,048 NVIDIA H800 GPU chips (compared with OpenAI's GPT4, which was trained in 2022 using an estimated 20,000 A100 GPUs), and therefore at much lower cost than other foundation models. Using fewer, lower-powered GPU chips means less energy may be required, and calls into question whether many of the datacenters planned with the previous models in mind will be needed, potentially reducing the demand for energy from utilities as well.
要点
The Take
自ChatGPT推出以来，长期所需的AI基础设施规模一直是未知数。最大的IT和AI公司，尤其是在美国，似乎更倾向于过度建设，而不想陷入产能不足的困境，这让投资者对这些AI投资的回报感到不安。在一定程度上，这解释了投资者对DeepSeek模型的反应，仿佛泡沫正在破裂。这可能是真的，关于AI在社会中的应用，以及它所需的资源，仍然是重大的未知因素。然而，一项新技术的发展总会有高潮和低谷。许多推动数据中心需求增长的趋势仍然存在，尽管总是存在过度建设的可能性，但数据中心的实际建设速度比以前的过度建设时期要快得多。现在，数据中心行业更加灵活，能够相对快速地应对过剩产能并停止建设，尽管顶级建设者可能会损失一些钱。然而，对于大型发电厂来说，情况并非如此，因为它们的发展周期要长得多。
Since the launch of ChatGPT, the amount of infrastructure required for AI over the long term has been anyone's guess. The largest IT and AI firms, particularly in the US, have seemed to lean toward overbuilding rather than being caught without capacity, making investors nervous about the returns on these AI investments. This accounts (in part) for the investor response to the DeepSeek model as if a bubble were bursting. That may be true, and there are still major unknowns about AI's use in society, as well as the resources it will require. However, there are always highs and lows as a new technology takes off. Many of the trends behind the growing demand for datacenters are still in place, and although there is always the potential for overbuilding, the actual construction of facilities is much faster than during previous eras of overbuilding. The industry is more flexible now, and can respond to overcapacity fairly quickly and stop building, although top builders may lose some money. The same cannot be said for large-scale power plants, though, with their much longer time horizons.
Added datacenter capacity — all for AI?
自2022年11月推出ChatGPT以来，规划的数据中心建设数量激增，建设者们向美国各地的电力公司请求电力供应，达到了前所未有的水平。这种情况也开始在美国以外发生。DeepSeek模型的推出是否意味着数据中心将闲置？这可能取决于预计增加的数据中心有多少是为GPU准备的。在我们的《GPU对数据中心市场的监控与预测》报告中，我们估计，从2025年到2029年，全球数据中心计划每年增加15-18GW，其中30%-40%的容量预计将用于容纳用于AI工作负载的GPU芯片（各种类型）。剩余的百分比估计用于其他类型的设备——用于AI的CPU芯片、用于非AI工作负载（例如典型的云部署）的存储、网络和IT设备。并非所有这些GPU芯片都被期望用于开发大型语言模型。AI提供商（和分析师）一直期待着一种转变——从训练非常大的语言模型到训练和维护其中的一些（但越来越多的小型专业模型）小型专业模型，而推理将占AI工作负载的更大比例。如果DeepSeek的开源模型需要更少的高性能芯片和潜在的更少的培训资金，它可能会加速向更小、更分散的模型转变。
Since the launch of ChatGPT in November 2022, planned datacenter builds have soared, with builders requesting power feeds from utilities throughout the US at a level not seen before. This wasstarting to happen outside the US as well. Does the launch of DeepSeek's model mean that datacenters will sit around empty? That could depend on what amount of the projected datacenter additions was expected to be for GPUs. In our GPU Impact on Datacenters Market Monitor & Forecast, we estimate that datacenters plan to add 15-18 GW per year globally from 2025 to 2029, with 30%-40% of that capacity expected to be built to house GPU chips (of various types) for AI workloads. The remaining percentage is estimated to be for other types of equipment — CPU chips used for AI, storage, networking and IT equipment used for non-AI workloads (e.g. typical cloud deployments). Not all of these GPU chips were expected to be used for developing large language models. AI providers (and analysts) have been expecting a shift — from training very large language models to training and maintaining some of those (but a growing number of smaller specialized models), while inferencing becomes a larger percentage of AI workloads. If DeepSeek's open-source model requires fewer high-powered chips and potentially less capital to train, it may accelerate the move toward smaller, more distributed models.
分布式的、更小的AI模型可能意味着，在大规模集中式位置，需要的数据中心容量更少，而且目前顶级的AI公司可能不需要像他们想象的那么多容量。AI和IT公司巨头似乎已经在尝试避免因容量不足而失去机会（并因此输给竞争对手）的“不惜一切代价以求增长”的方法。其中一个潜在的容量限制因素看起来将是电力的可用性，这可能会带来电力公司过度建设的连锁反应
Distributed, smaller AI models may mean that less datacenter capacity will be needed in very large-scale centralized locations, and that perhaps the current top AI companies will not need as much capacity as they think. The top AI and IT firms already seem to be trying to err on the side of overbuilding rather than being caught without capacity (and losing out to competitors as a result) in a "spend what it takes in order to grow" approach. One of the potential constraints on capacity looked like it would be the availability of electricity, with the potential knock-on effect of power companies overbuilding as well.
Are companies overbuilding?
在讨论科技巨头是否过度建设（或已经过度建设）时，批评的中心观点通常涉及对未来效率提高的担忧，无论是在硬件还是软件方面，这可能会减少对如此多基础设施的需求。DeepSeek就是一个例子，由于该公司将其作为开源提供，这导致了对AI计划的另一个常见批评，即长期财务可行性。
A central point of critique when debating whether the tech giants are overbuilding (or already have) generally involves concern over future efficiency gains, whether in hardware and/or software, which could reduce the need for so much infrastructure. DeepSeek could be an example of that, and since the company is offering it as open source, this leads to another common critique of AI initiatives, that of long-term financial viability.
自从计算机出现以来，一直存在技术瓶颈，无论是处理能力、内存速度/容量、存储速度/容量等，以及软件适当利用硬件进步的能力。当一个瓶颈被移除后，就会出现更多的瓶颈。对于通用计算来说，世界各地的系统都没有得到充分利用，这可能是它们被过度构建的一个迹象。公司继续扩建新的基础设施，这些设施最终可能也会被未充分利用或未完全利用。然而，当前的系统利用率并不能很好地反映未来的需求。
Since the advent of the computer, there have existed technological bottlenecks, whether processing power, memory speed/capacity, storage speed/capacity, etc., in addition to software's ability to properly harness hardware advancements. As a bottleneck was removed, advancements were made, and more bottlenecks would appear. For general-purpose computing, it is true that systems all over the world sit underutilized, which one could argue is a sign they have been overbuilt. Companies continue to build out new infrastructure, which may also be ultimately underutilized or not fully utilized. However, current system utilization is not a good indicator of future demand.
在基于GPU的计算领域，我们仍处于曲线的起点。公司继续寻找AI的新用途，而训练模型以获得结果的过程仍然需要长时间才能完成——在某些情况下是几周，甚至几个月。似乎有无穷无尽的作业队列等待运行。的确，优化仍然有很多机会，但目前并不缺乏新的想法和应用场景。在这个阶段，效率的提升（比如DeepSeek的提升）似乎更有可能成为跨越式发展的催化剂，而不是游戏结束时的蜂鸣器。因此，虽然DeepSeek声称已经创造了一种用更少的资源做更多事情的方法，但美国IT公司也将在这些进步的基础上继续竞争，但拥有更强大的计算能力。
In the world of GPU-based computing, we are still at the beginning of that curve. Companies continue to find new uses for AI, while the training of models to obtain results continues to take long periods of time to complete — weeks, or months in some cases. There seem to be endless queues of jobs to be run. It is true that there is still plenty of opportunity for optimization, but there is currently no shortage of new ideas and use cases. At this stage, it seems far more likely that efficiency gains, such as those from DeepSeek, will simply be a catalyst for leaps forward, rather than the buzzer at the end of the game. So, while DeepSeek claims to have created a way to do a lot with less, US IT firms will continue the race as well, building on those advancements, but with remarkably more compute power.
此外，大型公有云提供商也有一些后备计划。如果数百或数千个较小的模型正在运行，并且推理正在增长，那么共享基础设施的公共云模型可能是对运行AI的小型公司有吸引力的选择。如果AI用例不需要以非常低的延迟进行推理，那么该基础设施仍可能位于集中式数据中心。在规模更大、互联程度更高的数据中心安装设备仍然有好处，这些数据中心具有规模经济效益，能够获得绿色能源和其他可持续性效益。
In addition, the large public-cloud providers have a bit of a back-up plan. If hundreds or thousands of smaller models are operating and inferencing is growing, the public cloud model of shared infrastructure could be an appealing option for smaller companies running AI. And if the AI use case does not require inferencing to happen with very low latency, that infrastructure could still be in centralized datacenters. There will still be benefits to having equipment in larger, moreinterconnected facilities that have economies of scale and the ability to obtain green energy plus other sustainability benefits.
因此，对大型数据中心的需求仍然很大，它们可能需要更长的时间才能发展到最初计划的园区规模，并且可能会同时出现更小、更分散的数据中心。这在很大程度上取决于AI的用例，以及AI推理需要多少近乎即时的响应，因此，在城市中心分配较小的容量，而不是集中在高效的绿色数据中心。
So, there could still be plenty of demand for large-scale datacenters, they may just take longer to grow into the campus size originally planned, and there could be concurrent growth of smaller, more distributed facilities. A lot will depend on the use cases for AI and how much AI inferencing will require near-instantaneous response, and therefore distributed smaller amounts of capacity in, say, urban centers as opposed to being centralized in highly efficient green datacenters.
Impact on energy companies
AI应用的迅速扩张以及由此导致的数据中心能耗激增，与向全电力社会的过渡发生了冲突。电力公司已经面临着电力容量翻番的巨大挑战——甚至还没有考虑到数据中心带来的额外压力。在北美，大多数电力公司预计负荷增长来自有机需求，工业和运输电气化进一步加剧了能源需求。
The rapid expansion of AI applications and the resulting surge in datacenter energy consumption clash with the transition to an all-electric society. Utilities already face the monumental challenge of doubling power capacity — without even accounting for the added strain from datacenters. In North America, most utilities anticipate load growth from organic demand, with industrial and transportation electrification further intensifying energy needs.
资料来源：451 Research
问：考虑到你们对负载增长的预期，您预计哪一部分业务的增长比例最大？
基础：北美受访者。
从这个角度来看，DeepSeek的指标是令人充满希望的。仅使用用10%的GPU（比OpenAI使用GPU能耗更低）,就能训练出一个性能相当的模型，这些GPU的能耗更低，这可能会使AI的总能耗降低多达95%。然而，更高的效率和更低的推理成本可能会加速各行业的采用，而DeepSeek、IBM公司和Red Hat的Granite公司的Granite等开源模型可能会将工作负载从云计算转移到边缘计算。能源需求将同样分散，而不是具有专用需求的集中式AI集群。
From this perspective, DeepSeek's metrics are promising. Training a model with comparable performance with just 10% the number of GPUs, which have a lower energy demand than the GPUs used by OpenAI, could drive AI's net energy demand down by as much of 95%. However, more efficiency and lower inferencing cost will likely accelerate adoption across industries, and open-source models such as DeepSeek or IBM Corp. and Red Hat's Granite will likely shift workloads from cloud to edge. Rather than a centralized AI cluster with a dedicated demand, the energy demand will be equally decentralized.
随着AI、大科技、数据中心行业和电力公司的爆炸式增长，为不断增长的需求提供电力都面临着挑战。鉴于风能和太阳能等分布式能源的间歇性以及它们对电网稳定性的影响，在财力雄厚的大型科技公司的支持下，已经启动了雄心勃勃的计划，以恢复核电站或委托新的小型模块化反应堆为电网提供足够的电力和惯性。AI的民主化和去中心化从根本上改变了这一商业案例。然而，认为大型科技公司将为能源转型买单的说法是一种谬论。虽然核能可能比天然气或燃煤火力发电更清洁，但最可持续的选择仍然是降低总体电力需求。
With the explosive growth of AI, Big Tech, the datacenter industry and utilities have all been challenged to provide power to the increasing demand. Given the intermittency of distributed energy resources, such as wind and solar, and their impact on grid stability, ambitious plans have been launched to revive nuclear plants or commission new small modular reactors to provide enough power and inertia to the grid, backed by the deep pockets of Big Tech. The democratization and decentralization of AI radically changes this business case. However, the argument that Big Tech would pay for the energy transition is a fallacy. While nuclear power may be cleaner than gas or coal-fired thermal power, the most sustainable option still is to drive down overall power demand.
由于知识产权、隐私或地缘政治方面的考虑，许多企业可能会对部署DeepSeek犹豫不决，但通常致力于净零战略的大型工业企业可能会率先采用类似的分布式AI集群。通过将AI的能源需求整合到他们的减排计划中，这些公司通过智能空间和微电网解决方案增强了他们的能源弹性，平衡了分布式能源和电池存储的消耗。这种转变并没有将发电的责任从大型科技公司转移到电力公司，而是转移到最终用户身上。
Many enterprises may hesitate to deploy DeepSeek due to IP, privacy or geopolitical concerns, but large industrial players — often committed to net-zero strategies — will likely lead the adoption of similar decentralized AI clusters. By integrating AI's energy needs into their emission reduction plans, these companies enhance their energy resilience through smart spaces and microgrid solutions, balancing consumption with distributed energy and battery storage. This shift does not transfer responsibility for power generation from Big Tech to utilities, but rather to the end user.
免责声明
：本文由451Research授权并认可的中文版本，仅供读者学习参考，不得用于任何商业用途，本文不代表CDCC观点。
“
第三届数据中心液冷大会”将于2025年3月27日在杭州召开
，这场行业盛事将集结数据中心领域的顶尖专家、技术大牛和学术研究者，共同探讨创新趋势，携手谋划合作机遇，助力行业蓬勃发展。
关键词：
液冷、全栈、最佳实践、芯片、服务器、基础设施
嘉宾规模：500+专业人士
会议形式：
展览展示+主旨演讲+技术分享+应用案例参观
